FIELD OF THE INVENTION
[001] The present invention relates to a low power keyword based speech recognition
scheme for hands free wakeup of devices. More specifically, the present invention relates to a
low power keyword based speech recognition wake up scheme for hands free wakeup of devices
that can be used in Always-ON domain by virtue of its very low power consumption.
BACKGROUND OF THE RELATED ART
[002] Speech recognition systems allow a user to control the device with speech
recognition capability using natural language interface in a hands free manner.
[003] Generally, in most of the devices like cell phones or PNDs, user needs to use his/her
hands in order to start interacting with the device. For instance, by pushing a button or by turning
on the power delivered to the device. The electronic devices tend to move in a dormant state or
"sleep mode" when not used for a pre-specified time. For example, mobile phones when not used
for a pre-specified time, transition to a dormant state and remain there unless prompted by the
user or any other external signal. The tendency of devices to move in "sleep mode" enables them
to save significant amount of power.
[004] However, waking up the device from sleep mode to active state requires an input
from the user terminal generally by turning on an external switch or pushing a button. For
instance a cell phone in sleep mode comes out of it when any key is pressed by the user. Hence,
to make these devices more convenient and user friendly, there is a need for a mechanism that
allows hands free wake up of devices without the need for the user to turn on the switch or press
the button every time.
[005] Key word based wake up of devices is a new paradigm in speech recognition
technology that enables the wakeup of devices such as cell-phones, PNDs etc. using speech
recognition technology or natural speech input. The system remains in sleep mode until a prespecified
keyword is enunciated by the user. Upon recognition of the keyword, the system
transitions from the sleep mode to the active mode. Thus, the user activates the device using a
spoken word or phrase that makes the device more convenient and easy to use.
2
[006] However, systems incorporating speech recognition based wake up control must
continuously hunt for any voice activity or continuously listen to any keyword uttered by the
user in order to activate the device upon user's request. Whereas, speech recognition being a
computationally intensive technology requiring several million operations per second, consumes
significant amount of power that makes it impossible for the low power operated devices to keep
the keyword based hands free voice detection system in always active mode.
[007] Moreover, software solutions for speech recognition are not particularly designed to
be power efficient. They consume significant amount of power during the time device is looking
for spoken keyword. This is due to the fact that they have to run at an operating frequency of
upwards of 100 MHz and also has to have a large DDR Memory footprint.
[008] In light of the foregoing limitations, a keyword based speech recognition scheme for
hands free wake up of devices is needed that consumes less power and remains in Always-On
domain to hunt for voice activity.
BRIEF DESCRIPTION OF THE DRAWINGS
[009] Figure 1 is a block diagram for schematic representation of the hardware architecture
for speech recognition in accordance with an embodiment of the present invention.
[0010] Figure 2A is a block diagram representing the front end and its components in
accordance with an embodiment of the present invention.
[0011] Figure 2B is a block diagram representing the back end and its components in
accordance with an embodiment of the present invention.
[0012] Figure 3 is a schematic representation of the application processor that utilizes the
speech recognition hardware system in accordance with an embodiment of the present invention.
The highlighted region in Figure 3 shows the active, Always ON domain region. This domain
needs to always remain active in order to do voice activity detection.
[0013] Figure 4 is a schematic representation of the application processor that utilizes the
speech recognition hardware system in accordance with an embodiment of the present invention.
The highlighted region in FIGURE 4 shows the ON Domain for keyword detection after voice
activity is detected where system works out of SRAM for keyword detection after the voice
activity is detected.
3
•
[0014] Figure 5 is a schematic representation of the application processor that utilizes the
speech recognition hardware system in accordance with an embodiment of the present invention.
The highlighted region in FIGURE 5 shows the ON Domain for keyword detection after voice
activity is detected where system works out of the DDR for keyword detection after the voice
activity is detected.
[0015] Figure 6 is a flowchart illustrating the mechanism for low power keyword based
hands free wake-up in accordance with an embodiment of the present invention.
DETAILED DESCRIPTION OF THE PRESENT INVENTION
[0016] The present invention proposes a system and the mechanism for keyword based
hands free wake up that stays active all the time and consumes minimal amount of power.
| [0017] The keyword recognition approach is done in two stages that allow the system to go
into a low power state while simultaneously hunting for voice activity. The hardware based
scheme is embedded in the application processor chip that puts a segment of digital circuitry of
the application processor in Always-ON domain enabling it to consume very little power while
hunting for the voice while the rest segment of the application processor chip has been poweredoff.
[0018] The system goes into low power state if no activity is detected for a pre-specified
time and the system is in idle state, by deactivating various modules of application processors.
[0019] At this state the back end clock to domain2 134 is stopped while lowering down the
frequency of clock domain 1 130 and domain3 132 up to quite a significant level, while still
hunting for voice activity.
[0020] Upon detection of the voice activity there is a sudden escalation in the frequency of
the clock to domainl 130 and domain3 132. Along-with this proliferation the clock to domain2
134 is activated and the system runs into the full activated mode if the detected voice signal is
found to be a valid keyword.
[0021] However, if the detected voice activity or audio signal is found to be invalid i.e. do
not match with the keyword in the database, then the system gets back into the low power mode,
by shutting down all the unrequired modules of the application processors while still hunting for
the voice activity.
4
•
[0022] Figure 1 is a schematic representation of the hardware architecture for speech
recognition in accordance with an embodiment of the present invention. The system 100
comprising: a speech recognition hardware 110, a viterbi decoder 124, a senone scorer 122, an
arithmetic logic unit (ALU-FE) 128, an arithmetic logic unit (ALU-BE) 136, a backend 126, a
silence filter 114, a feature creator 116, a frontend 112, an arbiter 118, a host interface 120, a
DDR memory of backend 104, a SRAM of backend 102, a SRAM of frontend 106 and a
memory interface switch 108.
[0023] In accordance with this and the related objects, the system and the mechanism used
to fulfill the purpose as described in the present invention includes: a front end 112 consisting of
a silence filter 114 or a voice activity detector for detecting the voice activity and a feature
creator 116 in communication with silence filter for splitting the utterance into overlapping
frames of 25ms with an overlap of 15ms; a back end 126 consisting of two functional blocks that
are senone scorer 122 and viterbi decoder 124 used for processing the data; The system 100 has
three clock domains: front end 112 along with its SRAM(i.e. FE memory SRAM) works as clock
domainl 130 , back end 126 works as clock domain2 134, and host interface 120 works on clock
domain3 132.
[0024] In an embodiment of the proposed invention a speech recognition system 100
incorporating a frontend 112 is provided. The frontend 112 is the part responsible for detection
of voice activity and generation of feature vectors that are further used for determining whether
keyword was present in the detected voice activity or not. The said front end 112 comprises the
silence filter 114, the feature creator 116, the frontend memory 106 and the ALU-FE 128.
[0025] The silence filter 114 also known as voice activity detector (VAD), takes the audio
inputs in form of 16 Bits data (16 KHz or 8 KHz). It detects the voice activity and propagates
those parts of speech further that have voice activity in it. For example a command phrase like
"HELLO PND" when spoken preceded and followed by pauses will have its preceding and
following pauses removed by silence filter. Typically the silence filter 114 will keep calibrating
itself to account for ambient noise and will start passing speech audio downstream when it hears
voice beyond preset thresholds over ambient noise This is called voice activity detection or
VAD. It'll keep passing the speech audio downstream till it encounters a long programmable
pause in speech. The output of silence filter is a full utterance delimited by start and end flags.
5
•
[0026] After the detection of the voice activity, feature vectors are extracted from the
incoming utterance by the feature creator 116. Feature extraction is a step to reduce the
dimensionality of the input utterance. The feature creator 116 splits the utterance into frames and
extracts features from each frame. The utterance is then changed into a sequence of feature
vectors. The feature creator 116 splits the utterance into overlapping 25 ms frames with an
overlap of 15 ms. The frames are then subjected to pre-emphasis. Pre emphasis is done in order
to compensate the high-frequency part of the speech signal as the voiced segments have more
energy at lower frequencies than higher frequencies. A window is then applied to each frame in
order to minimize the signal discontinuities at the edges of the frame. Each frame of the speech
signal is then subjected to Mel Frequency Cepstral Coefficient (MFCC) generation. The MFCC
extraction process generates 13 MFCCs for each frame. These 13 MFCCs are then converted to
39 Dynamic Feature vectors, for each frame, by doing delta and delta-delta operations on them
across each frames. Thus, the utterance is converted into a sequence of feature vectors. MFCC
are generally used as features in speech recognition systems, such as the systems that can
automatically recognize the spoken words, like the numbers spoken into a telephone. These are
also used to recognize the speakers based on their voice. MFCCs are also increasingly finding
uses in music information retrieval applications such as genre classification, audio similarity
measures and many more.
[0027] The back end 126 is the part where bulk of processing happens. It has primarily two
functional blocks senone scorer 122 and viterbi decoder 124.
[0028] The senone scorer 122 calculates scores of active senones i.e. senones corresponding
to active hmms in each frame, based on the feature vector values of the frame calculated by front
end.
[0029] The viterbi decoder 124 processes frames one after other in a time synchronous
manner for complete search. It works on the lexical tree and null transaction databases using
senone scores calculated by the senone scorer 122. Search space pruning is done at each frame to
keep search space within reasonable limits. An intermediate output of this stage is a history entry
table. Once the decoding is over, hardware analyzes history entry table by using simple viterbi
backtrace. It interrupts the system and provides indication to system if keyword detection was
successful or not. This last step (of Back End running Viterbi backtrace can be enabled or
disabled). In a situation when this feature is disabled, Output of Back End is a History Entry
6
Table. This table has the complete information to arrive at the spoken utterance and host
software uses it to find the spoken phrase or a list of most probable spoken phrases (nBest list).
This mode will be used when the Speech Recognition hardware 110 is used in full functional
mode i.e if the system has detected the keyword successfully.
[0030] Figure 2A is a block diagram representing the front end and its components in
accordance with an embodiment of the present invention. Referring to Figure 2 A, a front end 112
consists of a silence filter 114 or a voice activity detector for detecting the voice activity and a
feature creator 116 in communication with silence filter for splitting the utterance into
overlapping frames of 25ms with an overlap of 15ms.
[0031] The silence filter 114, also known as voice activity detector, is a part of the frontend
112 of speech recognition hardware that remains in always-ON domain in order to detect any
voice activity in the spoken audio input. The silence filter 114 takes the audio input in the form
of 16 bit data. It keeps calibrating itself to account for the ambient noise and presets a threshold
value above the ambient noise. When voice activity above the preset threshold level is detected
in the audio input, the parts of the speech having the voice activity in them are then propagated
to the feature creator 116. For example a command phrase like "HELLO PND" when spoken
preceded and followed by pauses will have its preceding and following pauses removed by
silence filter.
[0032] After receiving the utterance having the voice activity, the feature creator 116 splits
the utterance into overlapping frames of 25ms with an overlap of 15ms. After pre emphasis and
windowing, 13 MFCCs are generated for each frame. The first and second derivatives (delta and
delta-delta operation) of these MFCCs then result in 39 dynamic feature vectors for each of the
frame based on the feature vector values calculated by the front end for each frame.
[0033] Figure 2B is a block diagram representing the back end and its components in
accordance with an embodiment of the present invention. Referring to Figure 2B, the back end
126 consists of two functional blocks that are senone scorer 122 and viterbi decoder 124 used for
processing the data.
[0034] The senone scorer 122, calculates the scores of the active senones that is the senones
corresponding to the active HMMS in each frame; the viterbi decoder 124 processes the frames
one after other in a time synchronous manner. Using the senone scores calculated by senone
scorer 122 it works on Lexical Tree and Null transaction databases and completes the search.
7
Search space pruning is done at each frame to keep search space within reasonable limits. The
output of this stage is a history entry table. This table has the complete information to arrive at
the spoken utterance. If the viterbi back trace is enabled the hardware analyzes the history entry
table by using viterbi back trace that is tracking back the best path to the beginning. It interrupts
the system and provides indication to system if keyword detection was successful or not. If the
viterbi back trace is not enabled then the output of the Back End 126 is a History Entry Table.
The host software then uses this table to find the spoken phrase or a list of most probable spoken
phrases using some sophisticated DAG (directed acyclic graph) based algorithms.
[0035] Figure 3 is a schematic representation of the application processor that utilizes the
speech recognition hardware system in accordance with an embodiment of the present invention.
The highlighted region 302 in Figure 3 shows the active, Always ON domain region. This
domain needs to always remain active in order to do voice activity detection. The highlighted
domain 302 represents the active part of the system 300 that always remains in active mode
hunting for the voice activity in the low power state. In this state as shown in Figure 3 the MIC
308, the audio codec 306, the power manager 304, the speech recognition hardware 110 and the
FE memory (SRAM) 106 remains active for voice input.
[0036] The system 300 has 3 clock domains. The Front end 112 along with the SRAM 106
works as the clock domainl 130. The Back end 126 works as clock domain2 134 and host
interface works as clock domain 3 132. The Clock domainl 130 and domain2 134 are same, the
only difference is gating. Clock to domain2 134 is a gated version of clock to domainl 130 and
so can be independently disabled.
[0037] According to the Keyword recognition scheme, in order to reduce the power
consumption to quite a lower level when the system remains in idle state for more than a prespecified
duration, the system gracefully deactivates different modules of application processor
and the clock domain2 134 is stopped (gated), frequency of clocks to domainl 130 and domain3
132 are reduced to a range of about lOOKhz. At this stage, hardware i.e. the Front End 112 stays
in always active mode to hunt for voice activity. Audio data is continuously pumped into the
Front End 112 under the control system of power manager 304 and it keeps doing calibration and
voice activity detection.
[0038] Figure 4 is a schematic representation of the application processor that utilizes the
speech recognition hardware system in accordance with an embodiment of the present invention.
8
The highlighted region in FIGURE 4 shows the ON Domain for keyword detection after voice
activity is detected where system works out of SRAM for keyword detection after the voice
activity is detected. Referring to Figure 4, the highlighted domain 402 shows the component that
goes into the active state when voice signal are detected activating the memory interface switch
108 and BE memory (SRAM) 102 for detection of the keyword.
[0039] Upon detection of the voice activity by the system 400, an indication from the front
End 112 is provided to the system power manager 304, after that clock to the domain 1 130 and
the domain3 132 are jacked up to the range of about 50Mhz from the range of about lOOKhz.
Here voltage jacking will also be done if voltage scaling is used, followed by the activation of
the back end clock to domain2 134. After domain2 134 is started, the BE SRAM 102 is
activated. Memory subsystem is started with appropriate clock with Bandwidth of approx.
20MB/Sec.
[0040] Figure 5 is a schematic representation of the application processor that utilizes the
speech recognition hardware system in accordance with an embodiment of the present invention.
The highlighted region in FIGURE 5 shows the ON Domain for keyword detection after voice
activity is detected where system works out of the DDR for keyword detection after the voice
activity is detected. The highlighted domain 502 of the system 500 represents the active state of
the system 500 after the detection of the keyword. Here in this stage system works out of DDR
310 for keyword detection after the voice activity detection has happened. After the activation of
the memory subsystem and the BE SRAM 102 or BE DDR 104, Back End databases are
initialized in either BE SRAM 102 or BE DDR 104 (as the case may be) for recognition of the
keyword inputted by the audio codec and finally handshaking between the back end 126 and the
front end 112 is started for data input and utterance decoding is started.
[0041] After utterance decoding is completed, the hardware interrupts the power manager
304 to indicate the detection of the keyword in form of decoded utterance. If the decoded
utterance is found to be the keyword the system then enters into the full performance mode and
is further ready for doing more sophisticated speech recognition.
[0042] Furthermore, if decoded utterance is not found to be the keyword or if no activity is
detected again for a preset duration, then system again goes back to the low power state by
stopping down the back end clock domain2 134 followed by the reduction in the frequencies of
clocks to the domainl 130 and the domain3 132.
9
[0043] Figure 6 is a flowchart illustrating the mechanism for low power keyword based
hands free wake-up in accordance with an embodiment of the present invention. Referring to
figure 6, the system is in active state 602 and is tracked whether the system remains idle for more
than pre-specified time, in step 604. The system continuously remains to be in active state if it
does not remain idle for more than a pre-specified time. If the system remains idle for more than
a pre-specified time, then the various modules of the application processor are gracefully
deactivated 606. The backend clock (clock domain2) is stopped 608 and the frequency of clock
domain 1 and3 is reduced 610. In the next step 612, the hardware is enabled to hunt for voice
activity and the application processor chip is put into low power mode by turning off all other
power domains. This will lead to the system to come down in low power state as shown in step
614.
[0044] If the system is in low power step 614, than it continuously hunt for voice activity, if
no voice activity is detected, than the system maintain itself in the low power state 614.
However, If the voice activity is detected, than an indication will be sent from front end to power
manager in step 618, that results in jacking up of clock domainl and clock domain3 upto 50
MHz approx as shown in step 620. Further the clock to domain2 get started 622. In the next step
624 memory subsystem get started. The back end SRAM (or DDR as the case may be)is
powered up and back end databases are initialized in back end SARM (or DDR) for keyword
recognition. Further the utterance is decoded in step 626 and checked for keyword detection in
step 628. If the keyword is detected then the power manager is interrupted 630 and system is
brought to full performance mode 632 and it will remains in active state 602. However if the
keyword is not detected in the utterance than the hardware interrupts power manager 634 and the
system goes to step 608 where backend clock (clock domain2) is stopped and frequency of clock
domainl and clock domain3 is reduced.
[0045] Example 1: Dictionary file for a keyword application (keyword is HELLO SIMSIM):
HELLO HH AH L OW
HELLO (2) HH EH L OW
SIMSIM S IH M S IH M
Gl AA
G2AE
G3AH
10
G4A0
G5AW
G6AY
G7B
G8CH
G9 D
G10DH
Gil EH
G12ER
G13EY
G14F
G15G
G16HH
G17IH
G18IY
G19JH
G20K
G21 L
G22M
G23N
G24NG
G25 OW
G26 0Y
G27P
G28R
G29S
G30 SH
G31T
G32TH
G33UH
G34UW
11
G35V
G36W
G37Y
G38Z
G39ZH
[0046] Example 2: Grammar file for a keyword application (keyword is HELLO SIMSIM):
#JSGFV1.0;
grarnmarkewword_test;
public = [] [] [];
= ;
= ;
= ;
= ;
= ;
= (HELLO SIMSIM);
= (Gl | G2 | G3 | G4 | G5 | G6 | G7 | G8 | G9 | G10 | Gl 1 | G12 | G13 |
G14 | G15 | G16 | G17 | G18 | G19 | G20 | G21 | G22 | G23 | G24 | G25 | G26 | G27 | G28
| G29 | G30 | G31 | G32 | G33 | G34 | G35 | G36 | G37 | G38 | G39) +;
[0047] Example 3: Dictionary file for a simple camera application:
AM EY EH M
APRIL EY P R AH L
AUGUST AA G AH S T
AUGUST (2) AO G AH S T
AUTO AO T OW
BEACH B IY CH
CLICK K L IH K
DATE D EY T
DECEMBER D IH S EH M B ER
DISPLAY D IH S P L EY
12
EASY IY ZIY
EIGHT EY T
EIGHTEEN EY T IY N
EIGHTEENTH EY T IY N TH
EIGHTH EY T TH
EIGHTH(2) EY TH
EIGHTY EY T IY
ELEVEN IH L EH V AH N
ELEVEN(2) IY L EH V AH N
ELEVENTH IH L EH V AH N TH
ELEVENTH (2) IY L EH V AH N TH
FEBRUARY F EH B Y UW W EH R IY
FEBRUARY (2) F EH B R UW W EH R IY
FIFTEEN F IH F T IY N
FIFTEENTH F IH F T IY N TH
FIFTH F IH F TH
FIFTH (2) F IH TH
FIFTY F IH F T IY
FIREWORKS F AY R W ER K S
FIRST F ER S T
FIVE F AY V
FORTY FAORTIY
FOUR F AO R
FOURTEEN F AO R T IY N
FOURTEENTH F AO R T IY N TH
FOURTH F AO R TH
GOURMET G UH R M EY
ISO AY AE S OW
JANUARY JH AE N Y UW EH R IY
JULY JH UW L AY
JULY (2) JH AH L AY
13
JUNE JH UW N
LANDSCAPE L AE N D S K EY P
LANDSCAPE (2) L AE N S K EY P
MARCH M AA R CH
MAY M EY
MODE M OW D
MOVIE M UW V IY
NINE N AY N
NINETEEN N AY N TIY N
NINETEENTH N A Y N T IY N TH
NINETY N AY N T IY
NINTH N AY N TH
NOVEMBER N OW V EH M B ER
OCTOBER AA K T OW B ER
ONE W AH N
ONE (2) HH W AH N
PANORAMA P AE N ER AE M AH
PETSPEHTS
PICTURE PIH K CH ER
PM P IY EH M
PORTRAIT P AO R T R AH T
READY R EH D IY
REDO R IY D UW
SCENE SIYN
SECOND S EH K AH N D
SECOND (2) S EH K AH N
SELECTION S AH L EH K SH AH N
SEPTEMBER S EH P T EH M B ER
SET S EH T
SEVEN S EH V AH N
SEVENTEEN S EH V AH N T IY N
14
SEVENTEENTH S EH V AH N TIY N TH
SEVENTH S EH V AH N TH
SEVENTY S EH V AH N T IY
SEVENTY (2) S EH V AH N IY
SHOOT SH UW T
SIX S IH K S
SIXTEEN S IH K S T IY N
SIXTEENTH S IH K S T IY N TH
SIXTH S IH K S TH
SIXTY S IH K S T IY
SNAP S N AE P
SNOW S N OW
SOFT S AA F T
SOFT (2) S AO F T
SPORTS S P AO R T S
TEN T EH N
TENTH T EH N TH
THIRD TH ER D
THIRTEEN TH ER T IY N
THIRTEENTH TH ER T IY N TH
THIRTIETH TH ER T IY AH TH
THIRTIETH (2) TH ER T IY IH TH
THIRTY TH ER D IY
THIRTY (2) TH ER T IY
THREE THRIY
TIME T AY M
TWELFTH T W EH L F TH
TWELVE T W EH L V
TWENTIETH T W EH N T IY AH TH
TWENTIETH (2) T W EH N T IY IH TH
TWENTIETH (3) T W EH N IY AH TH
15
TWENTIETH (4) T W EH N IY IH TH
TWENTY TW EH NTIY
TWENTY (2) T W EH N IY
TWILIGHT T W AY L AY T
TWO T UW
ZERO Z IH R OW
ZERO (2) Z IY R OW
[0048] Example 4: The Phoneme set: The current phoneme set has 39 phonemes. This
phoneme (or more accurately, phone) set is based on the ARPAbet symbol set developed for
speech recognition uses.
Phoneme Example Translation
AA odd AA D
AE at AE T
AH hut HH AH T
AO ought AO T
AW cow K AW
AY hide HH AY D
B be B IY
CH cheese CH IY Z
D dee D IY
DH thee DH IY
EH Ed EH D
ERhurtHHERT
EY ate EY T
F fee F IY
G green G R IY N
HH he HH IY
IH it IH T
IY eat IY T
16
JH gee JHIY
K key KIY
L lee L IY
M me M IY
N knee NIY
NG ping P IH NG
OW oat OW T
OY toy TOY
P pee P IY
R read R IY D
S sea S IY
SH she SH IY
T tea T IY
TH theta TH EY T AH
UH hood HH UH D
UW two T UW
V vee VIY
W we W IY
Y yield Y IY L D
Z zee Z IY
ZH seizure S IY ZH ER
[0049] Example 5: Grammar file for a simple camera application:
#JSGFV1.0;
grammarsonyenhanced;
public = | | | |
|
| | | | | | |
| | | | | | | |

|
17
I
| | | ;
= PICTURE [MODE];
= DISPLAY [MODE];
= [] [];
= [];
= SET TIME;
= SET DATE;
= | | ;
= SCENE [SELECTION];
= [SCENE] SELECTION;
= SCENE SELECTION;
= | | ;
= EASY [SHOOT];
= [EASY] SHOOT;
= EASY SHOOT;
= AUTO;
= PANORAMA;
= MOVIE;
- ISO;
= | | ;
= SOFT SNAP;
= [SOFT] SNAP;
= SOFT [SNAP];
= SPORTS;
= LANDSCAPE;
= PETS;
- GOURMET;
= TWILIGHT;
= PORTRAIT;
= BEACH;
18
= SNOW;
= FIREWORKS;
= AM | PM;
= | | ;
= | | | |
( []);
= | | | ( []) |

| | | | (
[]);
= | | | | | | |
|
| | | ;
= | | | |
( [])
| ( []);
= ;
= ZERO;
= ONE | TWO | THREE | FOUR | FIVE | SIX | SEVEN | EIGHT | NINE ;
= FIRST | SECOND | THIRD | FOURTH | FIFTH | SIXTH | SEVENTH |
EIGHTH | NINTH;
= TEN | ELEVEN | TWELVE;
= TENTH | ELEVENTH | TWELFTH;
= THIRTEEN | FOURTEEN | FIFTEEN | SIXTEEN | SEVENTEEN |
EIGHTEEN | NINETEEN;
= THIRTEENTH | FOURTEENTH | FIFTEENTH | SIXTEENTH |
SEVENTEENTH
| EIGHTEENTH | NINETEENTH;
= TWENTY | THIRTY | FORTY | FIFTY;
= TWENTY | THIRTY;
= TWENTIETH | THIRTIETH;
19
= SIXTY | SEVENTY | EIGHTY | NINETY;
= JANUARY;
= FEBRUARY;
= MARCH;
= APRIL;
= MAY;
= JUNE;
= JULY;
= AUGUST;
= SEPTEMBER;
= OCTOBER;
= NOVEMBER;
= DECEMBER;
= READY;
= CLICK;
= REDO;
[0050] In accordance with an embodiment of the present invention, the invention finds
application in areas including voice dialing, robotics, voice activated consumer products,
interactive voice response applications, low power high performance voice enabled embedded
applications, video games, hands free computing etc.

We claim:
1. A method for voice based activation of an electronic system comprising:
putting the system into low power mode when the system remains idle for more than a
pre-specified time;
maintaining a database of preselected keywords;
continuously searching for voice activity in low power mode;
capturing the voice activity and determining whether a match exists between said voice
activity and at least one of said keywords while remaining in low power mode;
activating the electronic system if at least one match exists between said voice activity
and keywords;
remaining in low power mode if the match does not exist between said voice activity and
said keywords.
2. The method of claim 1 wherein the voice activity is captured using a specialized speech
recognition hardware.
3. The method of claim 1 wherein the low power mode is attained through by keeping only
the voice activity detector ON in low performance.
: 4. The method of claim 1 wherein the keywords are the words to be used for activation of
the electronic device.
5. The method of claim 1 wherein the keywords are generated by the user and are stored in
the storing database.
6. A low power keyword based speech recognition system for activating an electronic
device comprising:
a first module for detecting a voice activity;
a second module for keyword recognition;
a processor in communication with the first module and the second module, wherein the
said processor deactivates the said second module and reduce the frequency of said first
module when the system remains idle beyond a pre-specified time;
21
a power manager for receiving the feedback from the said first module, wherein the said
power manager activates the said second module and jacks up the frequency of said first
module on detection of said voice activity;
an application programming interface in the said second module to determine whether a
match exist between the said voice activity and said keywords, wherein on a match
detection, the said application programming interface brings the electronic device to fiill
power mode.
7. The system as claimed in claim 6 wherein the frequency of the said first module is in
range of 100 kHz and requires SRAM in the range of 80 KB with a bandwidth of 200
KB/s for doing voice activity detection.
8. The system as claimed in claim 6 wherein the frequency of the said second module is in
the range of 50 Mhz and requires memory in the range of 2.7 MB with a bandwidth of 20
MB/s.
9. The system of claim 6 wherein the said first module remains in ON state to hunt for voice
activity.
10. The system of claim 6 wherein the said second module gets activated on detection of
voice activity.
11. The system of claim 6 wherein the power manager activates the said second module on
detection of the voice activity by the said first module.
12. The system of claim 6 wherein the application programming interface brings the device
to full power mode if the match occurs between said keywords and said voice activity.
13. A method for activating an electronic device using a speech recognition system
comprising:
Maintaining a database of preselected keywords;
when the electronic device remains idle for a pre-specified time, bringing the system in
sleep mode by keeping a first module meant for voice activity detection in low frequency
mode and deactivating a second module meant for keyword recognition;
continuously searching for voice activity by said first module in low frequency mode;
22
activating the said second module on detection of said voice activity;
determining whether a match exists between said voice activity and at least one of said
keywords;
bringing the electronic device to fiall power mode, if a match exist between said voice
activity and said keywords;
putting back the system in sleep mode, if the match does not exist between said voice
activity and at least one keywords.
14. The method of claim 13 wherein the frequency of the said first module is in range of 100
kHz and requires around 80 KB SRAM with a bandwidth of 200 KB/s.
15. The method of claim 13 wherein the frequency of the said second module is in the range
of 50 Mhz and requires memory in the range of 2.7 MB with a bandwidth of 20 MB/s .
16. The method of claim 13 wherein the first module is a voice activity detector.
17. The method of claim 13 wherein keywords are the words used to activate the electronic
device.
18. The method of claim 13 wherein the keywords are predefined by the user and are stored
in the database.