Audio Encoder for Encoding an Audio Signal Having an Impulse-like Portion and
Stationary Portion, Encoding Methods, Decoder, Decoding Method, and Encoded
Audio Signal
Specification
The present invention relates to source coding, and particularly, to audio source coding, in
which an audio signal is processed by at least two different audio coders having different
coding algorithms.
In the context of low bitrate audio and speech coding technology, several different coding
techniques have traditionally been employed in order to achieve low bitrate coding of such
signals with best possible subjective quality at a given bitrate. Coders for general music /
sound signals aim at optimizing the subjective quality by shaping spectral (and temporal)
shape of the quantization error according to a masking threshold curve which is estimated
from the input signal by means of a perceptual model ("perceptual audio coding")- On the
other hand, coding of speech at very low bitrates has been shown to work very efficiently
when it is based on a production model of human speech, i.e. employing Linear Predictive
Coding (LPC) to model the resonant effects of the human vocal tract together with an effi-
cient coding of the residual excitation signal.
As a consequence of these two different approaches, general audio coders (like MPEG-1
Layer 3, or MPEG-2/4 Advanced Audio Coding, AAC) usually do not perform as well for
speech signals at very low data rates as dedicated LPC-based speech coders due to the lack
of exploitation of a speech source model. Conversely, LPC-based speech coders usually do
not achieve convincing results when applied to general music signals because of their ina-
bility to flexibly shape the spectral envelope of the coding distortion according to a mask-
ing threshold curve. In the following, embodiments are described which provide a concept
that combines the advantages of both LPC-based coding and perceptual audio coding into a
single framework and thus describe unified audio coding that is efficient for both general
audio and speech signals.
Traditionally, perceptual audio coders use a filterbank-based approach to efficiently code
audio signals and shape the quantization distortion according to an estimate of the masking
curve.
Fig. 16a shows the basic block diagram of a monophonic perceptual coding system. An
analysis filterbank 1600 is used to map the time domain samples into subsampled spectral
components. Dependent on the number of spectral components, the system is also referred
to as a subband coder (small number of subbands, e.g. 32) or a transform coder (large
number of frequency lines, e.g. 512). A perceptual ("psychoacoustic") model 1602 is used
to estimate the actual time dependent masking threshold. The spectral ("subband'" or "fre-
quency domain") components are quantized and coded 1604 in such a way that the quanti-
zation noise is hidden under the actual transmitted signal, and is not perceptible after de-
coding. This is achieved by varying the granularity of quantization of the spectral values
over time and frequency.
The quantized and entropy-encoded spectral coefficients or subband values are, in addition
with side information, input into a bitstream formatter 1606, which provides an encoded
audio signal which is suitable for being transmitted or stored. The output bitstream of
block 1606 can be transmitted via the Internet or can be stored on any machine readable
data carrier.
On the decoder-side, a decoder input interface 1610 receives the encoded bitstream. Block
1610 separates entropy-encoded and quantized spectral/subband values from side informa-
tion. The encoded spectral values are input into an entropy-decoder such as a Huffman
decoder which is positioned between 1610 and 1620. The output of this entropy decoder is
quantized spectral values. These quantized spectral values are input into a re-quantizer
which performs an "inverse" quantization as indicated at 1620 in Fig. 16a. The output of
block 1620 is input into a synthesis filterbank 1622, which performs a synthesis filtering
including a frequency/time transform and, typically, a time domain aliasing cancellation
operation such as overlap and add and/or a synthesis-side windowing operation to finally
obtain the output audio signal.
Figs. 16b, 16c indicate an alternative to the entire filterbank based perceptual coding con-
cept of Fig. 16a, in which a pre-filtering approach on the encoder-side and a post-filtering
approach on the decoder-side are implemented.
In [EdlOO], a perceptual audio coder has been proposed which separates the aspects of irre-
levance reduction (i.e. noise shaping according to perceptual criteria) and redundancy re-
duction (i.e. obtaining a mathematically more compact representation of information) by
using a so-called pre-filter rather than a variable quantization of the spectral coefficients
over frequency. The principle is illustrated in Fig. 16b. The input signal is analyzed by a
perceptual model 1602 to compute an estimate of the masking threshold curve over fre-
quency. The masking threshold is converted into a set of pre-filter coefficients such that
the magnitude of its frequency response is inversely proportional to the masking threshold.
The pre-filter operation applies this set of coefficients to the input signal which produces
an output signal in which all frequency components are represented according to their per-
ceptual importance ("'perceptual whitening"). This signal is subsequently coded by any
kind of audio coder 1632 which produces a "white'" quantization distortion, i.e. does not
apply any perceptual noise shaping. The transmission / storage of the audio signal include
both, the coder's bitstream and a coded version of the pre-filtering coefficients. In the de-
coder of Fig. 16c, the coder bitstream is decoded (1634) into the perceptually whitened
audio signal which contains additive white quantization noise. This signal is then subjected
to a post-filtering operation 1640 according to the transmitted filter coefficients. Since the
post-filter performs the inverse filtering process relative to the pre-filter, it reconstructs the
original audio input signal from the perceptually whitened signal. The additive white quan-
tization noise is spectrally shaped like the masking curve by the post-filter and thus appears
perceptually colored at the decoder output, as intended.
Since in such a scheme perceptual noise shaping is achieved via the pre-/post-filtering step
rather than frequency dependent quantization of spectral coefficients, the concept can be
generalized to include non-filterbank-based coding mechanism for representing the pre-
filtered audio signal rather than a filterbank-based audio coder. In [Sch02] this is shown for
time domain coding kernel using predictive and entropy coding stages.
In order to enable appropriate spectral noise shaping by using pre-/post-filtering tech-
niques, it is important to adapt the frequency resolution of the pre-/post-filter to that of the
human auditor} system. Ideally, the frequency resolution would follow well-known per-
ceptual frequency scales, such as the BARK or ERB frequency scale [Zwi]. This is espe-
cially desirable in order to minimize the order of the pre-/post-filter model and thus the
associated computational complexity and side information transmission rate.
The adaptation of the pre-/post-filter frequency resolution can be achieved by the well-
known frequency warping concept [KHL97]. Essentially, the unit delays within a filter
structure are replaced by (first or higher order) allpass filters which leads to a non-uniform
deformation ("warping") of the frequency response of the filter. It has been shown that
even by using a first-order allpass filter, e.g.
a quite accurate approximation of perceptual frequency scales is possible by an appropriate
choice of the allpass coefficients [SA99]. Thus, most known systems do not make use of
higher-order allpass filters for frequency warping. Since a first-order allpass filter is fully
determined by a single scalar parameter (which will be referred to as the '"warping factor' -
1 < < 1), which determines the deformation of the frequency scale. For example, for a
warping factor of =0, no deformation is effective, i.e. the filter operates on the regular
frequency scale. The higher the warping factor is chosen, the more frequency resolution is
focused on the lower frequency part of the spectrum (as it is necessary to approximate a
perceptual frequency scale), and taken away from the higher frequency part of the spec-
trum).
Using a warped pre-/post-filter, audio coders typically use a filter order between 8 and 20
at common sampling rates like 48 kHz or 44.1 kHz [WSKH05].
Several other applications of warped filtering have been described, e.g. modeling of room
impulse responses [HKS00] and parametric modeling of a noise component in the audio
signal (under the equivalent name Laguerre / Kauz filtering) [SOB03]
Traditionally, efficient speech coding has been based on Linear Predictive Coding (LPC)
to model the resonant effects of the human vocal tract together with an efficient coding of
the residual excitation signal [VM06], Both LPC and excitation parameters are transmitted
from the encoder to the decoder. This principle is illustrated in Figs. 17a and 17b.
Fig. 17a indicates the encoder-side of an encoding/decoding system based on linear predic-
tive coding. The speech input is input into an LPC analyzer 1701 which provides, at its
output, LPC filter coefficients. Based on these LPC filter coefficients, an LPC filter 1703 is
adjusted. The LPC filter outputs a spectrally whitened audio signal which is also termed
"prediction error signal". This spectrally whitened audio signal is input into a resi-
dual/excitation coder 1705 which generates excitation parameters. Thus, the speech input
is encoded into excitation parameters on the one hand, and LPC coefficients on the other
hand.
On the decoder-side illustrated in Fig. 17b, the excitation parameters are input into an exci-
tation decoder 1707 which generates an excitation signal which can be input into an in-
verse LPC filter. The inverse LPC filter is adjusted using the transmitted LPC filter coeffi-
cients. Thus, the inverse LPC filter 1709 generates a reconstructed or synthesized speech
output signal.
Over time, many methods have been proposed with respect to an efficient and perceptually
convincing representation of the residual (excitation) signal, such as Multi-Pulse Excitation
(MPE), Regular Pulse Excitation (RPE), and Code-Excited Linear Prediction (CELP).
Linear Predictive Coding attempts to produce an estimate of the current sample value of a
sequence based on the observation of a certain number of past values as a linear combina-
tion of the past observations. In order to reduce redundancy in the input signal, the encoder
LPC filter "whitens" the input signal in its spectral envelope, i.e. it is a model of the in-
verse of the signal's spectral envelope. Conversely, the decoder LPC filter is a model of
the signal's spectral envelope. Specifically, the well-known auto-regressive (AR) linear
predictive analysis is known to model the signal's spectral envelope by means of an all-
pole approximation.
Typically, narrow band speech coders (i.e. speech coders with a sampling rate of 8kHz)
employ an LPC filter with an order between 8 and 12. Due to the nature of the LPC filter, a
uniform frequency resolution is effective across the full frequency range. This does not
correspond to a perceptual frequency scale.
Noticing that a non-uniform frequency sensitivity, as it is offered by warping techniques,
mav offer advantages also for speech coding, there have been proposals to substitute the
regular LPC analysis by warped predictive analysis, e.g. [TMK.94] [KTK.95]. Other com-
binations of warped LPC and CELP coding are known, e.g. from [HLM99].
In order to combine the strengths of traditional LPC/CELP-based coding (best quality for
speech signals) and the traditional filterbank-based perceptual audio coding approach (best
for music), a combined coding between these architectures has been proposed. In the
AMR-WB+ coder [BLS05] two alternate coding kernels operate on an LPC residual signal.
One is based on ACELP (Algebraic Code Excited Linear Prediction) and thus is extremelj
efficient for coding of speech signals. The other coding kernel is based on TCX (Trans-
form Coded Excitation), i.e. a filterbank based coding approach resembling the traditional
audio coding techniques in order to achieve good quality for music signals. Depending on
the characteristics of the input signal signals, one of the two coding modes is selected for a
short period of time to transmit the LPC residual signal. In this way, frames of 80 ms dura-
tion can be split into subframes of 40 or 20 ms in which a decision between the two coding
modes is made.
A limitation of this approach is that the process is based on a hard switching decision be-
tween two coders / coding schemes which possess extremely different characteristics re-
garding the type of introduced coding distortion. This hard switching process may cause
annoying discontinuities in perceived signal quality when switching from one mode to
another. For example, when a speech signal is slowly cross-faded into a music signal (such
as after an announcement in a broadcasting program), the point of switching may be de-
tectable. Similarly, for speech over music (like for announcements with music back-
ground), the hard switching may become audible. With this architecture, it is thus hard to
obtain a coder which can smoothly fade between the characteristics of the two component
coders.
Recently, also a combination of switched coding has been described that permits the filter-
bank-based coding kernel to operate on a perceptually weighted frequency scale by fading
the coder's filter between a traditional LPC mode (as it is appropriate for CELP-based
speech coding) and a warped mode which resembles perceptual audio coding based on pre-
/post-filtering as discussed on EP 1873754.
Using a filter with variable frequency warping, it is possible to build a combined speech /
audio coder which achieves both high speech and audio coding quality in the following
way as indicated in Fig. 17c:
The decision about the coding mode to be used ("Speech mode" or "'Music mode") is per-
formed in a separate module 1726 by carrying out an analysis of the input signal and can
be based on known techniques for discriminating speech signals from music. As a result,
the decision module produces a decision about the coding mode / and an associated opti-
mum warping factor for the filter 1722. Furthermore, depending on this decision, it deter-
mines a set of suitable filter coefficients which are appropriate for the input signal at the
chosen coding mode, i.e. for coding of speech, an LPC analysis is performed (with no
warping, or a low warping factor) whereas for coding of music, a masking curve is esti-
mated and its inverse is converted into warped spectral coefficients.
The filter 1722 with the time varying warping characteristics is used as a common encoder
/ decoder filter and is applied to the signal depending on the coding mode decision / warp-
ing factor and the set of filter coefficients produced by the decision module.
The output signal of the filtering stage is coded by either a speech coding kernel 1724 (e.g.
CELP coder) or a generic audio coder kernel 1726 (e.g. a filterbank-based coder, or a pre-
dictive audio coder), or both, depending on the coding mode.
The information to be transmitted / stored comprises the coding mode decision (or an indi-
cation of the warping factor), the filter coefficients in some coded form, and the informa-
tion delivered by the speech / excitation and the generic audio coder.
In the corresponding decoder, the outputs of the residual / excitation decoder and the ge-
neric audio decoder are added up and the output is filtered by the time varying warped syn-
thesis filter, based on the coding mode, warping factor and filter coefficients.
Due to the hard switching decision between two coding modes, the scheme is. however,
still subject to similar limitations as the switched CELP / filterbank-based coding as they
were described previously. With this architecture, it is hard to obtain a coder which can
smoothly fade between the characteristics of the two component coders.
Another way of combining a speech coding kernel with a generic perceptual audio coder is
used for MPEG-4 Large-Step Scalable Audio Coding [Gri97] [Her02]. The idea of scalable
coding is to provide coding/decoding schemes and bitstream formats that allow meaningful
decoding of subsets of a full bitstream, resulting in a reduced quality output signal, in this,
the transmitted/decoded data rate can be adapted to the instantaneous transmission channel
capacity without a re-encoding of the input signal.
The structure of an MPEG-4 large-step scalable audio coder is depicted by Fig. 18 [Gri97].
This configuration comprises both a so-called core coder 1802 and several enhancement
layers based on perceptual audio coding modules 1804. The core coder (typically a narrow
band speech coder) operates at a lower sampling rate than the subsequent enhancement
layers. The scalable combination of these components works as follows:
The input signal is down-sampled 1801 and encoded by the core coder 1802. The produced
bitstream constitutes the core layer portion 1804 of the scalable bitstream. It is decoded
locally 1806 and upsampled 1808 to match the sampling rate of the perceptual enhance-
ment layers and passed through the analysis filterbank (MDCT) 1810.
In a second signal path, the delay (1812) compensated input signal is passed through the
analysis filterbank 1814 and used to compute the residual coding error signal.
The residual signal is passed through a Frequency Selective Switch (FSS) tool 1816 which
permits to fall back to the original signal on a scalefactor band basis if this can be coded
more efficiently than the residual signal.
The spectral coefficients are quantized/coded by an AAC coding kernel 1804, leading to an
enhancement layer bitstream 1818.
Further stages of refinement (enhancement layers) by re-coding of the residual coding error
signal can follow.
Fig. 19 illustrates the structure of the associated core-based scalable decoder. The compo-
site bit-stream is decomposed 1902 into the individual coding layers. Decoding 1904 of the
core coder bitstream (e.g. a speech coder bitstream) is then performed and its output signal
may be presented via an optional post filter stage. In order to use the core decoder signal
within the scalable decoding process, it is upsampled 1908 to the sampling rate of the scal-
able coder, delay compensated 1910 with respect to the other layers and de-composed by
the coder analysis filterbank (MDCT) 1912.
Higher layer bitstreams are then decoded 1916 by applying the AAC noiseless decoding
and inverse quantization, and summing up 1918 all spectral coefficient contributions. A
Frequency Selective Switch tool 1920 combines the resulting spectral coefficients with the
contribution from the core layer by selecting either the sum of them or only the coefficients
originating from the enhancement layers as signaled from the encoder. Finally, the result is
mapped back to a time domain representation by the synthesis filterbank (IMDCT) 1922.
As a general characteristic, the speech coder (core coder) is always used and decoded in
this configuration. Only if a decoder has access not only to the core layer of the bitstream
but also to one or more enhancement layers, also contributions from the perceptual audio
coders in the enhancement layers are transmitted which can provide a good quality for non-
speech / music signals.
Consequently, this scalable configuration always includes an active layer containing a
speech coder which leads to some drawbacks regarding its performance to provide best
overall quality for both speech and audio signals:
If the input signal is a signal that predominantly consists of speech, the perceptual audio
coder in the enhancement layer(s) code a residual / difference signal that has properties that
may be quite different from that of regular audio signals and are thus hard to code for this
type of coder. As one example, the residual signal may contain components which are im-
pulsive of nature and therefore provoke pre-echoes when coded with a filterbank-based
perceptual audio coder.
if the input signal is not predominantly speech, the residual signal frequently requires more
bitrate to code than the input signal. In these cases, the FSS selects the original signal for
coding by the enhancement layer rather than the residual signal. Consequently, the core
layer does not contribute to the output signal and the bitrate of the core layer is spent in
vain since it does not contribute to an improvement of the overall quality. In other words,
in such cases the result sounds worse that if the entire bitrate would have simply been allo-
cated to a perceptual audio coder only.
In http://www.hitech-projects.com/euprojects/ardor/summary.htm
the ARDOR (Adaptive Rate-Distortion Optimised sound codeR) codec is described as fol-
lows:
Within the project, a codec is created that encodes generic audio with the most appropriate
combination of signal models, given the imposed constraints as well as the available sub-
coders. The work can be divided into three parts corresponding to the three codec compo-
nents as illustrated in Fig. 20.
A rate-distortion-theory based optimization mechanism 2004 that configures the ARDOR
codec such that it operates most efficiently given the current, time-varying, constraints and
type of input signal. For this purpose it controls: a set of 'subcoding' strategies 2000. each
of which is highly efficient for encoding a particular type of input-signal component, e.g..
tonal, noisy, or transient signals. The appropriate rate and signal-component allocation for
each particular subcoding strategy is based on: an advanced, new perceptual distortion
measure 2002 that provides a perceptual criterion for the rate-distortion optimization me-
chanism. In other words, a perceptual model, which is based on state-of-the-art knowledge
about the human auditor}' system, provides the optimization mechanism with information
about the perceptual relevance of different parts of the sound. The optimization algorithm
could for example decide to leave out information that is perceptually irrelevant. Conse-
quently, the original signal cannot be restored, but the auditory system will not be able to
perceive the difference.
The above discussion of several prior art systems underlines that there does not yet exist an
optimum encoding strategy which, on the one hand provides optimum quality for general
audio signals as well as speech signals, and which on the other hand, provides a low bitrate
for all kinds of signals. Particularly, the scalable approach as discussed in connection with
Fig. 18 and Fig. 19 which has been standardized in MPEG-4 continuously processes the
whole audio signal using a speech coder core without paying attention to the audio signal
and, specifically, to the source of the audio signal. Therefore, when the audio signal is not
speech-like, the core encoder will introduce heavy coding artifacts and, consequently, the
frequency selective switch tool 1816 in Fig. 18 will make sure that the full audio signal is
encoded using the AAC encoder core 1804. Thus, in this instance, the bitstream includes
the useless output of the speech core coder, and additionally includes the perceptually en-
coded representation of the audio signal. This not only results in a waste of transmission
bandwidth, but also results in a high and useless power consumption, which is particularly
problematic when the encoding concept is to be implemented in mobile devices which are
battery-powered and have limited resources of energy.
Generally stated, the transform-based perceptual encoder operates without paying attention
to the source of the audio signal, which results in the fact that, for all available sources of
signals, the perceptual audio encoder (when having a moderate bit rate) can generate an
output without too many coding artifacts, but for non-stationary signal portions, the bitrate
increases, since the masking threshold does not mask as efficient as in stationary sounds.
Furthermore, the inherent compromise between time resolution and frequency resolution in
transform-based audio encoders renders this coding system problematic for transient or
impulse-like signal components, since these signal components would require a high time
resolution and would not require a high frequency resolution.
The speech coder, however, is a prominent example for a coding concept, which is heavily
based on a source model. Thus, a speech coder resembles a model of the speech source,
and is, therefore, in the position to provide a highly efficient parametric representation for
signals originating from a sound source similar to the source model represented by the cod-
ing algorithm. For sounds originating from sources which do not coincide with the speech
coder source model, the output will include heavy artifacts or, when the bitrate is allowed
to increase, will show up a bitrate which is drastically increased and substantially higher
than a bitrate of a general audio coder.
It is the object of the present invention to provide an improved concept for audio coding,
which provides advantages from both a specific coder concept and a general coder concept,
and which avoids, as far as possible, the disadvantages of these coders.
This object is achieved by an audio coder in accordance with claim 1, a method of audio
encoding in accordance with claim 24, a decoder for decoding an encoded audio signal in
accordance with claim 25, a method of decoding in accordance with claim 32, an enhanced
audio signal in accordance with claim 33, or a computer program in accordance with claim
34.
The present invention is based on the finding that a separation of impulses from an audio
signal will result in a highly efficient and high quality audio encoding concept. By extract-
ing impulses from the audio signal, an impulse audio signal on the one hand and a residual
signal corresponding to the audio signal without the impulses is generated. The impulse
audio signal can be encoded by an impulse coder such as a highly efficient speech coder,
which provides extremely low data rates at a high quality for speech signals. On the other
hand, the residual signal is freed of its impulse-like portion and is mainly constituted of the
stationary portion of the original audio signal. Such a signal is very well suited for a signal
encoder such as a general audio encoder and, preferably, a transform-based perceptually
controlled audio encoder. An output interface outputs the encoded impulse-like signal and
the encoded residual signal. The output interface can output these two encoded signals in
any available format, but the format does not have to be a scalable format, due to the fact
that the encoded residual signal alone, or the encoded impulse-like signal alone, may under
special circumstances not be of significant use by itself. Only both signals together will
provide a high quality audio signal.
On the other hand, however, the bitrate of this combined encoded audio signal can be con-
trolled to a high degree, when preferably a fixed rate impulse coder such as an CELP or
ACELP encoder is used, which can be tightly controlled with respect to its bitrate. On the
other hand, the signal encoder is, when for example, implemented as an MP3 or MP4 en-
coder, controllable so that it outputs a fixed bitrate, although performing a perceptual cod-
ing operation which inherently outputs a variable bitrate, based on an implementation of a
bit reservoir as known in the art for MP3 or MP4 coders. This will make sure that the bi-
trate of the encoded output signal is a constant bitrate.
Due to the fact that the residual audio signal does not include the problematic impulse-like
portions anymore, the bitrate of the encoded residual signal will be low. since this residual
signal is optimally suited for the signal encoder.
On the other hand, the impulse encoder will provide an excellent and efficient operation,
since the impulse encoder is fed with a signal which is specifically shaped and selected
from the audio signal to fit perfectly to the impulse coder source model. Thus, when the
impulse extractor is not able to find impulse portions in the audio signal, then the impulse
encoder will not be active and will not try to encode any signal portions which are not at all
suitable for being coded with the impulse coder. In view of this, the impulse coder will also
not provide an encoded impulse signal and will also not contribute to the output bitrate for
signal portions where the impulse coder would require a high bitrate or would not be in the
position to provide an output signal having an acceptable quality. Specifically, for mobile
applications, the impulse coder will also not require any energy resources in such a situa-
tion. Thus, the impulse coder will only become active when the audio signal includes an
impulse-like portion and the impulse-like portion extracted by the impulse extractor will
also be perfectly in line with what the impulse encoder expects.
Thus, the distribution of the audio signal to two different coding algorithms will result in a
combined coding operation, which is specifically useful in that the signal encoder will be
continuously active and the impulse coder will work as a kind of a fallback module, which
is only active and only produces output bits and only consumes energy, if the signal actual-
ly includes impulse-like portions.
Preferably, the impulse coder is adapted for advantageously encoding sequences of im-
pulses which are also called "impulse trains" in the art. Theses "pulses"' or "impulse trains'1
are typical pattern obtained by modeling the human vocal tract. A pulse train has impulses
at time-distances between adjacent impulses. Such a time distance is called a "pitch lag",
and this value corresponds with the "pitch frequency".
Preferred embodiments of the present invention are subsequently discussed in connection
with the accompanying drawings, in which:
Fig. 1 is a block diagram of an audio encoder in accordance with an embodiment
of the present invention;
Fig. 2 is a block diagram of a decoder for decoding an encoded audio signal;
Fig. 3a illustrates an open-loop embodiment;
Fig. 3b illustrates a specific embodiment of a decoder;
Fig. 4a illustrates another open-loop embodiment of the encoder-side:
Fig. 4b illustrates a closed-loop embodiment of the encoder-side;
Fig. 4c illustrates an embodiment in which the impulse extractor and the impulse
coder are implemented within a modified ACELP coder;
Fig. 5a illustrates a wave form of a time domain speech segment as an impulse-like
signal segment;
Fig. 5b illustrates a spectrum of the segment of Fig. 5a;
Fig. 5c illustrates a time domain speech segment of unvoiced speech as an example
for a stationary segment;
Fig. 5d illustrates a spectrum of the time domain wave form of Fig. 5c;
Fig. 6 illustrates a block diagram of an analysis by synthesis CELP encoder:
Figs. 7a to 7d illustrate voiced/unvoiced excitation signals as an example for im-
pulse-like and stationary signals;
Fig. 7e illustrates an encoder-side LPC stage providing short-term prediction infor-
mation and the prediction error signal;
Fig. 8 illustrates a preferred embodiment of the Fig. 4a open-loop embodiment;
Fig. 9a illustrates a wave form of a real impulse-like signal;
Fig. 9b illustrates an enhanced or more ideal impulse-like signal as generated by the
impulse characteristic enhancement stage of Fig. 8;
Fig. 10 illustrates a modified CELP algorithm implementable in the Fig. 4c embo-
diment;
Fig. 11 illustrates a more specific implementation of the algorithm of Fig. 10;
Fig. 12 illustrates a specific implementation of the algorithm of Fig. 1 I;
Fig. 13 illustrates another modified CELP algorithm implemented in Fig. 4c;
Fig. 14 illustrates the operation modes illustrating the continuous operation of the
signal decoder and the intermittent operation of the impulse decoder;
Fig. 15 illustrates an encoder embodiment in which the signal encoder includes a
psychoacoustic model;
Fig. 16a illustrates an MP3 or MP4 coding/decoding concept;
Fig. 16b illustrates a pre-filter encoding concept;
Fig. 16c illustrates a post-filter decoding concept;
Fig. 17a illustrates an LPC encoder;
Fig. 17b illustrates and LPC decoder;
Fig. 17c illustrates a coder implementing switched coding with a dynamically varia-
ble warped LPC filter;
Fig. 18 illustrates an MEPG-4 scalable encoder;
Fig. 19 illustrates an MPEG-4 scalable decoder; and
Fig. 20 illustrates a schematic diagram of an ARDOR encoder.
It is an advantage of the following embodiments to provide a unified method that extends a
perceptual audio coder to allow coding of not only general audio signals with optimal qual-
ity, but also provide significantly improved coded quality for speech signals. Furthermore,
they enable the avoidance of problems associated with a hard switching between an audio
coding mode (e.g. based on a filterbank) and a speech coding mode (e.g. based on the
CELP approach) that were described previously. Instead, below embodiments allow for a
smooth / continuous combined operation of coding modes and tools, and in this way
achieves a more graceful transition / blending for mixed signals.
The following considerations form a basis for the following embodiments:
Common perceptual audio coders using filterbanks are well-suited to represent signals that
may have considerable fine structure across frequency, but are rather stationary over time.
Coding of transient or impulse-like signals by filterbank-based coders results in a smearing
of the coding distortion over time and thus can lead to pre-echo artifacts.
A significant part of speech signals consists of trains of impulses that are produced by the
human glottis during voiced speech with a certain pitch frequency. These pulse train struc-
tures are therefore difficult to code by filterbank-based perceptual audio coders at low bi-
trates.
Thus, in order to achieve optimum signal quality with a filterbank-based coding system, it
is advantageous to decompose the coder input signal into impulse-like structures and other,
more stationary components. The impulse-like structures may be coded with a dedicated
coding kernel (hereafter referred to as the impulse coder) whereas the other residual com-
ponents may be coded with the common filterbank-based perceptual audio coder (hereafter
referred to as the residual coder). The pulse coder is preferably constructed from functional
blocks from traditional speech coding schemes, such as an LPC filter, information on pulse
positions etc. and may employ techniques such as excitation codebooks, CELP etc.
The separation of the coder input signal may be carried out such that two conditions are
met:
(Condition #1) Impulse-like signal characteristics for impulse coder input: Preferably, the
input signal to the impulse coder only comprises impulse-like structures in order to not
generate undesired distortion since the impulse coder is especially optimized to transmit
impulsive structures, but not stationary (or even tonal) signal components. In other words,
feeding tone-like signal components into the impulse coder will lead to distortions which
cannot be compensated easily by the filterbank-based coder.
(Condition #2) Temporally smooth impulse coder residual for the residual coder: The resi-
dual signal which is coded by the residual coder is preferably generated such that after the
split of the input signal, the residual signal is stationary over time, even at time instances
where pulses are coded by the pulse coder. Specifically, it is preferred that no "holes*' in
the temporal envelope of the residual are generated.
In contrast to the aforementioned switched coding schemes, a continuous combination be-
tween impulse coding and residual coding is achieved by having coders (the impulse coder
and the residual coder) and their associated decoders run in parallel, i.e. simultaneously, if
the need arises. Specifically, in a preferred way of operation, the residual coder is always
operational, while the impulse coder is only activated when its operation is found to be
beneficial.
A part of the proposed concept is to split the input signal into partial input signals that are
optimally adapted to the characteristics of each partial coder (impulse coder and residual
coder) in order to achieve optimum overall performance. In the following preferred embo-
diments, the following is assumed.
One partial coder is a filterbank-based audio coder (similar to common perceptual audio
coders). As a consequence, this partial coder is well-suited to process stationary and tonal
audio signals (which in a spectrogram representation correspond to "horizontal struc-
tures'"), but not to audio signals which contain many instationarities over time, such as
transients, onsets or impulses (which in a spectrogram representation correspond to "ver-
tical structures"). Trying to encode such signals with the filterbank-based coder will lead to
temporal smearing, pre-echoes and a reverberant characteristic of the output signal.
The second partial coder is an impulse coder which is working in the time domain. As a
consequence, this partial coder is well-suited to process audio signals which contain many
instationarities over time, such as transients, onsets or impulses (which in a spectrogram
representation correspond to "'vertical structures"), but not to represent stationary and tonal
audio signals (which in a spectrogram representation correspond to "horizontal struc-
tures"). Trying to encode such signals with the time-domain impulse coder will lead to
distortions of tonal signal components or harsh sounding textures due to the underlying
sparse time domain representation.
The decoded output of both the filterbank-based audio decoder and the time-domain im-
pulse decoder are summed up to form the overall decoded signal (if both the impulse and
the filterbank-based coder are active at the same time).
Fig. 1 illustrates an audio encoder for encoding an audio signal 10 having an impulse-like
portion and a stationary portion. Generally, a differentiation between an impulse-like por-
tion of an audio signal and a stationary portion of a stationary signal can be made by apply-
ing a signal processing operation, in which the impulse-like characteristic is measured and
the stationary-like characteristic is measured as well. Such measurements can, for example,
be done by analyzing the wave form of the audio signal. To this end, any transform-based
processing or LPC processing or any other processing can be performed. An intuitive way
of determining as to whether a portion is impulse-like or not is for example to look at a
time domain wave form and to determine whether this time domain wave form has peaks
at regular or irregular intervals, and peaks in regular intervals are even more suited for a
speech-like coder.
Exemplarily, reference is made to Figs. 5a to 5d. Here, impulse-like signal segments or
signal portions and stationary signal segments or signal portions are exemplarily discussed.
Specifically, a voiced speech as illustrated in Fig. 5a in the time domain and in Fig. 5b in
the frequency domain is discussed as an example for an impulse-like signal portion, and an
unvoiced speech segment as an example for a stationary signal portion is discussed in con-
nection with Figs. 5c and 5d. Speech can generally be classified as voiced, unvoiced, or
mixed. Time-and-frequency domain plots for sampled voice and unvoiced segments are
shown in Fig. 5a to 5d. Voiced speech is quasi periodic in the time domain and harmoni-
cally structured in the frequency domain, while unvoiced speed is random-like and broad-
band. In addition, the energy of voiced segments is generally higher than the energy of
unvoiced segments. The short-time spectrum of voiced speech is characterized by its tine
and formant structure. The fine harmonic structure is a consequence of the quasi-
periodicity of speech and may be attributed to the vibrating vocal chords. The formant
structure (spectral envelope) is due to the interaction of the source and the vocal tracts. The
vocal tracts consist of the pharynx and the mouth cavity. The shape of the spectral
envelope that "fits" the short time spectrum of voiced speech is associated with the transfer
characteristics of the vocal tract and the spectral tilt (6 dB /Octave) due to the glottal pulse.
The spectral envelope is characterized by a set of peaks which are called formants. The
formants are the resonant modes of the vocal tract. For the average vocal tract there are
three to five formants below 5 kHz. The amplitudes and locations of the first three for-
mants, usually occurring below 3 kHz are quite important both, in speech synthesis and
perception. Higher formants are also important for wide band and unvoiced speech repre-
sentations. The properties of speech are related to the physical speech production system as
follows. Voiced speech is produced by exciting the vocal tract with quasi-periodic glottal
air pulses generated by the vibrating vocal chords. The frequency of the periodic pulse is
referred to as the fundamental frequency or pitch. Unvoiced speech is produced by forcing
air through a constriction in the vocal tract. Nasal sounds are due to the acoustic coupling
of the nasal tract to the vocal tract, and plosive sounds are produced by abruptly releasing
the air pressure which was built up behind the closure in the tract.
Thus, a stationary portion of the audio signal can be a stationary portion in the time domain
as illustrated in Fig. 5c or a stationary portion in the frequency domain, which is different
from the impulse-like portion as illustrated for example in Fig. 5a, due to the fact that the
stationary portion in the time domain does not show prominent repeating pulses. As will be
outlined later on, however, the differentiation between stationary portions and impulse-like
portions can also be performed using LPC methods, which model the vocal tract and the
excitation of the vocal tracts. When the frequency domain representation of the signal is
considered, impulse-like signals show the prominent occurrence of the individual formants.
i.e., the prominent peaks in Fig. 5b, while a stationary spectrum has quite a white spectrum
as illustrated in Fig. 5d, or in the case of harmonic signals, quite a continuous noise floor
having some prominent peaks representing specific tones which occur, for example, in the
music signal, but which do not have such a regular distance from each other as the im-
pulse-like signal in Fig. 5b.
Furthermore, impulse-like portions and stationary portions can occur in a timely manner,
i.e.. which means that a portion of the audio signal in time is stationary and another portion
of the audio signal in time is impulse-like. Alternatively, or additionally, the characteristic
of a signal can be different in different frequency bands. Thus, the determination, whether
the audio signal is stationary or impulse-like, can also be performed frequency-selective so
that a certain frequency band or several certain frequency bands are considered to be sta-
tionary and other frequency bands are considered to be impulse-like. In this case, a certain
time portion of the audio signal might include an impulse-like portion and a stationary por-
tion.
The Fig. 1 encoder embodiment includes an impulse extractor 10 for extracting the im-
pulse-like portion from the audio signal. The impulse extractor 10 includes an impulse
coder for encoding the impulse-like portion to obtain an encoded impulse-like signal. As
will be shown later on, the impulse extraction and the actual encoding operation can be
separated from each other, or can be combined so that one obtains a single algorithm such
as the ACELP algorithm in its modified form as discussed in connection with Fig. 4c.
The output of the impulse extractor 10 is an encoded impulse signal 12 and, in some em-
bodiments, additional side information relating to the kind of impulse extraction or the
kind of impulse encoding.
The Fig. 1 encoder embodiment furthermore includes a signal encoder 16 for encoding a
residual signal 18 derived from the audio signal 10 to obtain an encoded residual signal 20.
Specifically, the residual signal 18 is derived from the audio signal 10 so that the impulse-
like portions in the audio signal are reduced or completely eliminated from the audio sig-
nal. Nevertheless, the audio signal still includes the stationary portion, since the stationary
portion has not been extracted by the impulse extractor 10.
Furthermore, the inventive audio encoder includes an output interface 22 for outputting the
encoded impulse signal 12, the encoded residual signal 20 and, if available, the side infor-
mation 14 to obtain an encoded signal 24. The output interface 22 does not have to be a
scalable datastream interface producing a scalable datastream which is written in a manner
that the encoded residual signal and the encoded impulse signal can be decoded indepen-
dent of each other, and a useful signal is obtained. Due to the fact that neither the encoded
impulse signal, nor the encoded residual signal will be an audio signal with an acceptable
audio quality, rendering of only one signal without the other signal does not make any
sense in preferred embodiments. Thus, the output interface 22 can operate in a completely
bit-efficient manner, without having to worry about the datastream, and whether it can be
decoded in a scalable way or not.
In a preferred embodiment, the inventive audio decoder includes a residual signal genera-
tor 26. The residual signal generator 26 is adapted for receiving the audio signal 10 and
information 28 relating to the extracted impulse signal portions, and for outputting the re-
sidual signal 18 which does not include the extracted signal portions. Depending on the
implementation, the residual signal generator 26 or the signal encoder 16 may output side
information as well. Output and transmission of side information 14. however, is not nec-
essarily required due to the fact that a decoder can be pre-set in a certain configuration and,
as long as the encoder operates based on these configurations, the inventive encoder does
not need to generate and transmit any additional side information. Should there, however,
be a certain flexibility on the encoder side and on the decoder side, or should there be a
specific operation of the residual signal generator which is different from a pure subtrac-
tion, it might be useful to transmit side information to the decoder so that the decoder and,
specifically, the combiner within the decoder, ignores portions of the decoded residual sig-
nal which have been introduced on the encoder side only to have a smooth and non-
impulse-like residual signal without any holes.
Fig. 2 illustrates a preferred decoder embodiment for decoding an encoded audio signal 24
which is the same signal as is output by the output interface 22. Generally, the encoded
audio signal 24 includes an encoded impulse-like signal and an encoded residual signal.
The decoder may comprise a decoder input interface 28 for extracting the encoded impulse
signal 12. the encoded residual signal 20, and the side information 14 from the encoded
audio signal 24. The encoded impulse signal 12 is input into an impulse decoder 30 for
decoding the encoded impulse-like signal using a decoding algorithm adapted to a coding
algorithm used for generating the encoded impulse-like signal, i.e., coding algorithm as
was used in block 10 of Fig. 12. The decoder in Fig. 2 furthermore comprises a signal de-
coder 32 for decoding the encoded residual signal using a decoding algorithm adapted to a
coding algorithm used for generating the encoded residual signal, i.e., a coding algorithm
used in block 16 of Fig. 1. The output signals of both decoders 30 and 32 are forwarded to
an input into a signal combiner 34 for combining the decoded impulse-like signal and the
decoded residual signal to provide a decoded output signal 36. Specifically, the signal de-
coder 32 and the impulse decoder 30 are operative to provide for selected portions of the
decoded audio signal output values relating to the same time instant of the decoded audio
signal.
This characteristic will be discussed in connection with Fig. 14. Fig. 14 schematically illu-
strates an output of the signal decoder 32 at 140. It is illustrated in Fig. 14 that the output
140 of the signal decoder continuously exists. This means that the signal decoder (and the
corresponding signal encoder) continuously operates and always provides an output signal
as long as the audio signal exists. Naturally, only when the audio track is over, the signal
decoder will stop its output as well, since there is no input signal to encode anymore.
The second line in Fig. 14 illustrates the impulse decoder output 142. Specifically, it is
outlined in Fig. 14 that there are portions 143, in which there does not exist an impulse
decoder output due to the fact that the original audio signal did not have any stationan
components in these time portions 143. However, in the other time portions, the signal had
stationary components and/or impulse-like components, and the impulse-like components
are generated by the impulse decoder output. Thus, in the time portions 142, both decoders
provide output values which are related to the same time instant of a decoded signal. How-
ever, in the time portions 143, the output signal only consists of the residual signal decoder
output and does not have any contribution from the impulse decoder.
Fig. 3a illustrates a preferred embodiment of an encoder in a so-called open-loop configu-
ration. The impulse extractor 10 includes a general impulse extractor for generating a non-
encoded impulse signal indicated on line 40. The impulse extractor is indicated at 10a. The
impulse signal 40 is forwarded to the impulse coder 10b which finally outputs the encoded
impulse signal 12. The information on the impulse signal on line 28 corresponds to the
non-encoded impulse signal as extracted by the impulse extractor 10a. The residual signal
generator 26 is implemented in Fig. 3a as a subtractor for subtracting the non-encoded im-
pulse signal on line 28 from the audio signal 10 to obtain the residual signal 18.
Preferably, the signal encoder 16 is implemented as a filterbank based audio encoder, since
such a filterbank based audio encoder is specifically useful for encoding a residual signal
which does not have any impulse-like portions anymore, or in which the impulse-like por-
tions are at least attenuated with respect to the original audio signal 10. Thus, the signal is
put through a first processing stage 10a which is designed to provide the input signals of
the partial coders at its output. Specifically, the splitting algorithm is operative to generate
output signals on line 40 and line 18 which fulfill the earlier discussed condition 1 (the
impulse coder receives impulse-like signals) and condition 2 (the residual signal for the
residual coder is temporarily smoothed). Thus, as illustrated in Fig. 3a, the impulse extrac-
tion module 10a extracts the impulse signal from the audio input signal 10.
The residual signal 18 is generated by removing the impulse signal from the audio input.
This removal can be done by subtraction as is indicated in Fig. 3a, but can also be per-
formed by other measures such as replacing the impulse-like region of the audio signal by
a less impulse-like ("flattened") signal that can be derived from the original audio signal 10
by appropriate time-variant scaling or interpolation between regions to the left and right of
the impulse-like region. In the consecutive parallel coding stages 10b, 16, the impulse sig-
nal (if present) is coded by a dedicated impulse coder 10b and the residual signal is coded
preferably by a filterbank-based audio coder 16.
In a different preferred embodiment, in which a time portion of the audio signal has been
detected as impulse-like, a pure cutting out operation of this time portion and encoding the
portion only with the impulse coder would result in a hole in the residual signal for the
signal coder. In order to avoid this hole, which is a problematic discontinuity for the signal
encoder, a signal to be introduced into the "hole" is synthesized. This signal can be, as dis-
cussed later, an interpolation signal or a weighted version of the original signal or a noise
signal having a certain energy.
In one embodiment, this interpolated/synthesized signal is subtracted from the impulse like
"cut-out" signal portion so that only the result of this subtraction operation (the result is an
impulse-like signal as well) is forwarded to the impulse coder. This embodiment will make
sure that - on the decoder side - the output of the residual coder and the output of the im-
pulse decoder can be combined in order to obtain the decoded signal. In this embodiment,
all signals obtained by both output decoders are always used and combined to obtain the
output signal, and any discarding of an output of any one of both decoders will not take
place.
Subsequently, other embodiments of the residual signal generator 26, apart from a subtrac-
tion, are discussed.
As stated before, a time-variant scaling of the audio signal can be done. Specifically, as
soon as an impulse-like portion of the audio signal is detected, a scaling factor can be used
for scaling the time domain samples of the audio signal with a scaling factor value of less
than 0.5 or, for example, even less than 0.1. This results in a decrease of the energy of the
residual signal at the time period in which the audio signal is impulse-like. However, in
contrast to simply setting to 0 the original audio signal in this impulse-like period, the resi-
dual signal generator 26 makes sure that the residual signal does not have any "holes",
which are again, instationarities which would be quite problematic for the filterbank based
audio coder 16. On the other hand, the encoded residual signal during the impulse-like time
portion which is the original audio signal multiplied by a small scaling factor might not be
used on the decoder-side, or might only to a small degree be used on the decoder-side. This
fact may be signaled by a certain additional side information 14. Thus, a side information
bit generated by such a residual signal generator might indicate which scaling factor was
used for down-scaling the impulse-like portion in the audio signal, or which scaling factor
is to be used on the decoder-side to correctly assemble the original audio signal after hav-
ing decoded the individual portions.
Another way of generating the residual signal is to cut out the impulse-like portion of the
original audio signal and to interpolate the cut out portion using the audio signal at the be-
ginning or at the end of the impulse-like portion in order to provide a continuous audio
signal, which is however, no longer impulse-like. This interpolation can also be signaled
by a specific side information bit 14, which generally provides information regarding the
impulse coding or signal coding, or residual signal generation characteristic. On the decod-
er side, a combiner can fully delete, or at least attenuate to a certain degree, the decoded
representation of the interpolated portion. The degree or indication can be signaled via a
certain side information 14.
Furthermore, it is preferred to provide the residual signal so that a fade-in and a fade-out
occurs. Thus, the time-variant scaling factor is not abruptly set to a small value, but is con-
tinuously reduced until the small value and, at the end or around the end of the impulse-
like portion, the small scaling factor is continuously increased to a scaling factor in the
regular mode, i.e., to a small scaling factor of 1 for an audio signal portion which does not
have an impulse-like characteristic.
Fig. 3b illustrates a decoder which corresponds to the encoder in Fig. 3a, where the signal
decoder 32 of Fig. 2 is implemented as a filterbank based audio decoder, and where the
signal combiner 34 is implemented as a sample-wise adder.
Alternatively, the combination performed by the signal combiner 34 can also be performed
in the frequency domain or in the subband domain provided that the impulse decoder 30
and the filterbank based audio decoder 32 provide output signals in the frequency domain
or in the subband domain.
Furthermore, the combiner 34 does not necessarily have to perform a sample-wise addi-
tion, but the combiner can also be controlled by side information such as the side informa-
tion 14 as discussed in connection with Figs. 1, 2 and 3a, in order to apply a time variant
scaling operation in order to compensate for encoder-side fade in and fade out operations,
and in order to handle signal portions which have been generated on the encoder-side to
flatten the residual signals, such as by insertion, interpolation, or time-variant scaling.
When, the residual signal generator 26 is operative to perform a sample-wise subtraction as
indicated in Fig. 3a, then the decoder-side combiner 34 will not require any additional side
information and will perform a sample-wise addition without any additional processing
steps such as fade, fade out, or signal scaling.
For voiced speech signals, the excitation signal, i.e., the glottal impulses are filtered by the
human vocal tracts which can be inverted by an LPC filter. Thus, the corresponding im-
pulse extraction for glottal impulses typically may include an LPC analysis before the ac-
tual impulse picking stage and an LPC synthesis before calculating the residual signal as is
illustrated in Fig. 4a, which is additionally, an open-loop implementation.
Specifically, the audio signal 8 is input into an LPC analysis block 10a. The LPC analysis
block produces a real impulse-like signal as is, for example, illustrated in Fig. 9a. This sig-
nal is input into an impulse picking stage 10c, which processes the real impulse-like signal,
as for example illustrated in Fig. 9a, in order to output an impulse signal which is an ideal
or at least a more ideal impulse-like signal compared to the real impulse-like signal at the
input into the impulse picking stage 10c. This impulse signal is subsequently input into the
impulse coder 10b. The impulse coder 10b provides a high quality representation of the
input impulse-like signal, due to the fact that this coder is specifically suited for such im-
pulse-like signals and due to the fact that the input impulse signal on line 48 is an ideal, or
almost ideal, impulse signal. In the Fig. 4a embodiment, the impulse signal on line 48,
which corresponds to the "information on impulse signal" 28 of Fig. 1, is input into an
LPC synthesis clock 26b in order to "transform" the ideal impulse-like signal which exists
in the "LPC domain" back into the time domain. Thus, the output of the LPC synthesis
block 26b is then input into the subtracter 26a, so that a residual signal 18 is generated,
which is the original audio signal, but which no longer includes the pulse structure
represented by the ideal impulse signal on line 48 or 28. Thus, the residual signal generator
26 of Fig. 1 is implemented in Fig. 4a as the LPC synthesis block 26b and the subtracter
26a.
The functionality of the LPC analysis 10a and the LPC synthesis 26b will subsequently be
discussed in more detail with respect to Figs. 7a to 7e, Fig. 8, and Fig. 9a to 9b.
Fig. 7a illustrates a model of a linear speech production system. This system assumes a
two-stage excitation, i.e., an impulse-train for voiced speech as indicated in Fig. 7a, and a
random-noise for unvoiced speech as indicated in Fig. 7d. The vocal tract is modeled as an
all-pole transform filter 70 which processes pulses of Fig. 7c or Fig. 7d, generated by the
glottal model 72. The all-pole transfer function is formed by a cascade of a small number
of two-pole resonators representing the formants. The glottal model is represented as a
two-pole low-pass filter, and the lip-radiation model 74 is represented by L(z)=l-z"'. Final-
ly, a spectral correction factor 76 is included to compensate for the low-frequency effects
of the higher poles. In individual speech representations the spectral correction is omitted
and the 0 of the lip-radiation function is essentially cancelled by one of the glottal poles.
Hence, the system of Fig. 7a can be reduced to an all pole-model of Fig. 7b having a gain
stage 77, a forward path 78, a feedback path 79, and an adding stage 80. In the feedback
path 79, there is a prediction filter 81, and the whole source-system synthesis model illu-
strated in Fig. 7b can be represented using z-domain functions as follows:
S(z)=g/(1-A(z))-X(z),
where g represents the gain, A(z) is the prediction filter as determined by an LPC analysis.
X(z) is the excitation signal, and S(z) is the synthesis speech output.
Figs. 7c and 7d give a graphical time domain description of voiced and unvoiced speech
synthesis using the linear source system model. This system and the excitation parameters
in the above equation are unknown and must be determined from a finite set of speech
samples. The coefficients of A(z) are obtained using linear prediction. In a p-th order for-
ward linear predictor, the present sample of the speech sequence is predicted from a linear
combination oft passed samples. The predictor coefficients can be determined by well-
known algorithms such as the Levinson-Durbin algorithm, or generally an autocorrelation
method or a reflection method.
Fig. 7e illustrates a more detailed implementation of the LPC analysis block 10a of Fig. 4a.
The audio signal is input into a filter determination block which determines the filter in-
formation A(z). This information is output as the short-term prediction information re-
quired for a decoder. In the Fig. 4a embodiment, i.e., the short-term prediction information
might be required for the impulse coder output signal. When, however, only the prediction
error signal at line 84 is required, the short-term prediction information does not have to be
output. Nevertheless, the short-term prediction information is required by the actual predic-
tion filter 85. In a subtractor 86, a current sample of the audio signal is input and a pre-
dieted value for the current sample is subtracted so that for this sample, the prediction error
signal is generated at line 84. A sequence of such prediction error signal samples is very
schematically illustrated in Fig. 9a, where, for clarity issues, any issues regarding AC/DC
components, etc. have not been illustrated. Therefore, Fig. 9a can be considered as a kind
of a rectified impulse-like signal.
Fig. 8 will subsequently be discussed in more detail. Fig. 8 is similar to Fig. 4a, but shows
block 10a and block 26b in more detail. Furthermore, a general functionality of the impulse
characteristic enhancement stage 10c is discussed. The LPC analysis stage 10a in Fig. 8
can be implemented as shown in detail in Fig. 7e, where the short-term prediction informa-
tion A(z) is input into the synthesis stage 26b, and the prediction error signal which is the
"real impulse-like signal" is output here at line 84. When it is assumed that the signal is
mixed, i.e., includes speech components and other components, then the real impulse-like
signal might be considered as a superposition of the excitation signals in Figs. 7c and 7d,
which in a rectified representation, correspond to Fig. 9a. One can see a real impulse-like
signal which, additionally, has stationary components. These stationary components are
removed by the impulse characteristic enhancement stage 10c, which provides at its output,
a signal which is for example similar to Fig. 9b. Alternatively, the signal output by block
10c can be the result of a pure peak picking which means that an impulse, starting from
some samples to the left of the peak and ending at some samples to the right of the peak, is
picked out from the signal in Fig. 9a, where signal samples of the signal in Fig. 9a between
the peaks are completely discarded. This would mean that a similar signal as the one
shown in Fig. 7c is generated by block 10c, with the difference that the impulses are not
ideal DIRAC pulses, but have a certain impulse width. Furthermore, the impulse characte-
ristic enhancement stage 10c can be operative to process the peaks so that each peak has
the same height and shape which is schematically illustrated in Fig. 9b.
The signal generated by block 10c will be ideally suited for the impulse coder 10b and the
impulse coder will provide an encoded representation requiring a small number of bits and
being a representation of the ideal impulse-like signal without, or only with a very small
amount of quantization errors.
The LPC synthesis stage 26b in Fig. 8 can be implemented in exactly the same manner as
the all-pole model in Fig. 7b, with a unity gain or a gain different from 1, so that the trans-
fer function as indicated in block 26b is implemented in order to have a representation of
the ideal pulse-like signal at the output of block 10c in the time domain, so that a sample-
wise combination such as a subtraction can be performed in block 26a. Then, the output of
block 26a will be the residual signal, which in an ideal case, only includes the stationary
portion of the audio signal and no longer includes the impulse-like portion of the audio
signal. Any information loss introduced by performing the impulse characteristic en-
hancement operation in block 10c such as peak picking is non-problematic, since this er-
ror' is accounted for in the residual signal and is not lost. Importantly, however, the posi-
tions of the impulses picked by stage 10c precisely represent the impulse positions in the
audio signal 8 so that the combination of both signals in block 26a, especially when made
using a subtraction, does not result in two pulses which are closely adjacent to each other,
but results in a signal without any pulses, since a pulse in the original audio signal 8 has
been cancelled due to the combination operation by block 26a.
This feature is an advantage of the so-called "open-loop embodiment" and might be a dis-
advantage of the so-called "closed-loop embodiment" which is illustrated in Fig. 4b. Fig.
4b is different from Fig. 4a in that the impulse coder output signal is input into an impulse
decoder 26e, which is a part of the residual signal generator 26 of Fig. 1. When the impulse
coder 10b introduces quantization errors into the positions of the pulses, and when these
errors are not compensated by the operation of the impulse decoder 26c, then the subtrac-
tion operation in block 26a will result in a residual signal which not only has the original
pulses in the audio signal, but has, in the neighborhood to these pulses, additional pulses
which have been introduced due to the subtraction operation. In order to avoid this situa-
tion, the combiner 26 can be operative to not just perform a sample-wise subtraction, but to
perform an analysis of the impulse decoder 26c output signal, so that a synchronized sub-
traction is obtained.
The "closed-loop" operation can also be considered as a cascaded splitting operation. One
of the two partial coders (preferably the impulse coder) is tuned to accept an appropriate
part of the input signal (preferably the glottal impulses). Then, the other partial coder 16 is
fed by the residua! signal consisting of the difference signal between the original signal and
the decoded signal from of the first partial coder. The impulse signal is first coded and de-
coded, and the quantized output is subtracted from the audio input in order to generate the
residual signal in the closed-loop approach, which is coded by the filterbank-based audio
coder.
As an example, a CELP or an ACELP coder can be used as an efficient impulse coder as
illustrated in Fig. 4c. which will be discussed later. Preferably, however, the CELP or
ACELP routine is modified such that the coder only models impulsive parts of the input
signal, rather than trying to also model tonal or very stationary signal components. In other
words, once a certain number of impulses are spent to model impulsive signal parts, the
allocation of more impulses to model the other parts of the signal would be counterproduc-
tive and would deteriorate the quality of the overall output signal. Thus, an appropriate
preprocessor or controller, as for example illustrated at 1000 in Fig. 10, terminates the im-
pulse allocation procedure once all actually occurring impulses are modeled.
Furthermore, it is preferred that the residual after removal from the impulse coder output
signal is constructed such that it becomes rather flat over time in order to fulfill condition
number 2, in order to be suitable for coding with the filterbank-based coder 16 of Fig. 4c.
Thus, Fig. 4c illustrates this approach, in which the modified ACELP coder 10 operates
both, as the impulse extractor and impulse coder. Again, the residual signal generator 26 of
Fig. 1 uses a subtraction 26a to remove the impulse-like portions from the audio signal, but
also other methods can be applied such as flattening or interpolation, as previously de-
scribed.
The disadvantage of the open-loop implementation of Fig. 4b, in which the signal is first
separated into an impulse signal and a residual signal, with both signal portions then being
coded individually, and which involves lossy coding, i.e. quantization in both the impulse
coder and the filterbank-based audio coder, is that the quantization errors of both coders
have to be controlled and perceptually minimized individually. This is due to the fact that
at the decoder output, both quantization errors add up.
However, the advantage of the open-loop implementation is that the impulse extraction
stage produces a clean impulse signal, which is not distorted by quantization errors. Thus
the quantization in the impulse coder does not affect the residual signal.
Both implementations can, however, be mixed in order to implement a kind of mixed
mode. Thus, components from both the open-loop and the closed-loop approaches are im-
plemented together.
An efficient impulse coder usually quantizes both the individual values and the positions of
the impulses. One option for a mixed open/closed-loop mode is to use the quantized im-
pulse values and the accurate/unquantized impulse positions for calculating the residual
signal. The impulse position is then quantized in an open-loop fashion. Alternatively, an
iterative CELP analysis-by-synthesis process for the detection of impulse-like signals can
be used, but a dedicated coding tool for the actual coding the impulse signal is imple-
mented, which quantizes or not, the position of the pulses with a small quantization error.
Subsequently, an analysis-by-synthesis CELP encoder will be discussed in connection with
Fig. 6 in order to illustrate the modifications applied to this algorithm, as illustrated in Figs.
10 to 13. This CELP encoder is discussed in detail in "Speech Coding: A Tutorial Re-
view", Andreas Spanias, Proceedings of the IEEE, Vol. 82, No. 10, October 1994, pages
1541-1582. The CELP encoder as illustrated in Fig. 6 includes a long-term prediction
component 60 and a short-term prediction component 62. Furthermore, a codebook is used
which is indicated at 64. A perceptual weighting filter W(z) is implemented at 66, and an
error minimization controller is provided at 68. s(n) is the excitation signal as, for example,
generated by the LPC analysis stage 10a. This signal is also called "prediction error signal"
as indicated at line 84 in Fig. 7e. After having been perceptually weighted, the weighted
prediction error signal is input into a subtractor 69, which calculates the error between the
synthesis signal at the output of block 66 and the actual weighted prediction error signal
s(w)(n). Generally, the short-term prediction A(z) is calculated by a LPC analysis stage as
indicated in Fig. 7e, and depending on this information, the long-term prediction informa-
tion AL(z) including the long-term prediction gain g and the vector quantization index, i.e..
codebook references are calculated. The CELP algorithm encodes the excitation using a
codebook of for example Gaussian sequences. The ACELP algorithm, where the "A"
stands for "Algebraic" has a specific algebraically designed codebook.
A codebook may contain more or less vectors where each vector is some samples long. A
gain factor g scales the excitation vector and the excitation samples are filtered by the
long-term synthesis filter and the short-term synthesis filter. The "optimum" vector is se-
lected such that the perceptually weighted mean square error is minimized. The search
process in CELP is evident from the analysis-by-synthesis scheme illustrated in Fig. 6.
Subsequently, an exemplary ACELP algorithm is described in connection with Fig. 10,
which additionally illustrates the modification performed in accordance with an embodi-
ment of the present invention discussed in connection with Fig. 4c.
The publication "A simulation tool for introducing Algebraic CELP (ACELP) coding con-
cepts in a DSP course", Frontiers in Education Conference, Boston, Massachusetts, 2002,
Venkatraman Atti and Andreas Spanias, illustrates a description of an educational tool for
introducing code excited linear prediction (CELP) coding concepts in University courses.
The underlying ACELP algorithm includes several stages, which include a pre-processing
and LPC analysis stage 1000, an open-loop pitch analysis stage 1002, a closed-loop pitch
analysis stage 1004, and an algebraic (fixed) codebook search stage 1006.
In the pre-processing and LPC analysis stage, the input signal is high-pass filtered and
scaled. A second order pole-zero filter with a cut-off frequency of 140 Hz is used to per-
form the high-pass filtering. In order to reduce the probability of overflows in a fixed-point
implementation, a scaling operation is performed. Then, the preprocessed signal is win-
dowed using a 30 ms (240 samples) asymmetric window. A certain overlap is implemented
as well. Then, using the Levinson-Durbin algorithm, the linear prediction coefficients are
computed from the autocorrelation coefficients corresponding to the windowed speech.
The LP coefficients are converted to line spectral pairs which are later quantized and
transmitted. The Levinson-Durbin algorithm additionally outputs reflection coefficients
which are used in the open-loop pitch analysis block for calculating an open-loop pitch Top
by searching the maximum of an autocorrelation of a weighted speech signal, and by read-
ing out the delay at this maximum. Based on this open-loop pitch, the closed-loop pitch
search stage 1004 is searching a small range of samples around Top to finally output a high-
ly accurate pitch delay and a long-term prediction gain. This long-term prediction gain is
additionally used in the algebraic fixed codebook search and finally output together with
other parametric information as quantized gain values. The algebraic codebook consists of
a set of interleaved permutation codes containing few non-zero elements which have a spe-
cific codebook structure in which the pulse position, the pulse number, an interleaving
depth, and the number of bits describing pulse positions are referenced. A search codebook
vector is determined by placing a selected amount of unit pulses at found locations where a
multiplication with their signs is performed as well. Based on the codebook vector, a cer-
tain optimization operation is performed which selects, among all available code vectors,
the best-fitting code vector. Then, the pulse positions and the times of the pulses obtained
in the best-fitting code vector are encoded and transmitted together with the quantized gain
values as parametric coding information.
The data rate of the ACELP output signal depends on the number of allocated pulses. For a
small number of pulses, such as a single pulse, a small bitrate is obtained. For a higher
number of pulses, the bitrate increases from 7.4 kb/s to a resulting bitrate of 8.6 kb/s for
five pulses, until a bitrate of 12.6 kb/s for ten pulses.
In accordance with a preferred embodiment of the present as discussed in Fig. 4c, the mod-
ified ACELP coder 10 includes a pulse number control stage 1000. Specifically, the pulse
number control stage measures the LTP gain as output by the closed-loop pitch analysis
and performs a pulse number control, if the LTP gain is low. A low LTP gain indicates that
the actually processed signal is not very much impulse-train like, and a high LTP gain in-
dicates that the actual signal is impulse-train like, and therefore, very suitable for the
ACELP encoder.
Fig. 11 illustrates a preferred implementation of a block 1000 of Fig. 10. Specifically, a
block 1010 determines, whether the LTP gain is greater than a predetermined LTP gain
threshold. When this is the case, it is determined that the signal is pulse-like at 1011. Then,
a predetermined or inherent number of pulses is used as indicated at 1012. Thus, a
straightforward pulse setting or a straightforward pulse number control of an ACELP en-
coding algorithm is applied without any modification, but a pulse position variation intro-
duced by this encoder is partly or completely restricted, to a periodic grid based on past
information in order to make sure that the disadvantage of the closed-loop embodiment is
eliminated, as indicated at block 1013. Specifically, if the long-term predictor (LTP) gain
is high, i.e., the signal is periodic and pulses were placed in the past frames, i.e., the signal
is impulse-like, the algebraic codebook is used to refine the impulse shapes by restricting
possible pulse positions to a periodic grid determined by past pulse positions and the LTP
lag. Specifically, the number of pulses placed by the algebraic codebook may be constant
for this mode, as indicated at block 1011.
If it is determined that the long-term predictor (LTP) gain is low, as indicated at 1014. the
number of pulses is varied in the codebook optimization, as indicated at 1015. Specifically,
the algebraic codebook is controlled such that it is allowed to place pulses in such a man-
ner that the energy of the remaining residual is minimized and the pulse positions form a
periodic pulse train with the period equal to the LTP lag. The process, however is stopped
when the energy difference is below a certain threshold, which results in a variable number
of pulses in the algebraic codebook.
Subsequently, Fig. 12 is discussed in order to provide a preferred embodiment of the varia-
tion of the number of pulses described in connection with block 1015. At the beginning,
the optimization is performed using a small number of pulses, such as a single pulse, as
indicated at 1016. Then, the optimization is performed with this small number of pulses, as
indicated at 1017. For the best matching code vector, the error signal energy is calculated
in block 1018 and is compared to an error energy threshold (THR) in block 1019. the thre-
shold is predetermined and may be suitably set to a value which makes sure that the
ACELP encoder only encodes the pulse portion of the signal with a certain accuracy, but
does not try to encode non-pulse-like portions of the signal as well, which the coder would
do when the inventive controller 1000 of Fig. 10 were not there.
When step 1019 determines that the threshold is met, the procedure is stopped. When,
however, the comparison in block 1019 determines that the error signal energy threshold is
not yet met. the number of pulses is increased, for example by 1. as indicated at 1020.
Then, steps 1017, 1018, and 1019 are repeated, but now with a higher number of pulses.
This procedure is continued until a final criterion such as a maximum number of allowed
pulses is met. Normally, however, the procedure will stop due to the threshold criterion, so
that generally the number of pulses for a non-pulse-like-signal will be smaller than the
number of pulses which the encoding algorithm would allocate in the case of a pulse-like
signal.
Another modification of an ACELP encoder is illustrated in Fig. 13. In such an encoder, a
voiced/unvoiced decision is performed as indicated at 1300. Such an encoder then uses
depending on this voice/unvoiced decision, a first codebook for voiced portions, and a
second codebook for unvoiced portions. In accordance with an embodiment of the present
invention, the CELP analysis-by-synthesis procedure is only used for determining impulse
code information when a voiced portion has been detected by block 1300 as is indicated at
1310. When, however, the CELP encoder determines an unvoiced portion, then the CELP
encoder output for these unvoiced portions is not calculated or at least ignored and not in-
cluded into the encoded impulse signal. In accordance with the present invention, these
unvoiced portions are encoded using the residual coder and, therefore, the modification of
such an encoder consists of ignoring the encoder output for unvoiced portions as indicated
at 1320.
The present invention is preferably combined to the concept of switched coding with a
dynamically variable warped LPC filter, as indicated in Fig. 17. The impulse coder em-
ploys an LPC filter, where the impulse coder is represented by block 1724. If the filter-
bank-based residual coder contains a pre/post-filtering structure, it is possible to use a uni-
fied time-frequency representation for both the pulse coder 1724 and the residual coder,
which is not indicated in Fig. 17c. since a processing of the audio input apart from apply-
ing the pre-filter 1722 is not performed, but would be performed in order to provide the
input into the generic audio coder 1726 which would correspond to the residual signal cod-
er 16 of Fig. 1. In this way one can avoid two analysis filters at the encoder-side and two
synthesis filters at the decoder-side. This may include a dynamic adaptation of a genera-
lized filter in its warping characteristics, as has been described with respect to Fig. 17c.
Thus, the present invention can be implemented into the framework of Fig. 17c b>
processing the pre-filter 1722 output signal before inputting this signal into the generic
audio coder 1726, and by additionally extracting the pulses from the audio signal before
the audio signal is input into a residual excitation coder 1724. Thus, blocks 10c, 26b, and
26a would have to be placed at the output of the time-varying warped filter 1722 and the
input into the residual/excitation coder 1724 which would correspond to the impulse coder
10b in Fig. 4a and the input of the generic audio coder 1726 which would correspond to the
filterbank-based audio coder 16 in Fig. 4a. Naturally, the closed-loop embodiment of Fig.
4b can additionally be implemented into the Fig. 17c encoding system.
Preferably, a psychoacoustically controlled signal encoder 16 of Fig. 1 is used. Preferably,
the psychoacoustic model 1602, which is for example similar to the corresponding block in
Fig. 16a is implemented in Fig. 15 so that its input is connected to the audio signal 8. This
makes sure that the psychoacoustic masking threshold information on line 1500 reflects the
situation of the original audio signal, rather than the residual signal at the output of the
residual signal generator 26. Thus, the quantizer 1604a is controlled by masking threshold
information 1500 which is not derived from the signal actually quantized, but which is
derived from the original audio signal before the residual signal 18 was calculated. This
procedure is preferably over a connection of psychoacoustic model input to the output of
the residual signal generator 26 due to the fact that the masking effect of the impulse-like
signal portion is utilized as well so that the bitrate can be further decreased. On the other
hand, however, a connection of the psychoacoustic model input to the output of the resi-
dual signal generator 18 might also be useful, since the residual signal is an actual audio
signal, and consequently, has a masking threshold. However, although this implementation
is generally possible and useful for certain applications, it will produce a higher bitrate
compared to the situation in which the psychoacoustic model 1602 is fed with the original
audio signal.
Generally, preferred embodiments of the present invention have several aspects which can
be summarized as follows.
Encoding side: Method of signal splitting; filterbank-based layer is always present; the
speech enhancement is an optional layer; performing a signal analysis (the impulse extrac-
tion) prior to the coding; the impulse coder handles only a certain component of the input
signal; the impulse coder is tuned to handle only impulses; and the Filterbank-based layer is
an unmodified filterbank-based coder. Decoding side: filterbank-based layer is always
present; and the speech enhancement is an optional layer.
Generally, the impulse coding method is selected in addition to the filterbank-based coding
mode if the underlying source model for the impulses (e.g. glottal impulse excitation) fits
well for the input signal, the impulse coding can start at any convenient point in time; the
impulse coding mode is selected in addition to the filterbank-based coding mode if the un-
derlying source model for the impulses (e.g. glottal impulse excitation) fits well for the
input signal; and this does not involve an analysis of the rate-distortion behavior of both
codec and is therefore vastly more efficient in the encoding process.
A preferred impulse coding or pulse train coding method is the technique of waveform
interpolation as described in "Speech coding below 4 kB/s using waveform interpolation",
W. B. Kleijn, Globecom '91, pages 1879 to 1883, or in "'A speech coder based on decom-
position of characteristic waveforms", W. B. Kleijn and J. Haagen, ICASSP 1995, pages
508 to 511.
The below-described embodiments are merely illustrative for the principles of the present
invention. It is understood that modifications and variations of the arrangements and the
details described herein will be apparent to others skilled in the art. It is the intent, there-
fore, to be limited only by the scope of the impending patent claims and not by the spe-
cific details presented by way of description and explanation of the embodiments herein.
Depending on certain implementation requirements of the inventive methods, the inventive
methods can be implemented in hardware or in software. The implementation can be per-
formed using a digital storage medium, in particular, a disc, a DVD or a CD having elec-
tronically-readable control signals stored thereon, which co-operate with programmable
computer systems such that the inventive methods are performed. Generally, the present
invention is therefore a computer program product with a program code stored on a ma-
chine-readable carrier, the program code being operated for performing the inventive me-
thods when the computer program product runs on a computer. In other words, the inven-
tive methods are, therefore, a computer program having a program code for performing at
least one of the inventive methods when the computer program runs on a computer.
Literature:
[EdlOO] B. Edler, G. Schuller: "Audio coding using a psychoacoustic pre- and post-
filter", ICASSP 2000, Volume 2, 5-9 June 2000 Page(s):11881 - II884 vol.2;
[Sch02] G. Schuller, B. Yu, D. Huang, and B. Edler, "Perceptual Audio Coding using
Adaptive Pre- and Post-Filters and Lossless Compression", IEEE Transactions
on Speech and Audio Processing, September 2002, pp. 379-390;
I'Zwi] Zwicker, E. and H. Fasti, "Psychoacoustics, Facts and Models", Springer
Verlag, Berlin:
[KHL97] M. Karjalainen, A. Harma, U.K. Laine, "Realizable warped IIR filters and their
properties", IEEE ICASSP 1997, pp. 2205-2208, vol.3;
[SA99] J.O. Smith, J.S. Abel, "Bark and ERB Bilinear Transforms", IEEE Transac-
tions on Speech and Audio Processing, Volume 7, Issue 6, Nov. 1999, pp. 697
-708:
[HKS00] Harma, Aki; Karjalainen, Matti; Savioja, Lauri; Valimaki, Vesa; Laine. Unto
K.; Huopaniemi, Jyri, "Frequency-Warped Signal Processing for Audio Appli-
cations'", Journal of the AES, Volume 48 Number 11 pp. 1011-1031; Novem-
ber 2000;
[SOB03] E. Schuijers, W. Oomen, B. den Brinker, J. Breebaart, "Advances in Parame-
tric Coding for High-Quality Audio", 114th Convention, Amsterdam. The
Netherlands 2003. preprint 5852;
[WSKH05]S. Wabnik, G. Schuller, U. KraTner, J. Hirschfeld. ..Frequency Warping in Lou
Delay Audio Coding", IEEE International Conference on Acoustics, Speech,
and Signal Processing, March 18-23, 2005, Philadelphia, PA, USA:
[TMK94] K. Tokuda, H. Matsumura, T. Kobayashi and S. Imai, "Speech coding based on
adaptive mel-cepstral analysis," Proc. IEEE ICASSP'94, pp. 197-200. Apr.
1994;
[KTK95] K. Koishida, K. Tokuda, T. Kobayashi and S. Imai, "CELP coding based on
mel-cepstral analysis," Proc. IEEE ICASSP*95, pp.33-36, 1995;
[HLM99] Aki Harma, Unto K. Laine, Matti Karjalainen, "Warped low-delay CELP for
wideband audio coding", 17th International AES Conference, Florence, Italy,
1999;
[BLS05] B. Bessette, R. Lefebvre, R. Salami, "UNIVERSAL SPEECH/AUDIO COD-
ING USING HYBRID ACELP/TCX TECHNIQUES," Proc. IEEE ICASSP
2005. pp. 301 -304,2005;
[Gri97] Grill. B.. "A Bit Rate Scalable Perceptual Coder for MPEG-4 Audio", 103rd
AES Convention, New York 1997, Preprint 4620; and
[Her02] J. Herre, H. Purnhagen: "General Audio Coding", in F. Pereira, T. Ebrahimi
(Eds.), "The MPEG-4 Book", Prentice Hall IMSC Multimedia Series, 2002.
ISBN 0-13-061621-4.
We claim:
1. Audio encoder for encoding an audio signal (8) having an impulse-like portion and
a stationary portion, comprising:
an impulse extractor (10) for extracting the impulse-like portion from the audio
signal, the impulse-extractor having an impulse coder for encoding the impulse-like
portions to obtain an encoded impulse-like signal;
a signal encoder (16) for encoding a residual signal derived from the audio signal to
obtain an encoded residual signal (20), the residual signal being derived from the
audio signal so that the impulse-like portion is reduced or eliminated from the au-
dio signal; and
an output interface (22) for outputting the encoded impulse-like signal (12) and the
encoded residual signal (20), to provide an encoded signal (24),
wherein the impulse encoder is configured for not providing an encoded impulse-
like signal, when the impulse extractor is not able to find an impulse portion in the
audio signal.
2. Audio encoder in accordance with claim 1, wherein the impulse coder (10b) and the
signal coder (16) are designed such that the impulse coder (10b) is better suited for
impulse-like signals than the signal encoder (16) and that the signal encoder (16) is
better suited for stationary signals than the impulse coder (10b).
3. Audio encoder in accordance with claim 1 or 2, further comprising a residual signal
generator (26). the residual signal generator being adapted for receiving the audio
signal (8) and information (28) relating to the extracted impulse-like signal portions
and for outputting the residual signal (18) which does not include the extracted sig-
nal portions.
4. Audio encoder in accordance with claim 3. in which the residual signal generator
(26) includes a subtractor (26a) for subtracting the extracted signal portions (28)
from the audio signal (8) to obtain the residual signal (18).
5. Audio encoder in accordance with claim 3,
in which the impulse extractor (10) is operative to extract a parametric representa-
tion of the impulse-like signal portions; and
in which the residual signal generator (26) is operative to synthesize (26c) the wave
form representation using the parametric representation, and to subtract (26a) the
wave form representation from the audio signal (8).
6. Audio encoder in accordance with claim 3, in which the residual signal generator
(26) includes an impulse decoder (26c) for calculating a decoded impulse-like sig-
nal, and a subtracter (26a) for subtracting the decoded impulse-like signal from the
audio signal (8).
7. Audio encoder in accordance with claim 3,
in which the impulse extractor (10) includes an LPC analysis stage (10a) for per-
forming a LPC analysis of the audio signal (8). the LPC analysis being such that a
prediction error signal (84) is obtained,
in which the impulse extractor (10) includes a prediction error signal processor
(10c) for processing the prediction error signal such that an impulse like characte-
ristic of this signal is enhanced, and
in which the residual signal generator (26) is operative to perform an LPC synthesis
(26b) using the enhanced prediction error signal and to subtract a signal resulting
from the LPC synthesis from the audio signal to obtain the residual signal (18).
8. Audio encoder in accordance with one of the preceding claims, in which the im-
pulse extractor (10) comprises an impulse/non-impulse decision stage (1300), and
in which a portion of the audio signal being detected as an impulse-like portion is
provided to the impulse coder (10b) and is not provided to the signal encoder (16).
9. Audio encoder in accordance with claim 8, in which the impulse/non-impulse stage
is a voiced/unvoiced decision stage.
10. Audio encoder in accordance with one of the preceding claims, in which the audio
signal has a formant structure and a fine structure, in which the impulse extractor
(10) is operative to process the audio signal so that a processed signal only
represents the fine structure, and to process the fine structure signal so that the im-
pulse-like characteristic of the fine structure signal is enhanced (10c), and
in which the enhanced fine structure signal is encoded by the impulse coder (10b).
11. Audio encoder in accordance with one of the preceding claims, in which the signal
encoder (16) is a transform or filterbank based general audio encoder, and in which
the impulse coder is a time domain based coder.
12. Audio encoder in accordance with one of the preceding claims, in which the im-
pulse extractor (10) includes an ACELP coder (10) having an LPC analysis stage
(1000) to obtain short-term predictor information, a pitch determination stage
(1002, 1004) for obtaining pitch information and a long-term predictor gain, and a
codebook stage (1006) for determining codebook information relating to pulse po-
sitions of a number of pulses used for the parametric representation of a residual
signal, and
wherein the impulse extractor (10) is operative to control (1000) the ACELP coder
(10) depending on the long-term prediction gain to allocate either a variable num-
ber of pulses for the first long-term prediction gain or a fixed number of pulses for
a second long-term prediction gain, wherein the second long-term prediction gain is
greater than the first long-term prediction gain.
13. Audio encoder in accordance with claim 12, in which a maximum of the variable
number of pulses is equal or lower than the fixed number.
14. Audio encoder in accordance with claim 12 or 13, wherein the impulse extractor
(10) is operative to control (1000) the ACELP coder so that a gradual allocation
starting from a small number of pulses (1016) and proceeding to a higher number
of pulses (1020) is performed, and wherein the gradual allocation is stopped, when
an error energy is below a predetermined energy threshold (1019).
15. Audio encoder in accordance with one of the claims 12 to 14, in which the impulse
extractor (10) is operative to control (1000) the ACELP coder (10). so that in case
of a long-term predictor gain being higher than a threshold, possible pulse positions
are determined to be in a grid which is based on at least one pulse position from a
preceding frame (1013).
16. Audio encoder in accordance with claim 3,
in which the impulse coder (10b) is a code excited linear prediction (CELP) encod-
er calculating impulse positions and quantized impulse values, and
in which the residual signal generator (26) is operative to use unquantized impulse
positions and quantized impulse values for calculating a signal to be subtracted
from the audio signal (8) to obtain the residual signal.
17. Audio encoder in accordance with claim 3,
in which the impulse extractor includes a CELP analysis by synthesis process for
determining unquantized impulse positions in the prediction error signal, and
in which the impulse coder (10b) is operative to code the impulse position with a
precision higher than a precision of a quantized short-term prediction information.
18. Audio encoder in accordance with claim 3,
in which the impulse extractor (10) is operative to determine a signal portion as
impulse-like, and
in which the residual signal generator (26) is operative to replace the signal portion
of the audio signal by a synthesis signal having a reduced or no impulse-like struc-
ture.
19. Audio encoder in accordance with claim 18, in which the residual signal generator
(26) is operative to calculate the synthesis signal by extrapolation from a border be-
tween an impulse-like signal and the non-impulse-like signal.
20. Audio encoder in accordance with claim 18, in which the residual signal generator
is operative to calculate the synthesis signal by weighting the audio signal in the
impulse-like portion using a weighting factor smaller than 0.5.
21. Audio encoder in accordance with one of the preceding claims, in which the signal
encoder (16) is a psychoacoustically driven audio encoder, wherein a psychoacous-
tic masking threshold (1500) used from quantizing (1604a) audio values is calcu-
lated using the audio signal (8), and wherein the signal encoder (16) is operative to
convert the residua) signal in a spectral representation (1600) and to quantize
(1604a) values of the spectral representation using the psychoacoustic masking
threshold (1500).
22. Audio encoder in accordance with one of the preceding claims,
in which the impulse extractor (10) is operative to extract an impulse-like signal
from the audio signal to obtain an extracted impulse-like signal,
in which the impulse extractor (10) is operative (10c) to manipulate the extracted
impulse-like signal to obtain an enhanced impulse-like signal with a more ideal im-
pulse-like shape compared to a shape of the extracted impulse-like signal.
in which the impulse coder (10b) is operative to encode the enhanced impulse-like
signal to obtain an encoded enhanced impulse-like signal, and
in which the audio encoder includes a residual signal calculator (26) for subtracting
the extracted impulse-like signal or the enhanced impulse-like signal or a signal de-
rived by decoding the encoded enhanced impulse-like signal from the audio signal
to obtain the residual signal.
23. Audio encoder in accordance with one of the preceding claims, in which the im-
pulse extractor is operative for extracting an impulse train, and
in which the impulse coder is adapted for encoding an impulse-train like signal with
higher efficiency or less encoding error than a non-impulse-train like signal.
24. Method of encoding an audio signal (8) having an impulse-like portion and a sta-
tionary portion, comprising:
extracting (10) the impulse-like portion from the audio signal, the step of extracting
comprising a step of encoding the impulse-like portions to obtain an encoded im-
pulse-like signal;
encoding (16) a residual signal derived from the audio signal to obtain an encoded
residual signal (20), the residual signal being derived from the audio signal so that
the impulse-like portion is reduced or eliminated from the audio signal; and
outputting (22), by transmitting or storing, the encoded impulse-like signal (12) and
the encoded residual signal (20), to provide an encoded signal (24),
wherein the step of impulse encoding is not performed, when the step of impulse-
extracting does not find an impulse portion in the audio signal.
25. Decoder for decoding an encoded audio signal (24) having an encoded impulse-like
signal (12) and an encoded residual signal (20), comprising:
an impulse decoder (30) for decoding the encoded impulse-like signal using a de-
coding algorithm adapted to a coding algorithm used for generating the encoded
impulse-like signal, wherein a decoded impulse-like signal is obtained;
a signal decoder (32) for decoding the encoded residual signal (20) using a decod-
ing algorithm adapted to a coding algorithm used for generating the encoded resi-
dual signal (20), wherein a decoded residual signal is obtained; and
a signal combiner (34) for combining the decoded impulse-like signal and the de-
coded residual signal to provide a decoded output signal (36), wherein the signal
decoder (32) and the impulse decoder (30) are operative to provide output values
related to the same time instant of a decoded signal,
wherein the impulse decoder (30) is operative to receive the encoded impulse-like
signal and provide the decoded impulse-like signal at specified time portions (142)
separated by periods (143) in which the signal decoder provides the decoded resi-
dual signal and the impulse decoder (30) does not provide the decoded impulse-like
signal, so that the decoded output signal has the periods (143) in which the decoded
output signal is identical to the decoded residual signal and the decoded output sig-
nal has the specified time portions (142) in which the decoded output signal con-
sists of the decoded residual signal and the decoded impulse-like signal or consists
of the decoded impulse-like signal only.
26. Decoder in accordance with claim 25. in which the impulse decoder (30) is a time
domain decoder and the signal decoder (32) is a filterbank or transform based de-
coder.
27. Decoder in accordance with one of claims 25 to 26, in which the encoded audio
signal includes side information (14) indicating information relating to an encoding
or decoding characteristic pertinent to the residual signal, and
in which the combiner (34) is operative to combine the decoded residual signal and
the decoded impulse-like signal in accordance with the side information (14).
28. Decoder in accordance with one of claims 25 to 26. in which the side information
(14) indicates that, at an impulse-like portion, a synthetic signal has been generated
in the residual signal, and
in which the combiner (32) is operative to suppress or at least attenuate the decoded
residual signal during the impulse-like portion in response to the side information.
29. Decoder in accordance with one of claims 25 to 26, in which the side information
indicates that an impulse-like signal has been attenuated by an attenuation factor
before being subtracted from the audio signal, and
in which the combiner (34) is operative to attenuate the decoded residual signal
based on the attenuation factor and to use the attenuated decoded signal for a com-
bination with the decoded impulse-like signal.
30. Decoder in accordance with one of claims 25 to 29. in which the encoded impulse-
like signal comprises an impulse-train like signal, and
in which the decoder for decoding (30) the encoded impulse-like signal is operative
to use a decoding algorithm adapted to a coding algorithm, wherein the coding al-
gorithm is adapted for encoding an impulse-train like signal with higher efficiencor less encoding error than a non-impulse-train like signal.
31. Method of decoding an encoded audio signal (24) having an encoded impulse-like
signal (12) and an encoded residual signal (20), comprising:
decoding (30) the encoded impulse-like signal using a decoding algorithm adapted
to a coding algorithm used for generating the encoded impulse-like signal, wherein
a decoded impulse-like signal is obtained;
decoding (32) the encoded residual signal (20) using a decoding algorithm adapted
to a coding algorithm used for generating the encoded residual signal (20), wherein
a decoded residual signal is obtained; and
combining (34) the decoded impulse-like signal and the decoded residual signal to
provide a decoded output signal (36), wherein the steps of decoding (32. 30) are
operative to provide output values related to the same time instant of a decoded
signal,
wherein, in the step of decoding (30) the encoded impulse-like signal, the encoded
impulse-like signal is received and the decoded impulse-like signal is provided at
specified time portions (142) separated by periods (143) in which the step of decod-
ing (32) the encoded residual signal provides the decoded residual signal and the
step of decoding (30) the encoded impulse-like signal does not provide the decoded
impulse-like signal, so that the decoded output signal has the periods (143). in
which the decoded output signal is identical to the decoded residual signal and the
decoded output signal has the specified time portions (142) in which the decoded
output signal consists of the decoded residual signal and the decoded impulse-like
signal or consists of the impulse-like signal only.
32. Encoded audio signal (24) having an encoded impulse-like signal (12), an encoded
residual signal (20), and side information (14) indicating information relating to an
encoding or decoding characteristic pertinent to the encoded residual signal or the
encoded impulse-like signal, wherein the encoded impulse-like signal represents
specified time portions (142) of the audio signal, in which the audio signal is
represented by the encoded impulse-like signal only or is represented by the en-
coded residual signal and the encoded impulse-like signal, the specified time por-
tions (142) being separated by periods (143), in which the audio signal is only
represented by the encoded residual signal and not by the encoded impulse-like
signal (12).
33. Computer program having a program code adapted for performing the method of
claim 24 or 3 I. when running on a processor.

An audio encoder for encoding an audio signal includes an impulse extractor (10) for extracting an impulse-like portion from the audio signal. This impulse-like portion is encoded and forwarded to an output interface (22). Furthermore, the audio encoder includes a signal encoder (16) which encodes a residual signal derived from the original audio signal so that the impulse-like portion is reduced or eliminated in the residual audio signal. The output interface (22) forwards both, the encoded signals, i.e., the encoded impulse signal (12) and the encoded residual signal (20) for transmission or storage. On the decoder-side, both signal portions are separately decoded and then combined to obtain a decoded audio signal.