Optimized scale factor for frequency band extension in an audio frequency signal
decoder
The present invention relates to the field of the coding/decoding and the processing
5 of audio frequency signals (such as speech, music or other such signals) for their
transmission or their storage.
More particularly, the invention relates to a method and a device for determining an
optimized scale factor that can be used to adjust the level of an excitation signal or, in an
equivalent manner, of a filter as part of a frequency band extension in a decoder or a
10 processor enhancing an audio frequency signal.
Numerous techniques exist for compressing (with loss) an audio frequency signal
such as speech or music.
The conventional coding methods for the conversational applications are generally
classified as waveform coding (PCM for "Pulse Code Modulation", ADCPM for "Adaptive
15 Differential Pulse Code Modulation", transform coding, etc.), parametric coding (LPC for
"Linear Predictive Coding", sinusoidal coding, etc.) and parametric hybrid coding with a
quantization of the parameters by "analysis by synthesis" of which CELP ("Code Excited
Linear Prediction") coding is the best known example.
For the non-conversational applications, the prior art for (mono) audio signal coding
20 consists of perceptual coding by transform or in subbands, with a parametric coding of the
high frequencies by band replication.
A review of the conventional speech and audio coding methods can be found in the works by
W.B. Kleijn and K.K. Paliwal (eds.), Speech Coding and Synthesis, Elsevier, 1995; M. Bosi,
R.E. Goldberg, Introduction to Digital Audio Coding and Standards, Springer 2002; J. Benesty,
25 M.M. Sondhi, Y. Huang (Eds.), Handbook of Speech Processing, Springer 2008.
The focus here is more particularly on the 3GPP standardized AMR-WB ("Adaptive
Multi-Rate Wideband") codec (coder and decoder), which operates at an input/output
frequency of 16 kHz and in which the signal is divided into two subbands, the low band (0-
30 6.4 kHz) which is sampled at 12.8 kHz and coded by CELP model and the high band (6.4-
7 kHz) which is reconstructed parametrically by "band extension" (or BWE, for "Bandwidth
Extension") with or without additional information depending on the mode of the current
frame. It can be noted here that the limitation of the coded band of the AMR-WB codec at
7 kHz is essentially linked to the fact that the frequency response in transmission of the
35 wideband terminals was approximated at the time of standardization (ETSI/3GPP then rru-T)
according to the frequency mask defined in the standard ITU-T P.341 and more specifically
by using a so-called "P341" filter defined in the standard ITU-T G.191 which cuts the
frequencies above 7kHz (this filter observes the mask defined in P.341). However, in theory,
wo 2015/004373 2 PCT/FR2014/051720
it is well known that a signal sampled at 16 kHz can have a defined audio band from 0 to
8000 Hz; the AMR-WB codec therefore introduces a limitation of the high band by comparison
with the theoretical bandwidth of 8 kHz.
The 3GPP AMR-WB speech codec was standardized in 2001 mainly for the circuit
5 mode (CS) telephony applications on GSM (2G) and UMTS (3G). This same codec was also
standardized in 2003 by the ITU-T in the form of recommendation G.722.2 "Wideband coding
speech at around 16 kbit/s using Adaptive Multi-Rate Wideband (AMR-WB)".
It comprises nine bit rates, called modes, from 6.6 to 23.85 kbit/s, and comprises
continuous transmission mechanisms (DTX, for "Discontinuous Transmission") with voice
10 activity detection (VAD) and comfort noise generation (CNG) from silence description frames
(SID, for "Silence Insertion Descriptor"), and lost frame correction mechanisms (FEC for
"Frame Erasure Concealment", sometimes called PLC, for "Packet Loss Concealment").
The details of the AMR-WB coding and decoding algorithm are not repeated here; a
detailed description of this codec can be found in the 3GPP specifications (TS 26.190, 26.191,
15 26.192, 26.193, 26.194, 26.204) and in ITU-T-G.722.2 (and the corresponding annexes and
appendix) and in the article by B. Bessette et al. entitled "The adaptive multirate wideband
speech codec (AMR-WB)", IEEE Transactions on Speech and Audio Processing, vol. 10, no. 8,
2002, pp. 620-636 and the source code of the associated 3GPP and ITU-T standards.
The principle of band extension in the AMR-WB codec is fairly rudimentary. Indeed,
>o the high band (6.4-7 kHz) is generated by shaping a white noise through a time (applied in
the form of gains per subframe) and frequency (by the application of a linear prediction
synthesis filter or LPC, for "Linear Predictive Coding") envelope. This band extension
technique is illustrated in figure 1.
A white noise u HBI (n) , n = 0, · · ·, 79 is generated at 16 kHz for each 5 ms subframe
>5 by linear congruential generator (block 100). This noise u HBI (n) is formatted in time by
application of gains for each subframe; this operation is broken down into two processing
steps (blocks 102, 106 or 109):
• A first factor is computed (block 101) to set the white noise u HBI (n) (block 102) at a
level similar to that of the excitation, u(n) , n =0, .. ·,63, decoded at 12.8 kHz in
30 the low band:
63
L;u(l)'
!/H B2 (11) = !/H BI (11) j-,7;;;;:='=0'----LUHBI({)
2
f,_O
wo 2015/004373 3 PCT/FR2014/051720
It can be noted here that the normalization of the energies is done by comparing
blocks of different size (64 for u(n) and 80 for uHBI (n)) without compensation of the
differences in sampling frequencies (12.8 or 16kHz).
• The excitation in the high band is then obtained (block 106 or 109) in the form:
5 liHB(11) = gHB!IHB2(1l)
10
15
20
in which the gain g HB is obtained differently depending on the bit rate. If the bit rate
of the current frame is < 23.85 kbit/s, the gain gH8 is estimated "blind" (that is to
say without additional information); in this case, the block 103 filters the signal
decoded in low band by a high-pass filter having a cut-off frequency at 400 Hz to
obtain a signal s,P (11) 1 II= 0, · · ·, 63 - this high-pass filter eliminates the influence of
the very low frequencies which can skew the estimation made in the block 104 -then
the "tilt" (indicator of spectral slope) denoted e,~< of the signal s,/n) is computed
by normalized self-correlation (block 104):
63 2>,p (n)s,p (n-1)
n=! ealt = =----:c,,,-------
2),P (n )2
n=O
and finally, gH8 is computed in the form:
gHB = )\ISPgSP + (1- Wsp)gBG
in which g8P = 1-e,u, is the gain applied in the active speech (SP) frames,
g8G = 1.25g8P is the gain applied in the inactive speech frames associated with a
background (BG) noise and w8P is a weighting function which depends on the voice
activity detection (VAD). It is understood that the estimation of the tilt ( e,11,) makes it
possible to adapt the level of the high band as a function of the spectral nature of the
signal; this estimation is particularly important when the spectral slope of the CELP
decoded signal is such that the average energy decreases when the frequency
increases (case of a voiced signal where e,~< is close to 1, therefore g8P = 1-e,, is
thus reduced). It should also be noted that the factor gH8 in the AMR-WB decoding
is bounded to take values within the range [0.1, 1.0]. Indeed, for the signals whose
energy increases when the frequency increases ( e,, close to -1, g8P close to 2), the
gain g HB is usually underestimated.
wo 20151004373 4 PCTIFR2014I051720
At 23.85 kbitls, a correction information item is transmitted by the AMR-WB coder
and decoded (blocks 107, 108) in order to refine the gain estimated for each subframe ( 4 bits
every 5 ms, or 0.8 kbitls). The artificial excitation u HB (n) is then filtered (block 111) by
an LPC synthesis filter (block 111) of transfer function 11 AHn ( z) and operating at the
5 sampling frequency of 16 kHz. The construction of this filter depends on the bit rate of the
current frame:
• At 6.6 kbitls, the filter 1/ AH8 (z) is obtained by weighting by a factor r= 0.9 an
LPC filter of order 20, 1/ A'-"(z), which "extrapolates" the LPC filter of order 16,
1 I A(z), decoded in the low band (at 12.8 kHz)- the details of the extrapolation in
20 the realm of the ISF (Imittance Spectral Frequency) parameters are described in the
standard G.722.2 in section 6.3.2.1; in this case,
25
11 AHB(z)=ll A"'1(zly)
• at the bit rates > 6.6 kbitls, the filter 11 AHn (z) is of order 16 and corresponds
simply to:
1 I AHB(z) = 11 A(z I r)
in which r = 0.6. It should be noted that, in this case, the filter 11 A(z I y) is used
at 16 kHz, which results in a spreading (by proportional transformation) of the
frequency response of this filter from [0, 6.4 kHz] to [0, 8 kHz].
The result, sH8(n), is finally processed by a bandpass filter (block 112) of FIR ("Finite
20 Impulse Response") type, to keep only the 6 - 7 kHz band; at 23.85 kbitls, a low-pass filter
also of FIR type (block 113) is added to the processing to further attenuate the frequencies
above 7 kHz. The high frequency (HF) synthesis is finally added (block 130) to the low
frequency (LF) synthesis obtained with the blocks 120 to 122 and re-sampled at 16 kHz
(block 123). Thus, even if the high band extends in theory from 6.4 to 7 kHz in the AMR-WB
25 codec, the HF synthesis is rather contained in the 6-7 kHz band before addition with the LF
synthesis.
A number of drawbacks in the band extension technique of the AMR-WB codec can
be identified, in particular:
• the estimation of gains for each subframe (block 101, 103 to 105) is not optimal.
30 Partly, it is based on an equalization of the "absolute" energy per subframe (block
101) between signals at different frequencies: artificial excitation at 16 kHz (white
noise) and a signal at 12.8 kHz (decoded ACELP excitation). It can be noted in
particular that this approach implicitly induces an attenuation of the high-band
excitation (by a ratio 12.8116 = 0.8); in fact, it will also be noted no de-emphasis is
35 performed on the high band in the AMR-WB codec, which implicitly induces an
wo 2015/004373 5 PCT/FR2014/051720
amplification relatively close to 0.6 (which corresponds to the value of the frequency
response of 1 I ( 1-0.68z ,) at 6400 Hz). In fact, the factors of 1/0.8 and of 0.6 are
compensated approximately.
• Regarding speech, the 3GPP AMR-WB codec characterization tests documented in the
5 3GPP report TR 26.976 have shown that the mode at 23.85 kbit/s has a less good
quality than at 23.05 kbit/s, its quality being in fact similar to that of the mode at
15.85 kbit/s. This shows in particular that the level of artificial HF signal has to be
controlled very prudently, because the quality is degraded at 23.85 kbit/s whereas
the 4 bits per frame are considered to best make it possible to approximate the
10 energy of the original high frequencies.
• The low-pass filter at 7 kHz (block 113) introduces a shift of almost 1 ms between the
low and high bands, which can potentially degrade the quality of certain signals by
slightly desynchronizing the two bands at 23.85 kbit/s - this desynchronization can
also pose problems when switching bit rate from 23.85 kbit/s to other modes.
An example of band extension via a temporal approach is described in the 3GPP standard TS
26.290 describing the AMR-WB+ codec (standardized in 2005). This example is illustrated in
the block diagrams of figures 2a (general block diagram) and 2b (gain prediction by response
level correction) which correspond respectively to figures 16 and 10 of the 3GPP specification
20 TS 26.290.
In the AMR-WB+ codec, the (mono) input signal sampled at the frequency Fs (in Hz) is
divided into two separate frequency bands, in which two LPC filters are computed and coded
separately:
• one LPC filter, denoted A(z), in the low band (0-Fs/4) - its quantized version is
25 denoted A(z)
• another LPC filter, denoted AHF(z), in the spectrally aliased high band (Fs/4-Fs/2)-
its quantized version is denoted AHF (z)
The band extension is done in the AMR-WB+ codec as detailed in sections 5.4 (HF coding)
and 6.2 (HF decoding) of the 3GPP specification TS 26.290. The principle thereof is
30 summarized here: the extension consists in using the excitation decoded at low frequencies
(LFC excit.) and in formatting this excitation by a temporal gain per subframe (block 205) and
an LPC synthesis filtering (block 207); the processing operations to enhance (postprocessing)
the excitation (block 206) and smooth the energy of the reconstructed HF signal
(block 208) are moreover implemented as illustrated in figure 2a.
35 It is important to note that this extension in AMR-WB+ necessitates the transmission of
additional information: the coefficients of the filter AHF(z) in 204 and a temporal formatting
.I
~
5
wo 2015/004373 6 PCT/FR2014/051720
gain per subframe (block 201). One particular feature of the band extension algorithm in
AMR-WB+ is that the gain per subframe is quantified by a predictive approach; in other
words, the gains are not coded directly, but rather gain corrections which are relative to an
estimation of the gain denoted g"'""". This estimation, g"'""", actually corresponds to a level
equalization factor between the filters A(z) and AHF(z) at the frequency of separation
between low band and high band (Fs/4). The computation of the factor g"'""" (block 203) is
detailed in figure 10 of the 3GPP specification TS 26.290 reproduced here in figure 2b. This
figure will not be detailed more here. It will simply be noted that the blocks 210 to 213 are
used to compute the energy of the impulse response of
A(z)
_
1
, , while recalling
(1-0.9z )AHF(z)
10 that the filter AHF (z) models a spectrally alia sed high band (because of the spectral
properties of the filter bank separating the low and high bands). Since the filters are
interpolated by subframes, the gain g "'""" is computed only once per frame, and it is
interpolated by subframes.
The band extension gain coding technique in AMR-WB+, and more particularly the
15 compensation of levels of the LPC filters at their junction is an appropriate method in the
context of a band extension by LPC models in low and high band, and it can be noted that
such a level compensation between LPC filters is not present in the band extension of the
AMR-WB codec. However, it is in practice possible to verify that the direct equalization of the
level between the two LPC filters at the separation frequency is not an optimal method and
20 can provoke an overestimation of energy in high band and audible artifacts in certain cases; it
will be recalled that an LPC filter represents a spectral envelope, and the principle of
equalization of the level between two LPC filters for a given frequency amounts to adjusting
the relative level of two LPC envelopes. Now, such an equalization performed at a precise
frequency does not ensure a complete continuity and overall consistency of the energy (in
25 frequency) in the vicinity of the equalization point when the frequency envelope of the signal
fluctuates significantly in this vicinity. A mathematical way of positing the problem consists in
noting that the continuity between two curves can be ensured by forcing them to meet at one
and the same point, but there is nothing to guarantee that the local properties (successive
derivatives) coincide so as to ensure a more global consistency. The risk in ensuring a spot
30 continuity between low and high band LPC envelopes is of setting the LPC envelope in high
band at a relative level that is too strong or too weak, the case of a level that is too strong
being more damaging because it results in more annoying artifacts.
Moreover, the gain compensation in AMR-WB+ is primarily a prediction of the gain known to
the coder and to the decoder and which serves to reduce the bit rate necessary for the
35 transmission of gain information scaling the high-band excitation signal. Now, in the context
wo 2015/004373 7 PCT/FR2014/051720
of an interoperable enhancement of the AMR-WB coding/decoding, it is not possible to modify
the existing coding of the gains by subframes (0.8 kbit/s) of the band extension in the AMRWB
23.85 kbit/s mode. Furthermore, for the bit rates strictly less than 23.85 kbit/s, the
compensation of levels of LPC filters in low and high bands can be applied in the band
5 extension of a decoding compatible with AMR-WB, but experience shows that this sole
technique derived from the AMR-WB+ coding, applied without optimization, can cause
problems of overestimation of energy of the high band (> 6 kHz).
There is therefore a need to improve the compensation of gains between linear prediction
10 filters of different frequency bands for the frequency band extension in a codec of AMR-WB
type or an interoperable version of this codec without in any way overestimating the energy
in a frequency band and without requiring additional information from the coder.
The present invention improves the situation.
To this end, the invention targets a method for determining an optimized scale factor
15 to be applied to an excitation signal or to a filter in an audio frequency signal frequency band
extension method, the band extension method comprising a step of decoding or of extraction,
in a first frequency band, of an excitation signal and of parameters of the first frequency
band comprising coefficients of a linear prediction filter, a step of generation of an extended
excitation signal on at least one second frequency band and a step of filtering, by a linear
20 prediction filter, for the second frequency band. The determination method is such that it
comprises the following steps:
- determination of a linear prediction filter called additional filter, of lower order than the
linear prediction filter of the first frequency band, the coefficients of the additional filter
being obtained from the parameters decoded or extracted from the first frequency band;
25 and
- computation of the optimized scale factor as a function at least of the coefficients of the
additional filter.
Thus, the use of an additional filter of lower order than the filter of the first frequency
band to be equalized makes it possible to avoid the overestimations of energy in the high
30 frequencies which could result from local fluctuations of the envelope and which can disrupt
the equalization of the prediction filters.
The equalization of gains between the linear prediction filters of the first and second
frequency bands is thus enhanced.
In an advantageous application of the duly obtained optimized scale factor, the band
35 extension method comprises a step of application of the optimized scale factor to the
extended excitation signal.
In an appropriate embodiment, the application of the optimized scale factor is
combined with the step of filtering in the second frequency band.
wo 2015/004373 8 PCT/FR2014/051720
Thus, the steps of filtering and of application of the optimized scale factor are
combined in a single filtering step to reduce the processing complexity.
In a particular embodiment, the coefficients of the additional filter are obtained by
truncation of the transfer function of the linear prediction filter of the first frequency band to
5 obtain a lower order.
This lower order additional filter is therefore obtained in a simple manner.
Furthermore, so as to obtain a stable filter, the coefficients of the additional filter are
modified as a function of a stability criterion of the additional filter.
In a particular embodiment, the computation of the optimized scale factor comprises
10 the following steps:
- computation of the frequency responses of the linear prediction filters of the first and
second frequency bands for a common frequency;
- computation of the frequency response of the additional filter for this common
frequency;
15 - computation of the optimized scale factor as a function of the duly computed frequency
responses.
20
Thus, the optimized scale factor is computed in such a way as to avoid the annoying
artifacts which could occur should the higher order filter frequency response of the first band
in proximity to the common frequency show a signal peak or trough.
In a particular embodiment, the method further comprises the following steps,
implemented for a predetermined decoding bit rate:
- first scaling of the extended excitation signal by a gain computed per subframe as a
function of an energy ratio between the decoded excitation signal and the extended
excitation signal;
25 - second scaling of the excitation signal obtained from the first scaling by a decoded
correction gain;
- adjustment of the energy of the excitation for the current subframe by an adjustment
factor computed as a function of the energy of the signal obtained after the second
scaling and as a function of the signal obtained after application of the optimized scale
30 factor.
Thus, additional information can be used to enhance the quality of the extended
signal for a predetermined operating mode.
The invention also targets a device for determining an optimized scale factor to be
35 applied to an excitation signal or to a filter in an audio frequency signal frequency band
extension device, the band extension device comprising a module for decoding or extracting,
in a first frequency band, an excitation signal and parameters of the first frequency band
comprising coefficients of a linear prediction filter, a module for generating an extended
wo 2015/004373 9 PCf/FR2014/051720
excitation signal on at least one second frequency band and a module for filtering, by a linear
prediction filter, for the second frequency band. The determination device is such that it
comprises:
-a module for determining a linear prediction filter called additional filter, of lower order
5 than the linear prediction filter of the first frequency band, the coefficients of the
additional filter being obtained from the parameters decoded or extracted from the first
frequency band; and
10
- a module for computing the optimized scale factor as a function at least of the
coefficients of the additional filter.
The invention targets a decoder comprising a device as described.
It targets a computer program comprising code instructions for implementing the
steps of the method for determining an optimized scale factor as described, when these
instructions are executed by a processor.
Finally, the invention relates to a storage medium, that can be read by a processor,
15 incorporated or not in the device for determining an optimized scale factor, possibly
removable, storing a computer program implementing a method for determining an optimized
scale factor as described previously.
Other features and advantages of the invention will become more clearly apparent on
reading the following description, given purely as a nonlimiting example and with reference to
20 the attached drawings, in which:
25
30
35
figure 1 illustrates a part of a decoder of AMR-WB type implementing frequency
band extension steps of the prior art and as described previously;
figures 2a and 2b present the coding of the high band in the AMR-WB+ codec
according to the prior art and as described previously;
figure 3 illustrates a decoder that can interwork with the AMR-WB coding,
incorporating a band extension device used according to an embodiment of the
invention;
figure 4 illustrates a device for determining a scale factor optimized by a
subframe as a function of the bit rate, according to an embodiment of the
invention; and
figures Sa and Sb illustrate the frequency responses of the filters used for the
computation of the optimized scale factor according to an embodiment of the
invention;
figure 6 illustrates, in flow diagram form, the main steps of a method for
determining an optimized scale factor according to an embodiment of the
invention;
figure 7 illustrates an embodiment in the frequency domain of a device for
determining an optimized scale factor as part of a band extension;
wo 2015/004373 10 PCT/FR2014/051720
figure 8 illustrates a hardware implementation of an optimized scale factor
determination device in a band extension according to the invention.
Figure 3 illustrates an exemplary decoder, compatible with the AMR-WB/G.722.2
5 standard in which there is a band extension comprising a determination of an optimized scale
factor according to an embodiment of the method of the invention, implemented by the band
extension device illustrated by the block 309.
Unlike the AMR-WB decoding which operates with an output sampling frequency of
16 kHz, a decoder is considered here which can operate with an output signal (synthesis) at
10 the frequency fs = 8, 16, 32 or 48 kHz. It should be noted that it is assumed here that the
coding has been performed according to the AMR-WB algorithm with an internal frequency of
12.8 kHz for the CELP coding in low band and at 23.85 kbit/s a gain coding per subframe at
the frequency of 16 kHz; even though the invention is described here at the decoding level, it
is assumed here that the coding can also operate with an input signal at the frequency fs
15 = 8, 16, 32 or 48 kHz and suitable resampling operations, beyond the context of the
invention, are implemented in coding as a function of the value of fs. It can be noted that,
when fs = 8 kHz, in the case of a decoding compatible with AMR-WB, it is not necessary to
extend the 0-6.4 kHz low band, because the audio band reconstructed at the frequency fs is
limited to 0-4000 Hz.
20 In figure 3, the CELP decoding (LF for low frequencies) still operates at the internal
frequency of 12.8 kHz, as in AMR-WB, and the band extension (HF for high frequencies) used
for the invention operates at the frequency of 16 kHz, and the LF and HF syntheses are
combined (block 312) at the frequency fs after suitable resampling (block 306 and internal
processing in the block 311). In the variant embodiments, the combining of the low and high
25 bands can be done at 16 kHz, after having resampled the low band from 12.8 to 16 kHz,
before resampling the combined signal at the frequency fs.
The decoding according to figure 3 depends on the AMR-WB mode (or bit rate)
associated with the current frame received. As an indication, and without affecting the block
309, the decoding of the CELP part in low band comprises the following steps:
30 • demultiplexing of the coded parameters (block 300) in the case of a frame correctly
received (bfi=O where bfi is the "bad frame indicatol' with a value 0 for a frame
received and 1 for a frame lost);
• decoding of the !SF parameters with interpolation and conversion into LPC
coefficients (block 301) as described in clause 6.1 of the standard G.722.2;
35 • decoding of the CELP excitation (block 302), with an adaptive and fixed part for
reconstructing the excitation ( exc or u'(n) ) in each subframe of length 64 at 12.8
kHz:
u'(n)=gPv(n)+g,c(n), n=0, .. ·,63
5
wo 2015/004373 11 PCf/FR2014/051720
by following the notations of clause 7.1.2.1 of ITU-T recommendation G.718 of a
decoder interoperable with the AMR-WB coder/decoder, concerning the CELP
decoding, where v(n) and c(n) are respectively the code words of the adaptive and
fixed dictionaries, and g P and ftc are the associated decoded gains. This excitation
u '(n) is used in the adaptive dictionary of the next subframe; it is then postprocessed
and, as in G.718, the excitation u'(n) (also denoted exc) is distinguished
from its modified post-processed version u(n) (also denoted exc2) which serves as
input for the synthesis filter, 1/ A(z), in the block 303;
• synthesis filtering by II A(z) (block 303) where the decoded LPC filter A(z) is of
10 the order 16;
• narrow-band post-processing (block 304) according to clause 7.3 of G.718 if fs=8
kHz;
• de-emphasis (block 305) by the filter 1/ ( 1-0.68z ,);
• post-processing of the low frequencies (called "bass posfi!ter") (block 306)
15 attenuating the cross-harmonics noise at low frequencies as described in
clause 7.14.1.1 of G.718. This processing introduces a delay which is taken into
account in the decoding of the high band (> 6.4 kHz);
20
25
• resampling of the internal frequency of 12.8 kHz at the output frequency fs (block
307). A number of embodiments are possible. Without losing generality, it is
considered here, by way of example, that if fs=8 or 16 kHz, the resampling
described in clause 7.6 of G.718 is repeated here, and if fs=32 or 48kHz, additional
finite impulse response (FIR) filters are used;
• computation of the parameters of the "noise gate" (block 308) preferentially
performed as described in clause 7.14.3 of G.718 to "enhance" the quality of the
silences by level reduction.
In variants which can be implemented for the invention, the post-processing operations
applied to the excitation can be modified (for example, the phase dispersion can be
enhanced) or these post-processing operations can be extended (for example, a reduction of
the cross-harmonics noise can be implemented), without affecting the nature of the band
30 extension.
It can be noted that the use of blocks 306, 308, 314 is optional.
It will also be noted that the decoding of the low band described above assumes a so-called
"active" current frame with a bit rate between 6.6 and 23.85 kbit/s. In fact, when the DTX
35 mode is activated, certain frames can be coded as "inactive" and in this case it is possible to
either transmit a silence descriptor (on 35 bits) or transmit nothing. In particular, it will be
wo 2015/004373 12 PCT/FR2014/051720
recalled that the SID frame describes a number of parameters: !SF parameters averaged over
8 frames, average energy over 8 frames, "dithering" flag for the reconstruction of nonstationary
noise. In all cases, in the decoder, there is the same decoding model as for an
active frame, with a reconstruction of the excitation and of an LPC filter for the current frame,
5 which makes it possible to apply the band extension even to inactive frames. The same
observation applies for the decoding of "lost frames" (or FEC, PLC) in which the LPC model is
applied.
In the embodiment described here and with reference to figure 7, the decoder makes
it possible to extend the decoded low band (50-6400 Hz taking into account the 50 Hz high-
10 pass filtering on the decoder, 0-6400 Hz in the general case) to an extended band, the width
of which varies, ranging approximately from 50-6900Hz to 50-7700Hz depending on the
mode implemented in the current frame. It is thus possible to refer to a first frequency band
of 0 to 6400 Hz and to a second frequency band of 6400 to 8000 Hz. In reality, in the
preferred embodiment, the extension of the excitation is performed in the frequency domain
15 in a 5000 to 8000 Hz band, to allow a bandpass filtering of 6000 to 6900 or 7700 Hz width.
At 23.85 kbit/s, the HF gain correction information (0.8 kbit/s) transmitted at 23.85
kbit/s is here decoded. Its use is detailed later, with reference to figure 4. The highband
synthesis part is produced in the block 309 representing the band extension device used
for the invention and which is detailed in figure 7 in an embodiment.
20 In order to align the decoded low and high bands, a delay (block 310) is introduced
to synchronize the outputs of the blocks 306 and 307 and the high band synthesized at 16
kHz is resampled from 16kHz to the frequency fs (output of block 311). The value of the
delay T depends on how the high band signal is synthesized, and on the frequency fs as in
the post-processing of the low frequencies. Thus, generally, the value of Tin the block 310
25 will have to be adjusted according to the specific implementation.
The low and high bands are then combined (added) in the block 312 and the
synthesis obtained is post-processed by 50 Hz high-pass filtering (of IIR type) of order 2, the
coefficients of which depend on the frequency fs (block 313) and output post-processing with
optional application of the "noise gate" in a manner similar to G.718 (block 314).
Referring to figure 3, an embodiment of a device for determining an optimized scale
factor to be applied to an excitation signal in a frequency band extension process is now
described. This device is included in the band extension block 309 described previously.
Thus, the block 400, from an excitation signal decoded in a first frequency band
u(n) , performs a band extension to obtain an extended excitation signal u Hn (n) on at least
35 one second frequency band.
It will be noted here that the optimized scale factor estimation according to the
invention is independent of how the signal uH8(n) is obtained. One condition concerning its
energy is, however, important. Indeed, the energy of the high band from 6000 to 8000 Hz
wo 2015/004373 13 PCT/FR2014/051720
must be at a level similar to the energy of the band from 4000 to 6000 Hz of the decoded
excitation signal at the output of the block 302. Furthermore, since the low-band signal is deemphasized
(block 305), the de-emphasis must also be applied to the high-band excitation
signal, either by using a specific de-emphasis filter, or by multiplying by a constant factor
5 which corresponds to an average attenuation of the filter mentioned. This condition does not
apply to the case of the 23.85 kbit/s bit rate which uses the additional information
transmitted by the coder. In this case, the energy of the high-band excitation signal must be
consistent with the energy of the signal corresponding to the coder, as explained later.
The frequency band extension can, for example, be implemented in the same way as
10 for the decoder of AMR-WB type described with reference to figure 1 in the blocks 100 to
102, from a white noise.
15
In another embodiment, this band extension can be performed from a combination of
a white noise and of a decoded excitation signal as illustrated and described later for the
blocks 700 to 707 in figure 7.
Other frequency band extension methods with conservation of the energy level
between the decoded excitation signal and the extended excitation signal as described below,
can of course be envisaged for the block 400.
Furthermore, the band extension module can also be independent of the decoder and
can perform a band extension for an existing audio signal stored or transmitted to the
20 extension module, with an analysis of the audio signal to extract an excitation and an LPC
filter therefrom. In this case, the excitation signal at the input of the extension module is no
longer a decoded signal but a signal extracted after analysis, like the coefficients of the linear
prediction filter of the first frequency band used in the method for determining the optimized
scale factor in an implementation of the invention.
25 In the example illustrated in figure 4, the case of the bit rates < 23.85 kbit/s, for
which the determination of the optimized scale factor is limited to the block 401, is
considered first.
In this case, an optimized scale factor denoted g"82 (m) is computed. In one embodiment,
this computation is performed preferentially for each subframe and it consists in equalizing
30 the levels of the frequency responses of the LPC filters I I A(z) and I I A(zl y) used in low
and high frequencies, as described later with reference to figure 7, with additional
precautions to avoid the cases of overestimations which can result in an excessive energy of
the synthesized high band and therefore generate audible artifacts.
In an alternative embodiment, it will be possible to keep the extrapolated HF synthesis filter
35 I I A a' (zl y) as implemented in the AMR-WB decoder or a decoder that can interwork with
the AMR-WB coder/decoder, for example according to the ITU-T recommendation G.718, in
5
10
wo 2015/004373 14 PCT/FR2014/051720
place of the filter 1/ A(zl y). The compensation according to the invention is then performed
from the filters 1/ A(z) and 1/ A'" (zl y).
The determination of the optimized scale factor is also performed by the determination (in
401a) of a linear prediction filter called additional filter, of lower order than the linear
prediction filter of the first frequency band 11 A(z), the coefficients of the additional filter
being obtained from the parameters decoded or extracted from the first frequency band. The
optimized scale factor is then computed (in 401b) as a function at least of these coefficients
to be applied to the extended excitation signal u HB (n) .
The principle of the determination of the optimized scale factor, implemented in the
block 401, is illustrated in figures Sa and Sb with concrete examples obtained from signals
sampled at 16 kHz; the frequency response amplitude values, denoted R, P, Q below, of 3
filters are computed at the common frequency of 6000 Hz (vertical dotted line) in the current
subframe, of which the index m is not recalled here in the notations of the LPC filters
15 interpolated by subframe to lighten the text. The value of 6000 Hz is chosen such that it is
close to the Nyquist frequency of the low band, that is 6400 Hz. It is preferable not to take
this Nyquist frequency to determine the optimized scale factor. Indeed, the energy of the
decoded signal in low frequencies is typically already attenuated at 6400 Hz. Furthermore,
the band extension described here is performed on a second frequency band, called high
20 band, which ranges from 6000 to 8000 Hz. It should be noted that, in variants of the
invention, a frequency other than 6000 Hz will be able to be chosen, with no loss of
generality for determining the optimized scale factor. It will also be possible to consider the
case where the two LPC filters are defined for the separate bands (as in AMR-WB+). In this
case, R, P and Q will be computed at the separation frequency.
25 Figures Sa and Sb illustrate how the quantities R, P, Q are defined.
The first step consists in computing the frequency responses R and P respectively of the
linear prediction filter of the first frequency band (low band) and of the second frequency
band (high band) at the frequency of 6000Hz. The following is first computed:
R 1 1
'A
A(eio)j M
'L"..'J a.... ;e -jiO
;~o
30 in which M = 16 is the order of the decoded LPC filter, 1/ A(z), and Bcorresponds to the
frequency of 6000 Hz normalized for the sampling frequency of 12.8 kHz, that is:
B= 2tr 6000.
12800
Then, similarly, the following is computed:
wo 2015/004373
in which
e• = 2" 6ooo .
16000
15 PCT/FR2014/0517ZO
1
In a preferred embodiment, the quantities P and R are computed according to the
5 following pseudo-code:
px=py=O
rx=ry=O
fori=O to 16
px = px + Ap[i]*exp_tab_p[i]
10 py = py + Ap[i]*exp_tab_p[33-i]
rx = rx + Aq[i]*exp_tab_q[i]
ry = ry + Aq[i]*exp_tab_q[33-i]
end for
P = 1/sqrt(px*px+py*py)
15 R = 1/sqrt{rx*rx+ry*ry)
in which Aq[i]= a, corresponds to the coefficients of A(z) (of order 16), Ap[i]= r'a,
20
corresponds to the coefficient of A(zi y), sqrt() corresponds to the square root operation
and the tables exp_tab_p and exp_tab_q of size 34 contain the real and imaginary parts of
the complex exponentials associated with the frequency of 6000 Hz, with
exp_tab_p[i] =
exp_tab_q[i] =
(
cos 2tr60-0-0 l· )
12800
-sin(2tr
6000
(33-i))
12800
(
cos 2tr60-0-01 · )
16000
-sin(2tr
6000
(33-i))
16000
i=0, .. ·,16
i=17, .. ·,33
i = o, .. ·,16
i=17, .. ·,33
25 The additional prediction filter is obtained for example by suitably truncating the polynomial
A(z) to the order Z.
wo 2015/004373 16 PCT/FR2014/051720
In fact, the direct truncation to the order leads to the filter 1 + a1 +a,, which can pose a
problem because there is generally nothing to guarantee that this filter of order 2 is stable. In
a preferred embodiment, the stability of the filter 1 +a! +a, is therefore detected and a filter
1 + al '+a, ' is used, the coefficients of which are drawn from 1 + al +a, as a function of the
5 instability detection. More specifically, the following are initialized:
aA , ' ::::Aa ,, I' = 1, 2
The stability of the filter 1 + a1 +a, can be verified differently; here, a conversion is used in
the PARCOR coefficients (or reflection coefficients) domain by computing:
kl =a! 'f(l+a, ')
k A ' 10 2 = a2
15
The stability is verified if jk,j <1, i=1, 2. The value of k, is therefore conditionally modified
before ensuring the stability of the filter, with the following steps:
{
min(0.6,k2 )
k, fmax(-
0.6,k,)
{
min(0.99, k,)
kl fmax(-
0.99,k,)
k, > 0
k, <0
kl >0
kl <0
in which min{.,.) and max{.,.) respectively give the minimum and the maximum of 2
operands.
It should be noted that the threshold values, 0.99 for k1 and 0.6 for k2 , will be able to be
adjusted in variants of the invention. It will be recalled that the first reflection coefficient, k1 ,
20 characterizes the spectral slope (or tilt) of the signal modeled to the order 1; in the invention
the value of k1 is saturated at a value close to the stability limit, in order to preserve this
slope and retain a tilt similar to that of 1 I A(z) . It will also be recalled that the second
reflection coefficient, k2 , characterizes the resonance level of the signal modeled to the order
2; since the use of a filter of order 2 aims to eliminate the influence of such resonances
25 around the frequency of 6000 Hz, the value of k, is more strongly limited; this limit is set at
0.6.
The coefficients of 1 + a1 '+a, ' are then obtained by:
30 The frequency response of the additional filter is therefore finally computed:
:I wo 2015/004373 17 PCT/FR2014/051720
Q
I
2
"L ..a"'k 1 e -jkO
k=O
6000
with e = 2tr--. This quantity is computed preferentially according to the following
12800
pseudo-code:
qx=qy=O
5 for i=O to 2
10
qx = qx + As[i]*exp_tab_q[i];
qy = qy + As[i]*exp_tab_q{33-i];
end for
Q = 1/sqrt(qx*qx+qy*qy)
in which As[i] = a, '.
With no loss of generality, it will be possible to compute the coefficients of the filter of order
2 otherwise, for example by applying to the LPC filter A(z) of order 16 the reduction
procedure of the LPC order called "STEP DOWN" described in J.D. Markel and A.H. Gray,
15 Linear Prediction of Speech, Springer Verlag, 1976 or by performing two Levinson-Durbin (or
STEP-UP) algorithm iterations from the self-correlations computed on the signal synthesized
(decoded) at 12.8 kHz and windowed.
For some signals, the quantity Q, computed from the first 3 LPC coefficients decoded, better
takes account of the influence of the spectral slope (or tilt) in the spectrum and avoids the
20 influence of "spurious" peaks or troughs close to 6000 Hz which can skew or raise the value
of the quantity R, computed from all the LPC coefficients.
In a preferred embodiment, the optimized scale factor is deduced from the pre-computed
quantities R, P, Q conditionally, as follows:
If the tilt (computed as in AMR-WB in the block 104, by normalized self-correlation in the
form r(1)/r{O) in which r(i) is the self-correlation) is negative (tilt< 0 as represented in figure
5b ), the computation of the scale factor is done as follows:
to avoid artifacts due to excessively abrupt variations of energy of the high band, a
30 smoothing is applied to the value of R . In a preferred embodiment, an exponential
smoothing is performed with a fixed factor in time (0.5) in the form of:
R = O.SR + O.SRP""
wo 2015/004373 18 PCT/FR2014/051720
R -R prev
in which RP"" corresponds to the value of R in the preceding subframe and the factor 0.5 is
optimized empirically- obviously, the factor 0.5 will be able to be changed for another value
and other smoothing methods are also possible. It should be noted that the smoothing makes
5 it possible to reduce the temporal variants and therefore avoid artifacts.
10
The optimized scale factor is then given by:
gH82(m) = max(min(R, Q), P) I P
In an alternative embodiment, it will be possible to replace the smoothing of R with a
smoothing of g HBz (m) such that:
gHBZ (111) +--- Q.5gHBZ (111) + Q.5gHBZ (111-1)
If the tilt (computed as in AMR-WB in the block 104) is positive (tilt> 0 as in figure Sa), the
15 computation of the scale factor is done as follows:
20
the quantity R is smoothed adaptively in time, with a stronger smoothing when R is
low- as in the preceding case, this smoothing makes it possible to reduce the temporal
variants and therefore avoids artifacts:
R = (1-a)R+aRpm· with a= 1-R2
Rpm· -R
Then, the optimized scale factor is given by:
gHB2(m)=min(R,P,Q)! P
In an alternative embodiment, it will be possible to replace the smoothing of R with a
25 smoothing of gH82 (m) as computed above.
30
where g HB ( -1) is the scale or gain factor computed for the last subframe of the preceding
frame.
The minimum of R, P, Q is taken here in order to avoid overestimating the scale factor.
In a variant, the above condition depending only on the tilt will be able to be extended to
take account not only of the tilt parameter but also of other parameters in order to refine the
decision. Furthermore, the computation of gH82 (m) will be able to be adjusted according to
these said additional parameters.
5
wo 2015/004373 19 PCT/FR2014/051720
An example of additional parameter is the number of zero crossings (ZCR, zero crossing rate)
which can be defined as:
1 N-1
zc1~ =-:Lisgn[s(n)]-sgn[s(n-1)]1
2 IJool
in which
sgn(x) = {
1
-1
ifx'20
if X <0
The parameter zcr generally gives results similar to the tilt. A good classification criterion is
the ratio between zc1~ computed for the synthesized signal s(n) and zc1;, computed for the
excitation signal u(n) at 12 800 Hz. This ratio is between 0 and 1, where 0 means that the
signal has a decreasing spectrum, 1 that the spectrum is increasing (which corresponds to
10 (1- tilt) 12. In this case, a ratio zc1~ I zc1;, > 0.5 corresponds to the case tilt < 0, a ratio
zc1~ I zc1;, < 0.5 corresponds to tilt > 0.
In a variant, it will be possible to use a function of a parameter tilt,P where tilt,P is the tilt
computed for the synthesized signal s(n) filtered by a high-pass filter with a cut-off
frequency for example at 4800Hz; in this case, the response 1 I A(zl y) from 6 to 8kHz
15 (applied at 16kHz) corresponds to the weighted response of 1/ A(z) from 4.8 to 6.4 kHz.
Since 1 I A(zl y) has a more flattened response, it is necessary to compensate this change of
tilt. The scale factor function according to tilt,P is then given in an embodiment by:
(l-tilt,P)' +0.6. Q and R are therefore multiplied by min(l,(1-tilt,P)' +0.6) when
tilt >0 or by max ( 1,(1- tilt,P )' + 0.6) when tilt <0.
20 The case of the 23.85 kbit/s bit rate is now considered, for which a gain correction is
performed by the blocks 403 to 408. This gain correction could moreover be the subject of a
separate invention. In this particular embodiment according to the invention, the gain
correction information, denoted g11"'0"(m), transmitted by the AMR-WB (compatible) coding
with a bit rate of 0.8 kbit/s, is used to improve the quality at 23.85 kbit/s.
25 It is assumed here that the AMR-WB (compatible) coding has performed a correction gain
quantization on 4 bits as described in ITU-T clause G.722.2/5.11 or, equivalently, in the 3GPP
clause TS 26.190/5.11.
In the AMR-WB coder, the correction gain is computed by comparing the energy of the
original signal sampled at 16 kHz and filtered by a 6-7 kHz bandpass filter, s 118 (n), with the
30 energy of the white noise at 16 kHz filtered by a synthesis filter 1 I A( z I y) and a 6-7 kHz
wo 2015/004373 20 PCT/FR2014/051720
bandpass filter (before the filtering, the energy of the noise is set to a level similar to that of
the excitation at 12.8 kHz), s HBz (n). The gain is the root of the ratio of energy of the
original signal to the energy of the noise divided by two. In one possible embodiment, it will
be possible to change the bandpass filter for a filter with a wider band (for example from 6 to
s 7.6 kHz).
80(m+l)-l
L sHs(n)z
n-Wm 0 3
80(111+1)-1 I 11l = ' ... ,
L sHsz(n)z
n=80m
To be able to apply the gain information received at 23.85 kbit/s (in the block 407), it is
important to bring the excitation to a level similar to that expected of the AMR-WB
10 (compatible) coding. Thus, the block 404 performs the scaling of the excitation signal
according to the following equation:
u RBI (n) = gH83 (m)u HB(n), n = 80m, .. ·, SO(m + 1) -1
in which g HBJ (m) is a gain per subframe computed in the block 403 in the form:
63
Lu(n)2
n-O
79
5.Ll1Hs(n)z
n=O
15 in which the factor 5 in the denominator serves to compensate the bandwidth difference
between the signal u(n) and the signal uH8 (n), given that, in the AMR-WB coding, the HF
excitation is a white noise over the 0-8000 Hz band.
The index of 4 bits per subframe, denoted indexHF ,,;, (m), sent at 23.85 kbit/s is
demultiplexed from the bit stream (block 405) and decoded by the block 406 as follows:
20 gHBrorr (m) = 2.HP _gain(indexHF grun (111))
in which HP gain(.) is the HF gain quantization dictionary defined in the AMR-WB coding
and recalled below:
i HP_gair"iJ) I HP_gairi,J)
0 0.110595703125000 8 0.342102050781250
1 0.142608642578125 9 0.372497558593750
2 0.170806884765625 10 0.408660888671875
3 0.197723388671875 11 0.453002929687500
4 0.226593017578125 12 0.511779785156250
5 0.255676269531250 13 0.599822998046875f
wo 2015/004373 21 PCT/FR2014/051720
6 0.284545898437500 14 0.741241455078125
7 0.313232421875000 15 0. 998779296875000
Table 1 (gam d1ct10nary at 23.85 kb1t/s)
The block 407 performs the scaling of the excitation signal according to the following
equation:
5 u HB2 (n) = gH&o" (m)uHBI (n), n = 80m, .. · ,80(m + 1) -1
Finally, the energy of the excitation is adjusted to the level of the current subframe with the
following conditions (block 408). The following is computed:
79 2 L;( g(m )gH82 (m )uHB (n))
.fac(m) = "'"-"0'
--
7
"'
9
-----.2::
UHn2(n)'
n=O
The numerator here represents the high-band signal energy which would be obtained in the
10 mode 23.05. As explained before, for the bit rates < 23.85 kbit/s, it is necessary to retain the
level of energy between the decoded excitation signal and the extended excitation signal
uHB(n), but this constraint is not necessary in the case of the 23.85 kbit/s bit rate, since
ul/B(n) is in this case scaled by the gain gHB3(m). To avoid double multiplications, certain
multiplication operations applied to the signal in the block 400 are applied in the block 402 by
15 multiplying by g(m). The value of g(m) depends on the uHB(n)synthesis algorithm and
must be adjusted such that the energy level between the decoded excitation signal in low
band and the signal g(m)uH8 (n) is retained.
In a particular embodiment, which will be described in detail later with reference to figure 7,
g(m) = 0.6gHBl (m), where gHBl (m) is a gain which ensures, for the signal uH8 , the same
20 ratio between energy per subframe and energy per frame as for the signal u(n) and 0.6
corresponds to the average frequency response amplitude value of the de-emphasis filter
from 5000 to 6400 Hz.
It is assumed that, in the block 408, there is information on the tilt of the low-band signal -
in a preferred embodiment, this tilt is computed as in the AMR-WB codec according to the
25 blocks 103 and 104, but other methods for estimating the tilt are possible without changing
the principle of the invention.
If .fac(m) > 1 or tilt< 0, the following is assumed:
uliB '(n) = lluB2 (n), n = 80m, .. ·, 80(m + 1) -1
otherwise:
30 uHB '(n) =max ( --./1- tilt,.fac(m)) .lluB2 (n), n = 80m, .. · ,80(m + 1) -1
wo 2015/004373 22 PCT/FR2014/051720
It will be noted that the optimized scale factor computation described here, notably in the
blocks 401 and 402, is distinguished from the abovementioned equalization of filter levels
performed in the AMR-WB+ codec by a number of aspects:
• The optimized scale factor is computed directly from the transfer functions of the LPC
5 filters without involving any temporal filtering. This simplifies the method.
• The equalization is done preferentially at a frequency different from the Nyquist
frequency (6400 Hz) associated with the low band. Indeed, the LPC modeling
implicitly represents the attenuation of the signal typically caused by the resampling
operations and therefore the frequency response of an LPC filter may be subject at
~o the Nyquist frequency to a decrease which is not at the chosen common frequency.
~5
• The equalization here relies on a filter of lower order (here of order 2) in addition to
the 2 filters to be equalized. This additional filter makes it possible to avoid the
effects of local spectral fluctuations (peaks or troughs) which may be present at the
common frequency for the computation of the frequency response of the prediction
filters.
For the blocks 403 to 408, the advantage of the invention is that the quality of the signal
decoded at 23.85 kbit/s according to the invention is improved relative to a signal decoded at
23.05 kbit/s, which is not the case in an AMR-WB decoder. In fact, this aspect of the
invention makes it possible to use the additional information (0.8 kbit/s) received at
20 23.85 kbitjs, but in a controlled manner (block 408), to improve the quality of the extended
excitation signal at the bit rate of 23.85.
The device for determining the optimized scale factor as illustrated by the blocks 401 to 408
of figure 4 implements a method for determining the optimized scale factor now described
with reference to figure 6.
25 The main steps are implemented by the block 401.
Thus, an extended excitation signal uHE{n) is obtained in a frequency band extension
method E601 which comprises a step of decoding or of extraction, in a first frequency band
called low band, of an excitation signal and of parameters of the first frequency band such as,
for example, the coefficients of the linear prediction filter of the first frequency band.
30 A step E602 determines a linear prediction filter called additional filter, of lower order
than that of the first frequency band. To determine this filter, the parameters of the first
frequency band decoded or extracted are used.
In one embodiment, this step is performed by truncation of the transfer function of
the linear prediction filter of the low band to obtain a lower filter order, for example 2. These
35 coefficients can then be modified as a function of a stability criterion as explained previously
with reference to figure 4.
From the coefficients of the additional filter thus determined, a step E603 is
implemented to compute the optimized scale factor to be applied to the extended excitation
wo 2015/004373 23 PCT/FR2014/051720
signal. This optimized scale factor is, for example, computed from the frequency response of
the additional filter at a common frequency between the low band (first frequency band) and
the high band (second frequency band). A minimum value can be chosen between the
frequency response of this filter and those of the low-band and high-band filters.
5 This therefore avoids the overestimations of energy which could exist in the methods of the
prior art.
This step of computation of the optimized scale factor is, for example, described
previously with reference to figure 4 and figures Sa and Sb.
The step E604 performed by the block 402 or 409 (depending on the decoding bit
10 rate) for the band extension, applies the duly computed optimized scale factor to the
extended excitation signal so as to obtain an optimized extended extension signal uHa'(n).
In a particular embodiment, the device for determining the optimized scale factor 708
is incorporated in a band extension device now described with reference to figure 7. This
15 device for determining the optimized scale factor illustrated by the block 708 implements the
method for determining the optimized scale factor described previously with reference to
figure 6.
20
In this embodiment, the band extension block 400 of figure 4 comprises the blocks
700 to 707 of figure 7 that is now described.
Thus, at the input of the band extension device, a low-band excitation signal decoded
or estimated by analysis is received ( u(n) ). The band extension here uses the excitation
decoded at 12.8 kHz ( exc2 or u(n) ) at the output of the block 302 of figure 3.
It will be noted that, in this embodiment, the generation of the oversampled and
extended excitation is performed in a frequency band ranging from 5 to 8 kHz therefore
25 including a second frequency band (6.4-8 kHz) above the first frequency band (0-6.4 kHz).
30
Thus, the generation of an extended excitation signal is performed at least over the
second frequency band but also over a part of the first frequency band.
Obviously, the values defining these frequency bands can be different depending on
the decoder or the processing device in which the invention is applied.
For this exemplary embodiment, this signal is transformed to obtain an excitation
signal spectrum U(k) by the time-frequency transformation module 500.
In a particular embodiment, the transform uses a OCT-IV (for "Discrete Cosine Transform"type
IV) (block 700) on the current frame of 20 ms (256 samples), without windowing, which
amounts to directly transforming u(n) with n=0, .. ·,255 according to the following
35 formula:
N-l (7[ ( 1)( 1)) U(k)= ~u(n)cos N n+2 k+"2
WO 2015/004373 24 PCT/FR2014/051720
in which N = 256 and k = 0,- · ·, 255.
It should be noted here that the transformation without windowing (or, equivalently, with an
implicit rectangular window of the length of the frame) is possible because the processing is
performed in the excitation domain, and not the signal domain so that no artifact (block
5 effects) is audible, which constitutes an important advantage of this embodiment of the
invention.
In this embodiment, the OCT-IV transformation is implemented by FFT according to
the so-called "Evolved DCT(EOCT)" algorithm described in the article by O.M. Zhang, H.T. Li,
10 A Low Complexity Transform - Evolved OCT, IEEE 14th International Conference on
Computational Science and Engineering (CSE), Aug. 2011, pp. 144-149, and implemented in
the ITU-T standards G.718 Annex Band G.729.1 Annex E.
In variants of the invention, and without loss of generality, the OCT-IV transformation
will be able to be replaced by other short-term time-frequency transformations of the same
15 length and in the excitation domain, such as an FFT (for "Fast Fourier Transform") or a OCTII
(Discrete Cosine Transform- type II). Alternatively, it will be possible to replace the OCTIV
on the frame by a transformation with overlap-addition and windowing of length greater
than the length of the current frame, for example by using an MOCT (for "Modified Discrete
Cosine Transform"). In this case, the delay Tin the block 310 of figure 3 will have to be
20 adjusted (reduced) appropriately as a function of the additional delay due to the
analysis/synthesis by this transform.
The OCT spectrum, U(k), of 256 samples covering the 0-6400 Hz band (at
12.8 kHz), is then extended (block 701) into a spectrum of 320 samples covering the 0-
8000 Hz band (at 16kHz) in the following form:
!0 k = o,. .. '199
25 UHBI(k)= U(k) k=200,-··,239
U(k +start_ band- 240) k = 240,- · · ,319
in which it is preferentially taken that start_band = 160.
The block 701 operates as module for generating an oversampled and extended
excitation signal and performs a resampling from 12.8 to 16 kHz in the frequency domain, by
adding 1/• of samples ( k = 240,- · ·, 319) to the spectrum, the ratio between 16 and 12.8
30 being 5/4.
Furthermore, the block 701 performs an implicit high-pass filtering in the 0-5000 Hz
band since the first 200 samples of UH81 (k) are set to zero; as explained later, this highpass
filtering is also complemented by a part of progressive attenuation of the spectral values
of indices k = 200,- · · ,255 in the 5000-6400 Hz band; this progressive attenuation is
35 implemented in the block 704 but could be performed separately outside of the block 704.
5
wo 2015/004373 25 PCT/FR2014/051720
Equivalently, and in variants of the invention, the implementation of the high-pass filtering
separated into blocks of coefficients of index k = 0, · · ·, 199 set to zero, of attenuated
coefficients k = 200, 00
·, 255 in the transformed domain, will therefore be able to be
performed in a single step.
In this exemplary embodiment and according to the definition of U1181 (k), it will be
noted that the 5000-6000 Hz band of U 1181 (k) (which corresponds to the indices
k = 200, 00
·, 239) is copied from the 5000-6000 Hz band of U(k). This approach makes it
possible to retain the original spectrum in this band and avoids introducing distortions in the
5000-6000 Hz band upon the addition of the HF synthesis with the LF synthesis - in
10 particular the phase of the signal (implicitly represented in the DCT-IV domain) in this band is
preserved.
The 6000-8000 Hz band of U1181 (k) is here defined by copying the 4000-6000 Hz
band of U(k) since the value of start_bandis preferentially set at 160.
In a variant of the embodiment, the value of start_band will be able to be made
15 adaptive around the value of 160. The details of the adaptation of the start_ band value are
not described here because they go beyond the framework of the invention without changing
its scope.
For certain wide-band signals (sampled at 16 kHz), the high band (> 6 kHz) may be
noisy, harmonic or comprise a mixture of noise and harmonics. Furthermore, the level of
20 harmonicity in the 6000-8000 Hz band is generally correlated with that of the lower frequency
bands. Thus, the noise generation block 702 performs a noise generation in the frequency
domain, L{,BN(k) for k=240,oo·,319 (80 samples) corresponding to a second frequency
25
band called high frequency in order to then combine this noise with the spectrum U 1181 (k) in
the block 703.
In a particular embodiment, the noise (in the 6000-8000 Hz band) is generated
pseudo-randomly with a linear congruential generator on 16 bits:
u (k)- , ,
{
0 k=O 00
' 239
HBN - 31821U118N(k-1)+13849 k=240,oo•,319
with the convention that U HBN (239) in the current frame corresponds to the value
U118N (319) of the preceding frame. In variants of the invention, it will be possible to replace
30 this noise generation by other methods.
The combination block 703 can be produced in different ways. Preferentially, an
adaptive additive mixing of the following form is considered:
wo 2015/004373 26 PCT/FR2014/051720
in which G HBN is a normalization factor serving to equalize the level of energy between the
two signals,
319 L UHBI(k)' +s
k-240
319 L ul/BN(k)' +s
k=:240
s with s =0.01, and the coefficient a (between 0 and 1) is adjusted as a function of
parameters estimated from the decoded low band and the coefficient fJ (between 0 and 1)
depends on a .
10
15
In a preferred embodiment, the energy of the noise is computed in three bands:
2000-4000 Hz, 4000-6000 Hz and 6000-8000 Hz, with
EN2-4 = I U'2(k)
k EN(80,159)
EN4-6 = I U'2 (k)
kEN(l60,239)
EN4-6 = I U'2(k)
kEN(240,319)
in which
239 L U2(k)
k-160 U(k)
\ !59
,LU2(k)
k = 80, ... ,159
k"SO
U'(k)= U(k) k = 160, ... ,239
239 L U2(k)
k-160 UHBI(k) 319 k=240, ... ,319
L UHB12(k)
k=240
and N(k1,k2 ) is the set of the indices k for which the coefficient of index k is classified as
being associated with the noise. This set can, for example be obtained by detecting the local
IU'(k)l> IU '(k-1)1 and IU'(k)l > lrr '(k+ 1)1 peaks in U'(k) that verify - - fV and by considering
that these rays are not associated with the noise, i.e. (by applying the negation of the
20 preceding condition):
wo 2015/004373 27 PCf/FR2014/051720
N(a, b) ={asks h!IU '(k)l < IU '(k-1)1 or IU '(k)l < IU '(k+ Ill}
It can be noted that other methods for computing the energy of the noise are possible, for
example by taking the median value of the spectrum on the band considered or by applying a
smoothing to each frequency ray before computing the energy per band.
5 a is set such that the ratio between the energy of the noise in the 4-6 kHz and 6-8 kHz
bands is the same as between the 2-4 kHz and 4-6 kHz bands:
10
a= p-EN6-8
239 I U'(k)-EN6-S
k=160
in which
p = max(p,EN6_8)
In variants of the invention, the computation of a will be able to be replaced by other
methods. For example, in a variant, it will be possible to extract (compute) different
parameters (or "features") characterizing the signal in low band, including a "tilt" parameter
similar to that computed in the AMR-WB codec, and the factor a will be estimated as a
15 function of a linear regression from these different parameters by limiting its value between 0
and 1. The linear regression will, for example, be able to be estimated in a supervised
manner by estimating the factor a by exchanging the original high band in a learning base.
It will be noted that the way in which a is computed does not limit the nature of the
invention.
20
In a preferred embodiment, the following is taken
fJ=~l-a2
in order to preserve the energy of the extended signal after mixing.
In a variant, the factors fJ and a will be able to be adapted to take account of the fact that
25 a noise injected into a given band of the signal is generally perceived as stronger than a
harmonic signal with the same energy in the same band. Thus, it will be possible to modify
the factors fJ and a as follows:
30
fJ +--- fJ.f(a)
a+--- a.f(a)
in which f(a) is a decreasing function of a, for example f(a)=b-a-f;;, b=l.l,
a=1.2, /(a) limited from 0.3 to 1. It must be noted that, after multiplication by f(a),
wo 2015/004373 28 PCT/FR2014/051720
a'+ fJ' 8.85 bit/s.
Then, a bandpass filter is applied in the form:
0
Ghp(k- 200)UHB2 '(k)
UHB3(k) =
UHB2 '(k)
G,p(k-320-N,.)UHB2 '(k)
k = o, .. ·,199
k = 200, .. ·,255
k = 256, .. ·,319-N1"
k =320-N,., ... ,319
wo 2015/004373 30 Per /FR2014/051720
The definition of G,P (k), k = o, .. ·,55, is given, for example, in table 1 below.
K 9h/..k) K 9h/..k) K 9h/..k) K 9h/..k)
0 0.001622428 14 0.114057967 28 0.403990611 42 0.776551214
1 0.004717458 15 0.128865425 29 0.430149896 43 0.800503267
2 0.008410494 16 0.144662643 30 0.456722014 44 0.823611104
3 0.012747280 17 0.161445005 31 0.483628433 45 0.845788355
4 0.017772424 18 0.179202219 32 0.510787115 46 0.866951597
5 0.023528982 19 0.197918220 33 0.538112915 47 0.887020781
6 0.030058032 20 0.217571104 34 0.565518011 48 0.905919644
7 0.037398264 21 0.238133114 35 0.592912340 49 0.923576092
8 0.045585564 22 0.259570657 36 0.620204057 50 0.939922577
9 0.054652620 23 0.281844373 37 0.647300005 51 0.954896429
10 0.064628539 24 0.304909235 38 0.674106188 52 0.968440179
11 0.075538482 25 0.328714699 39 0.700528260 53 0.980501849
12 0.087403328 26 0.353204886 40 0.726472003 54 0. 991035206
13 0.100239356 27 0.378318805 41 0.751843820 55 1.000000000
Table 2
It will be noted that, in variants of the invention, the values of G,P (k) will be able to be
5 modified while keeping a progressive attenuation. Similarly, the low-pass filtering with
variable bandwidth, G1/k), will be able to be adjusted with values or a frequency medium
that are different, without changing the principle of this filtering step.
It will also be noted that the bandpass filtering will be able to be adapted by defining
a single filtering step combining the high-pass and low-pass filtering.
10 In another embodiment, the bandpass filtering will be able to be performed in an
equivalent manner in the time domain (as in the block 112 of figure 1) with different filter
coefficients according to the bit rate, after an inverse ocr step. However, it will be noted that
it is advantageous to perform this step directly in the frequency domain because the filtering
is performed in the domain of the LPC excitation and therefore the problems of circular
15 convolution and of edge effects are very limited in this domain.
20
It will also be noted that, in the case of the 23.85 kbit/s bit rate, the de-emphasis of
the excitation UH82 (k) is not performed to remain in agreement with t.he way in which the
correction gain is computed in the AMR-WB coder and to avoid double multiplications. In this
case, block 704 performs only the low-pass filtering.
The inverse transform block 705 performs an inverse ocr on 320 samples to find the
high-frequency excitation sampled at 16 kHz. Its implementation is identical to the block 700,
:i
5
wo 2015/004373 31 PCT/FR2014/051720
because the OCT-IV is orthonormal, except that the length of the transform is 320 instead of
256, and the following is obtained:
in which N16k =320 and k=0, .. ·,319.
This excitation sampled at 16 kHz is then, optionally, scaled by gains defined per subframe of
80 samples (block 707).
In a preferred embodiment, a gain g"81(m) is first computed (block 706) per subframe by
energy ratios of the subframes such that, in each subframe of index m=O, 1, 2 or 3 of the
10 current frame:
in which
63
e1 (m) = L u(n + 64m )2 + s
n=O
79
e2 (m) = 2:U HBo (n + 80m )2 + s
319
L;uHBo(n)2 +s
e3 (m) = e1( 111) "'-n==o:~5'o5;~--~
L;u(n)2 +s
IP=0
with s = 0.01. The gain per subframe gHB1 (m) can be written in the form:
63 L u(n + 64m )2 + s
255
L;u(n)2 + s
gHBl (Ill)= 11=0
79
L;ullB0(n +80m)2 +s
11=0
319
L;uHBo(n)2 +s
11=0
which shows that, in the signal u HB, the same ratio between energy per subframe and
energy per frame as in the signal u(n) is assured.
The block 707 performs the scaling of the combined signal according to the following
equation:
20 IIHB(n) =gHBI(IIl)UHBO(Il), II =80m, .. ·,80(m+l)-!
wo 2015/004373 32 PCT/FR2014/051720
It will be noted that the implementation of the block 706 differs from that of the
block 101 of figure 1, because the energy at the current frame level is taken into account in
addition to that of the subframe. This makes it possible to have the ratio of the energy of
each subframe in relation to the energy of the frame. The energy ratios (or relative energies)
5 are therefore compared rather than the absolute energies between low band and high band.
Thus, this scaling step makes it possible to retain, in the high band, the energy ratio
between the subframe and the frame in the same way as in the low band.
It will be noted here that, in the case of the 23.85 kbit/s bit rate, the gains gHBl (m)
are computed but applied in the next step, as explained with reference to figure 4, to avoid
10 the double multiplications. In this case u HB (n) = u HBo (n) .
According to the invention, the block 708 then performs a scale factor computation
per subframe of the signal (steps E602 to E603 of figure 6), as described previously with
reference to figure 6 and detailed in figures 4 and 5.
Finally, the corrected excitation uH8 '(n) is filtered by the filtering module 710 which
15 can be performed here by taking as transfer function I I A(zl y), in which y =0.9 at 6.6
kbit/s and y = 0.6 at the other bit rates, which limits the order of the filter to the order 16.
In a variant, this filtering will be able to be performed in the same way as is described for the
block 111 of figure 1 of the AMR-WB decoder, but the order of the filter changes to 20 at the
6.6 bit rate, which does not significantly change the quality of the synthesized signal. In
20 another variant, it will be possible to perform the LPC synthesis filtering in the frequency
domain, after having computed the frequency response of the filter implemented in the block
710.
In a variant embodiment, the step of filtering by a linear prediction filter 710 for the
second frequency band is combined with the application of the optimized scale factor, which
25 makes it possible to reduce the processing complexity. Thus, the steps of filtering I I A(zl y)
and of application of the optimized scale factor gH82 are combined in a single step of filtering
gHB21 A(zl y) to reduce the processing complexity.
In variant embodiments of the invention, the coding of the low band (0-6.4 kHz) will
be able to be replaced by a CELP coder other than that used in AMR-WB, such as, for
30 example, the CELP coder in G.718 at 8 kbit/s. With no loss of generality, other wide-band
coders or coders operating at frequencies above 16 kHz, in which the coding of the low band
operates with an internal frequency at 12.8 kHz, could be used. Moreover, the invention can
obviously be adapted to sampling frequencies other than 12.8 kHz, when a low-frequency
coder operates with a sampling frequency lower than that of the original or reconstructed
35 signal. When the low-band decoding does not use linear prediction, there is no excitation
signal to be extended, in which case it will be possible to perform an LPC analysis of the
il WO 2015/004373 33 PCT/FR2014/0517ZO
signal reconstructed in the current frame and an LPC excitation will be computed so as to be
able to apply the invention.
Finally, in another variant of the invention, the excitation ( u(n)) is resampled, for
example by linear interpolation or cubic "spline", from 1Z.8 to 16 kHz before transformation
5 (for example OCT-IV) of length 3ZO. This variant has the defect of being more complex,
because the transform (OCT-IV) of the excitation is then computed over a greater length and
the resampling is not performed in the transform domain.
Furthermore, in variants of the invention, all the computations necessary for the
estimation of the gains ( G HBN , g HBI (m) , g Ha2 (m) , g HBN, ... ) will be able to be performed
10 in a logarithmic domain.
15
20
In variants of the band extension, the excitation in low band u(n) and the LPC filter
11 A(z) will be estimated per frame, by LPC analysis of a low-band signal for which the band
has to be extended. The low-band excitation signal is then extracted by analysis of the audio
signal.
In a possible embodiment of this variant, the low-band audio signal is resampled
before the step of extracting the excitation, so that the excitation extracted from the audio
signal (by linear prediction) is already resampled.
The band extension illustrated in figure 7 is applied in this case to a low band which
is not decoded but analyzed.
Figure 8 represents an exemplary physical embodiment of a device for determining
an optimized scale factor 800 according to the invention. The latter can form an integral part
of an audio frequency signal decoder or of an equipment item receiving audio frequency
signals, decoded or not.
This type of device comprises a processor PROC cooperating with a memory block BM
25 comprising a storage and/or working memory MEM.
Such a device comprises an input module E suitable for receiving an excitation audio signal
decoded or extracted in a first frequency band called low band (u(n) or U(k)) and the
parameters of a linear prediction synthesis filter ( A(z) ). It comprises an output module S
suitable for transmitting the synthesized and optimized high-frequency signal (uH6(n)) for
30 example to a filtering module like the block 710 of figure 7 or to a resampling module like the
module 311 of figure 3.
The memory block can advantageously comprise a computer program comprising
code instructions for implementing the steps of the method for determining an optimized
scale factor to be applied to an excitation signal or to a filter within the meaning of the
35 invention, when these instructions are executed by the processor PROC, and notably the
steps of determination (E60Z) of a linear prediction filter, called additional filter, of lower
order than the linear prediction filter of the first frequency band, the coefficients of the
wo 2015/004373 34 PCT/FR2014/051720
additional filter being obtained from parameters decoded or extracted from the first
frequency band, and of computation (E603) of an optimized scale factor as a function at least
of the coefficients of the additional filter.
Typically, the description of figure 6 reprises the steps of an algorithm of such a
5 computer program. The computer program can also be stored on a memory medium that can
be read by a reader of the device or that can be downloaded into the memory space thereof.
The memory MEM stores, generally, all the data necessary for the implementation of
the method.
In a possible embodiment, the device thus described can also comprise functions for
10 application of the optimized scale factor to the extended excitation signal, of frequency band
extension, of low-band decoding and other processing functions described for example in
figures 3 and 4 in addition to the optimized scale factor determination functions according to
the invention.
15

CLAIMS
1. A method for determining an optimized scale factor to be applied to an excitation signal
or to a filter in an audio frequency signal frequency band extension method, the band
5 extension method (E601) comprising a step of decoding or of extraction, in a first
frequency band, of an excitation signal and of parameters of the first frequency band
comprising coefficients of a linear prediction filter, a step of generation of an extended
excitation signal on at least one second frequency band and a step of filtering, by a linear
prediction filter, for the second frequency band, the determination method being
10 characterized in that it comprises the following steps:
- determination (E602) of a linear prediction filter called additional filter, of lower order
than the linear prediction filter of the first frequency band, the coefficients of the
additional filter being obtained from the parameters decoded or extracted from the first
frequency band; and
15 -computation (E603) of the optimized scale factor as a function at least of the
coefficients of the additional filter.
2. The method as claimed in claim 1, characterized in that the band extension method
comprises a step of application (E604) of the optimized scale factor to the extended
20 excitation signal.
3. The method as claimed in claim 2, characterized in that the application of the optimized
scale factor is combined with the step of filtering in the second frequency band.
25 4. The method as claimed in claim 1, characterized in that the coefficients of the additional
filter are obtained by truncation of the transfer function of the linear prediction filter of
the first frequency band to obtain a lower order.
5. The method as claimed in claim 4, characterized in that the coefficients of the additional
30 filter are modified as a function of a stability criterion of the additional filter.
6. The method as claimed in claim 1, characterized in that the computation of the optimized
scale factor comprises the following steps:
- computation of the frequency responses of the linear prediction filters of the first and
35 second frequency bands for a common frequency;
- computation of the frequency response of the additional filter for this common
frequency;
- computation of the optimized scale factor as a function of the duly computed frequency
responses.
:j
wo 2015/004373 36 PCT/FR2014/051720
7. The method as claimed in claim 1, characterized in that it further comprises the following
steps, implemented for a predetermined decoding bit rate:
- first scaling of the extended excitation signal by a gain computed for each subframe as
5 a function of an energy ratio between the decoded excitation signal and the extended
excitation signal;
- second scaling of the excitation signal obtained from the first scaling by a decoded
correction gain;
- adjustment of the energy of the excitation for the current subframe by an adjustment
10 factor computed as a function of the energy of the signal obtained after the second
scaling and as a function of the signal obtained after application of the optimized scale
factor.
8. A device for determining an optimized scale factor to be applied to an excitation signal or
15 to a filter in an audio frequency signal frequency band extension device, the band
extension device ( 400) comprising a module for decoding or extracting, in a first
frequency band, an excitation signal and parameters of the first frequency band
comprising coefficients of a linear prediction filter, a module for generating an extended
excitation signal on at least one second frequency band and a module for filtering, by a
20 linear prediction filter, for the second frequency band, the determination device being
characterized in that it comprises:
-a module (401a) for determining a linear prediction filter called additional filter, of lower
order than the linear prediction filter of the first frequency band, the coefficients of the
additional filter being obtained from the parameters decoded or extracted from the first
25 frequency band; and
-a module (401b) for computing the optimized scale factor as a function at least of the
coefficients of the additional filter.
9. An audio frequency signal decoder, characterized in that it comprises a device for
30 determining an optimized scale factor according to claim 8.
35
10. A computer program comprising code instructions for implementing the steps of the
method for determining an optimized scale factor as claimed in one of claims 1 to 7,
when these instructions are executed by a processor.
11. A storage medium that can be read by a device for determining an optimized scale factor
on which is stored a computer program comprising code instructions for the execution of
steps of the method for determining an optimized scale factor as claimed in one of
claims 1 to 7.