121~
APPARATUS AND METHOD FOR
AUTOMATIC SPEECH ACTIVITY DETECTION
Background of the Invention
This invention relates to an apparatus and method
for speaker independent speech activity detection in an environ-
ment of relatively high level noise, and to automatic speech
recognizers which use such speake independent speech activity
detection.
Automatic speech recognition systems provide a means
for man to interface with communication equipment, computers
and other machines in a human's most natural and convenient
mode of communication. Where required, this will enable
operators of telephones, computers, etc. to call others, enter
data, request information and control systems when their hands
and eyes are busy, when they are in the dark, or when they are
unable to be stationary at a terminal.
One known approach to automatic speech recognition
involves the following: periodically sampling a bandpass
filtered (BPE') audio speech input signal to create frames of
data and then preprocessing the data to convert them to processed
frames of parametric values which are more suitable for speech
processing; storing a plurality of templates (each template
is a plurality of previously
--1-- ~ `~
, ,
4~8
~ S.E. Hutchlns et zl 1-1-4-3-3
cre~ted processed frames of parame-ric vallles repres~n~i-ng
a word, whicn when taken together form the reference
vocabulary of the automatic speech recogni2er); and
comparing the processed frames of speech with the
templates in accordance with a predetermined algorithml
such as the dynamic programming algorithm lDPA) described
in an article by F. Itakura, entitled "Minimum prediction
residual principle applied to speech recognition", IEEE
Tr~ns. Acoustics, Speech and Signal Processing, Vo.
ASSP-23, pp. 67-72, February 1975, to find the best time
alignment path or match between a given template and the
spoken word.
Aut-omatic Speech Recognizers depend on detecting tne
end points of speech based on measurements of energy.
Prior art speech activity detectors discriminate between
energy, assumed to be speech, and lack o- en.ergy, assumed
to be silence. Therefore, prior art Automatic Speech
Recognizers require a relatively quiet environment in
wnich to operate, otherwise, performance in terms of
recognition accuracy drops drastically. Requiring a auiet
environment restricts the uses to wnich a Speech
Recognizer can ~e put, for example, prior art recognizers
would have difficulty operating on a noisy factory floor
or in a coc~pit of a tactical aircraft, etc. Such noisy
environments as tnese can be characterized as having
bac~ground noise present whetner or not speecl~ is present
and noise events occurring when speecn is not present, tne
noise events sometimes having signal levels e~ual to or
greater than the speech signal levels. It is desiraDle,
therefore, to provide an apparatus and method for speaker
independent speech activity detection and for such speecn
activity detection for use in automatic speech recognizers
WhiCh must operate in an environment wherein noise events
with relatively high signal levels occur when speech is
not present.
SummarY of the Invention
The present invention relates to an apparatus and
--2--
~z~
, 5.E. ~utc~ins et ~ 3-3
me.hod for speech activity cetec.ion of s~eech sicnals in
the presence of noise, including noise events occurring
when speech is not present with signal le~els which may
have signal strengths e~ual to or greater than the speech
signals. The input signals are digitized and frames of
digital signal values associated with said digitized
signals are repeatedly formed. The speech signals and
noise event signals are automatically separated. In the
- preferred embodiment, this is done with a spea~er
0 independent predefined, fixed operation or transformation
performed on tne frames.
Also, in tne preferred embodiment, the input signals
are fre~uency filtered to provide a plurality of filter
output signals which are then digitized. The frames are
created from the digitized filter output signals. A
linear transformation is zpplied to the frames of digital
signal values to create a scal2r feature for each frame
whose magnitude will be larger for speech signals than for
noise event signals.
A detection tnreshold value is created for the scalar
feature magnitudes and repeatedly updated. Scalar
features are compared with the detection threshold value,
and the results of a plurality of successive comp2risons
are stored. The stored results are com~ined in a
2j predetermined manner to obtain an indication Ot when
speech signals are present.
When an indication that speech signals are present is
given, frames are further preprocessed before being
compared with stored templates representing the vocabulary
o~ recognizable words. The comparison is based on the
.ynamic prosramming algorithm (DPA).
Brief Description of the_Drawings
O~jects, features and advantages of the present
inven.ion will become more fully apparent from the
following de.ailed description of the preferred
emboàiment, the appended claims and the accompanying
drawings, in which:
~2~8~8
, S.E. :~;utcnins et al 1-1-4-3-3
Flg. 1 is a preLerred em~odiment bloc~ dia~r2~ of the
autom2tic speech recognition appar2tus of the present
invention.
Fig. 2 is a more detailed bloc~ diagram of the
bandpass filter portion of the invention of Fig. 1.
Fig. 3 is a table giving the filter characteristics
of the bandpass filter portion of Fig. 2.
Fig. 4 is a preferred emDodiment bloc~ diagram of the
operation of the speech recognition algorithm of the
present invention.
~ ig. 5 is a graph summarizing the time alignment and
matching of the recognition portion of the speech
recognition algoritnm of Fig. 4.
Fig. 6 shows three graphs of amplitude vs. frequency
for voice, jet noise and oxygen regulator noise.
Fig. 7 is a more detailed bloc~ diagram of the speech
activity detector portion of the speech recognition
algorithm of Fig. 4.
Detailed Description of tne Drawinqs
Fig. 1 is a bloc~ diagram of an automatic speech
recognizer apparatus designated generally 100. It
comprises a microphone 102; a microphone preamplifier
- circuit 104; a bandpass filter bank circuit 108 for
providing a digital spectrum sampling of the audio output
of circuit 104; a pair of processors 110 and 112
interconnected by inter-processor communication
circuits 114 and 116; and an external non-volatile memory
device 118. In the preferred embodiment, processors 110
and 112 are Motorola MC68000 microprocessors and
inter-processor communication circuits 114 and 116 are
conventionally designed circuits for handling interrupts
and data transfers between MC68000 microprocessors.
Interrupt procedures for the MC68000 are ade~uately
described in tne MC68000 specification.
The speech recognition algorithm is stored in the
EPROM memory portions 122 and 124 of the processors 110
and 112, respectively, wnile tne predefined vocabulary is
~Z184~8
S.E. Hutchins et âl 1~ -3-3
.
stored as previously created templates in the external
non-vola.ile memory device 118 which in the preferred
embodiment is an Intel bubble memory, Model No. 7110,
capable of storing one million bits. In the preferred
embodiment, there are only 36 words in the vocabulary,
and, hence, 36 templates with 4000 bits -equired per
template on the average. Hence, the bubble memory is
capa~le of storing approximately 250 templates. When
templates are needed for comparison with incoming frames
of speech data from BPF circuit 108, they are ~rousht from
memory 118 into working memory 126 in processor 112.
Referring now to Fig. 2, a more detailed block
diagram of the bandpass filter ~ank circuit 108 is shown.
The output from preamp 104 on lead 130 from Fig. 1 is
transmitted to an input amplifier stage 200 which has a
3 db bandwidth of lOkHz. This is followed ~y a 6 db per
octave preemphasis amplifier 202 having selectable cut in
frequencies of 500 or 5000 Hz. This is conventional
practice to provide more g2in at the higher frequencies
than at the lower frequencies since the higher frequencies
are generally lower in amplitude in speech data. At the
output of amplifier 202 the signal splits and is provided
to the inputs of anti-aliasing filters 204 (with a cutoff
frequency of 1.4 kHz) and 206 (with a cutoff frequency of
10.5 XHz). These are proviced to eliminate aliasing which
~ay result because of subsequent sampling.
The outputs of filters 20' and 206 are provided to
Dandpass ,ilter circuits (BPP) 208 and 210, respectively.
BPF 208 includes channels 1-9 while BPF 210 includes
cnannels 10-19. Each of channels 1~18 contains a
one/third octave filter. Channel 19 contains a full
octave filter. The channel filters are implemented in a
conventional manner using Reticon Model Numbers R5604 and
R56606 switched-capacitor devices. Fig. 3 gives tne clock
input frequency, center frequency and 3 db bandwidth of
tne 19 channels of the BPF circuits 208 and 210. The
bandpass filter clock frequency inputs required for the
BPF circuits 208 and 210 are generated in a conventional
--5--
.
~Z184~3
,C F~ r ch; rlS ~ ' i~
r,anner from a clock generator circuit 2i2 driven Dy a
1.632 ~.~z clock 21~.
The outputs of BPF circuits 208 and 210 are
rectified, iow pass filtered (cutoff frequency = 30 Hz)
and sampled simultaneously in 1~ sample and hold circuits
~National Semiconductor Model No. LF398) in sampling
circuitry 214. Tne 19 channel samples are then
multiplexed through multiplexers 216 and 218 (Siliconix
Model No. DG506) and converted from analog to digital
signals in lo~ A/D converter 220, a Siliconix device,
Model No. DF331. The converter 220 has an 8 bit serial
output which is converted to 2 parallel format in serial
to parallel register 222 (Nationzl Semiconductor Model
No. DM86LS62) for input to processor 110 via bus 132.
A 2 MHz cloc~ 224 generates various timing signals
for the circuitry 214, multiplexers 216 and 218 and for
A/D converter 220. A sample and hold,command is sent to
circuitry 214 once every io milliseconds over lead 215.
Then each of the sample and hold circuits is multiplexed
sequentially (one every 500 microseconds),in response to a
five bit selection'signal transmitted via bus 217 to
circuits 216 and 218 from timing circuit 226. Four bits
are used by each circuit while one bit is used to select
which circuit. It therefore takes 10 milliseconàs to A/D
convert 19 sampled channels plus a ground reference
sample. These 20 8-bit digital sisnals are callec a frame
of data and they zre-transmitted over bus 132 at
appropriate times to microprocessor 110. Once every frame
a status signal is generated from timing generator
circuit 226 and provided to processor 110 via lead 228.
This signal serves to sync the filter circuit 108 timing
to the processor 110 input. Timing generator circuiit 226
further provides a 2 kHz data reaày strobe via lead 230 to
processor 110. This provides 20 interrupt signals per
frame to processor 110.
Referring now to Fig. 4, a block diagram of the
automatic speech recognition algorithm 400 of the present
~18458
S.E H'tchins e. c~ 3-~
invention is presented. It can be divided into four
sub.asks: bancp2ss filter data transformation 402; speech
activity detection 404; variable frame rate encoding and
normalized mel-cepstral tr2nsformation 406; and
recognition 408. The speech activity detection
subtask 404 has been implemented in C lansuage for use on
2 VAX 11/780 and in assembly lansuage for use on an
MC68000. C language is a higher order language commonly
used in the technical community 2nd available from
Western Electric. The C language version of subtask 404
can be found on pages 12 throush 16 of the specification.
_ It will be described in more detail in connection with a
description of Fig. 7.
- As discussed earlier, every S00 microseconds the
microprocessor 110 is interrupted by the circuit 108 via
lead 230. The software which handles that interrupt is
the BPF transformation subtask 402. Usually, the new
8-bit filter value from bus 132 is stored into a buffer,
but every 10 miilisecond (the 20th interrupt) a new frame
sign21 is sent via lead 228. The BPF tr~nsformation
subt2sk 402 takes the 19 8-bit filter values that were
buffered, combines the first three values as the first
coefficient and the next two v21ues as the second
coefficient, and discards the l9th value beczuse it has
been founà to contain little if any useful information,
especizlly in a noisy environmen,. The resulting 15
coefficients characterize one 10 ms frame of the input
sign21
The transformed frzme of speech is passed onto
buffer 410 and then to the VERE and mel-ceostral
transformation subtask 406 if the speech activity detector
substask 404 has indicated that speech is present. The
speech activity detector subtask 404 will be explained in
more detail later. Assuming for tne moment that
su5task 404 indicates that speech is present, then in
subtask 406, the Euclidean distance between a previously
stored frame and tne current frame in buffer 410 is
determined. If the distance is small (large similarly)
i~ -7-
~Z18~8
and not more than two frames of data have been skipped, the
current frame is passed over, otherwise it is stored for future
comparison and passed onto the next step of normalized mel-
cepstral transformation. On the average one-half of the data
frames from the circuit 108 are passed on (i.e. 50 frames per
second).
To reduce the data to be processed, the 15 filter
coefficients are reduced to 5 coefficients by a linear
transformation matrix. A commonly used matrix comprises a
family of 5 "mel-cosine" vectors that transform the bandpass
filter data into an approximation of "mel-cepstral" coefficients.
Mel-cosine linear transformations are discussed in (1) Davis,
S. B. and Mermelstein, P. "Evaluation of Acoustic Parameters for
Monosyllable Word Identification", Journal Acoust. Soc. Am.,
Vol. 64, Suppl. 1, pp. S180-181, Fall 1978 (Abstract) and (2)
S. Davis and P. Mermelstein "Comparison of Parameter
Representations for Monosyllabic Word Recognition in Continuously
Spoken Sentences", IEEE Trans. Acoust., Speech, Signal Proc.,
Vol. ASSP-28, pp. 357-366. However, in the preferred embodiment
of the present invention, a variation on "mel-cosine" linear
transformation is used called normalized mel-cepstral transform-
ation, i.e., the raw BPF data is normalized to zero mean,
normalized to zero net slope above 500 Hz and mel-cosine
transformed in one step. The first mel-cepstral coefficient
(which is very sensitive to spectral slope) is not used.
Each frame which has undergone mel-cepstral
transformation is then compared with each of the templates
representing the vocabulary which are now stored in the
processor's working memory 126. The comparison is done in
accordance with a recognition portion 408 of an algorithm based
on the well-known dynamic programming algorithm (DPA) which is
--8--
~Z~8~L58
described in an article by F. Itakura entitled "Minimum Predic-
tion Residual Principle Applied to Speech Recognition", IEEE
Trans. Acoustics, Speech and Signal Processing, Vol. ASSP-23,
pp. 67-72, February 1975. A modified version of the DPA
may be used, called a windowed DPA with path boundary control.
A summary of the DPA is provided in connection with a description
of Fig. 5. A template is placed on the y-axis 502 and the input
word to be recognized is placed on the x-axis 504 to form a DPA
matrix 500. Every cell in the matrix corresponds to a one-to-one
mapping of a template frame with a word frame. Any time align-
ment between the frames of these patterns can be represented by
a path through the matrix from the lower-left corner to the
upper-right corner. A typical alignment path 506 is shown.
The DPA function finds the locally optimal path through the
matrix by progressively finding the best path to each cell, D,
in the matrix by extending the best path ending in the three
adjacent cells labeled by variables, A, B, and C. The path
that has the minimum score is selected to be extended to D
subject to the local path contraint: every horizontal or
vertical step must be followed by a diagonal step. For example,
if a vertical step was made into cell C, the path at cell C
cannot be chosen as the best path to cell D. The path score
at cell D is updated with the previous path score (from A, B,
or C) plus the frame-to-frame distance at cell D. This distance
is doubled before adding if a diagonal step was chosen to aid
in path score normalization. The movement of the DPA function
is along the template axis for each utterance frame. The func-
tion just described is repeated in the innermost loop of the
recognition algorithm by resetting the B variable to cell D's
score, the A variable to cell C's score and retrieving from
storage a new value for C.
However, before the subtasks 406 and 408 can operate,
the beginning and end of speech must be detected. Where
_g_
~841~8
S.B. ~ tcnins et al 1-1-4-3-3
s?eech recosni.ion is ta~ing place in a quiet environment
witn lit,le or no noise present, endpoint detection b2sed
on energy measurement can be used. However, in the
environment of tactical fighters, for example, there are
present two types of noise wnich render traditional speech
activity detectors useless. Background noise from engines
and wind is added to the speecn signal and results in tne
classical detection problem of separating signal and
additive nuise. See curve ~02 in Fig. 6. The use of an
oxygen regulator with a mask introduces noise from inhales
and exhales which are not concurrent with speech but
resemble speech in spectral shape and can cause spurious
deiection. See Curves 60~ and 606, respectively. The
amplitudes of the signals associated with these noise
events often exceed the speecn signal amplitudes in many
coc~pit conditions.
~ eferring now to Fig. 7, a more detailed description
of the speech activity detection subtas~ 404 is given. A
large number of frames or data ~rom subtask 4D2
representing both speech and noise event sounds from a
variety of spea~ers and oxygen regulators were studied to
determine a fixed transformation which when applied to the
frames would provide a good separation between speech and
noise events over a range of s?eakers. It was` determined
that a single 15 parameter feature vector 702 could be
found whose inner product 703 with modified frames 704
derived from the bandpass filter frame 70; would proviàe a
scalar feature 706 giving good separation of speech from
noise events. The frames coming from the BPF
transformation subtas~ 402 are logarithmically encodeQ
frames àue to the action of the log AjD Converter 220.
Better results are achieved, however, if frames
proportional to the energy of the noise event signals and
speech signals are formed. This is accomplished by
modifying the BPF frames from 705 via the operation of
squaring the inverse log of the frame components 707.
This step enhances speech activity detection by increasing
-10 -
~2~89L~8
S.E. -u.c~ ns et ai 1~ 3-3
the cynamic r2nse of the fe2turec, thus providing srec.er
separation between the pea~s of the speech spectra ana the
relatively broad band noise and non-speecn spectra.
To derive a good feature vector F, a collection of
frames of BPF data from a plur21ity of spea~ers and noise
events occurring when speech is not present are collected
and modified as described above. The data is divided into
sets of speech frames [S] and noise event frames [N]. By
inspection, a good intuitive guess at F is made and then
in accordance with the equation below, the inner products
of F with all of IS] and all of [N] is formed, and the
statistical overlap of the resulting two classes of scalar
features, [F S] and ¦F.N] is measured to form a separation
figure of merit. (- represents forming the inner product
of the two vectors.)
Separation = Mean (~F.S]) - Mean (IF-N])
Std Dev (lF.S]) I Std Dev ([F N])
Small changes in each of the feature veçtor components,
fj is made, for example, tne first component, fl, of F
is made a little larger and then a little smaller, then
the same is done for f2 and so on. For each small
cnange F.S and F.N is recomputed for all the frames lS]
and [N] and the separation remeasured. This i~entlfies
the direction to ta~e to cnange F for be-ter separation.
F is changed accordingly, obtaining a new vector for a
starting point and then the process is repeated. This
approach is ~nown as a gradient search.
When a feature vector F is formed ~nich appears to be
a significant improvement, it is tried in the recognizer
algoritnm to see how it wor~s. If certain types of noise
events are found to still trigger the detection, or if
certain speech sounds are consistently missed, samples of
them are ta~en and added to the data base ~S] and [N~.
Then a new feature vector is searched for that handles the
new data as well as the old.
12~84~.8
s.r. ~:~.chins et al 1-1-4-3-,
To assist in carrying out all the inner proauc~ and
sep~ration com~u.ations re~uired during the gradient
search, a program W25 created in C language for a VAX
computer. A listins of the program for a slightly
modified gradient searcn from that described a~ove is
founo on pages 21 to 23 of tne specification.
The preferred embodiment, 15 parameter feature vector
found by the gradient search as substanti~lly described
above is,
0 .0
2 13.9
3 5.9
4 1.2
1.4
6 1.4
7 1.5
8 1.6
9 2.4
1.3
11 2.0
12 1.2
13 4.8
14 -13.6
0.0
Once the optimum feature vector is determined, the
result nt scalar features formed by the inner product
operation with the modified frames are collected and
formed into a histogram designated generally 710 in
Fig. 7. The x-axis 712 is the magnitude of the scalar
feature while the y-axis 714 is the number of times a
particular magnitude occurs. Jet noise 716 and regulator
lZ1~4S~3
5.E. :.~.cnins et al 1~ -3-3
sounds 718 occu below z th.eshol2 720 while voice 727
occurs above the threshold 720.
When the speech recognizer is being used, e.g., in
flight in an aircr2ft cockpit, t~e speech 2ctivity
detection subtask 404 initially selects a detection
threshold but tnereafter continually gathers statistics
and updates the histogram on the feature 726. Every
1000 frames, the detection threshold is adjusted based on
the statistics in the histogram. For example, the
peak 750 is located in the histogram 710, and a search is
cond~cted forward from the peak 750 to locate the low
point 720. Tne threshold is set to the low point value
plus some bias sucn 2S one or two. Finally, each
histosram entry is divided by two to keep the histogram
values from growing t-oo large.
The magnitude of the detection threshold 708 is
subtrac.ed from the magnitude of the scalar feature 706 at
block 730 for each frame. A weighting function 732 is
applied to the outp~t vzlue of block 730 to smooth out the
values before they are filtered and clamped at 73~. The
weighting function reduces Iarge negative values from
block 730 and reduces smali positive values. Large
positive values are left substantially unaffected. The
weiahting function cooperates with the integra.ion process
performed by the filter and clamp function 734~to proviàe
sharp c~toff points between the beginning and end of
speech detection. Large nesztive values provide nc better
indication of non-speech than smaller values, b~t will
distort and delay the integration process from inGicating
when speech is present. Small positive values create
uncertainty ~s o whether speech is present and are better
left undetec~ed. An example of the preferred embodiment
weighting function and filter and clamping functions are
provided in C language on page 19 of the specification.
Four values from filter znd clamp 734 corresponding
to four successive frames from subtàsk 402 are stored in
-13-
~Z1891L~8
~ c~ s ~ 3~3
b~Lfers 736. Then multi-frzrle decision logic 738 is
employed to make a aecision whether speech is present.
For example, if no speech were present and if all four
buffers provide a positive indication, then a decision is
made that speech is present, and this is passed on to
block ~10 in Fig. 4, otnerwise a decision is made that
speech still is not present. On the other hand, if speech
is currently present, a decision is made tnat speech is
still present if any one of the buffers indicates that a
speech signal is present. Only if 211 four buffers
indicate no speech signals present will 2 decision be made
that speech is now over. The above-described decoding is
provided in C language at pages 19 and 20 of the
specification
It should be noted that in the preferred em30diment,
subtasks 402, 404 and 4~6 are performed in processor 110
while subtask 408-is performed in processor 112. ~owever,
there is no reason why the two processors could not be
combined as one. Although the present invention relates
to a 36 word vocabulary with isolated word recognition,
there is no reason why the speech activity detector could
not be used witn larser vocabulary continuous speech
recognition machines. Also, speech activity detection
through the use of the inner product between 2 predefined
feature vector and frames of speech can be perforr1ed on
frames of speech provided directly from the bandpcss
filter transformation subtask 402 even thOUgh thi~ frame
is proportional to the log of the value of the disital
signals. Similarly, the inner product could be performed
using frames whcse digital sisnals are proportional to the
masnitude of the digital signals and not the magnitude
squared.
Results to date on the performance of the recognizer
indicate recognition accuracy of ~5 to 95~ for worst cases
of cockpit sound pressur.e level.of 115 dB and acceleration
forces of 5G. In fact, the system shows no degradation
from low level ambient noise performance (95+~ accuracy)
~Z~8~S8
S.~ chins et al 1-1-4-3-3
.o noise levelc cf 2pproximately- 106 dB. It should be
pointed out, however, that the 115 dB sound levels at 5G
acceler2tion forces are often simulated. The pilot is
speaking into an oxygen regulator which parti211y seals
off the ambient cockpit noise. However, the stress of the
noise and acceleration forces causes the pilot to speak in
a less than normal speaking m~nner. Also, the noise
events caused by the stressed breathing of the pilot into
the oxygen regulator are also present.
REL:jn
June 28, 1983
- - -
S . E . Hutchins et al 1-1-4-3-3
epsqnew. c epsqnew. c
-
prosram n~me: epsqnew.c
cLassiJ~cc~um of voiced speech vs noise usir4g
bp~ energy in bir s 0 thTough 1 9 (0 &1 9 are usucl ly ~7r ored)
this is true energy (inverse mu squared~ ver~ion
4ptive voicing threshoLd and weighting
derived from a rolling histogram
r~eeds 10 seconds of noise or voice and roise to
ada~t to environment
nction: reads in ra7~J PP~le and durr~ps decisions
to ffLe of your choice
k)ad vw~ cc epsqnew.c--o epsqnew--lm
Input format is r the form.
ep /speech~csr/tac /neLson/wc95 0 33000 < da~a ep
"data.ep" is an input datafile with the folLowirg form~;
bias: detection threshold bias
thr; a~ptive threshoLcL
PcLamp. Positive cLamp vaLue
Ncl,amp. I~eg~;ive cLamp vaLue
options. threshoLd offset vaLue
dipdenom: determines ~of pe~Jc th~t defines aDpro2ima~e vaLley
coefs: 16 linear BP~ coe,fflcients
sc4Le. scale ractor for coefs
Or~in~L by: S.Æ. Hutchins.. va~ious mods in 1932
G7an~ed to reflect 68000 a~t~metic: S. S4~ew~rt and H.Koble
May, June I se3
~/
lude Cstdio.h>
#i~clude ~cu~es.h>
~define DIM 15/~number of features~/
main(argc~ergv) mai~
~Dt argc; char ~argv[];
/~voicing decision va~bles~i'
Rhart buf20r20] /* r~v~ bpf dat~ ~/
int stat~10~; /* rolLing nistogram ~/
~nt vyes,vyeso,vyesol,v,Yeso2lyyeso3; /~ voicing indicators ~/
illtthr,ncount; /~ threshold,frame counter ~/
~nt sphfig,glfflg,cnvcnt,ctpc;
int cdbar,cdsig; /* st~ts on ~ssumed voicing */
~llt Lmp,x,peak,decisiou; /~ temp variabLes and flr~l decision ~/
intPclamp- /~Posi~ive VyesocL~npvaLue ~/
~nt Nclamp, /~ Negative ~yeso cLamp va,ue ~/
int options /* Thre shoLd o~set vaLue /
intbias; /~ bvls forhistogramLowpoint:+4forchestreg, else a~
/~ logic for word and syLlaole lengths ~/
int state,oldstate,tempframe,fromframe,toframe;
int buf[20] /~ new op~ da~a ~/
1ung cd; /~ ,-aw fe~ure ~/
~oat ampscale,tf; /* used to v~eight decisi,on ~/
flDat scale; /~ scc,Les up coef ~rr~y ~/
P~ge 1 Df epsqne~.c
-16 -
~Z~84~3
S.E. Hutchins eL al 1-1-4-3-3
epsqnew. c epsqnew. c
. . . m~Ln
~at dipdenom; /~ ~etermines ~ of pe~c that ciefines relative valley * /
floatcoef[DlM~ buffer offea~ure coefiicients A/
fioat imu;
float imusq[128]-
double c,log(),exp();~t housekeep~ng andplotting vari~bles ~/
int i,j,frarnecount,firstfr.lastfr-
char filename[ôO];
F,LE ~fp,fptag;
* ~rgv,ments * /
if(argc < 2) ¦
printf("usage: 70S BPF--file ~firstframe~ ~lastframe~\n".argv[0~);
exit();
if((fp = fopen(argV[l],"r")) == o)
exit(printf("Unable to open %s~n",argv[l]));
if(argc > 2) firstfr = atoi(argv[2 )- else firstfr = 0;
if(argc > 3) lastfr = atoi(argv[3J ; else lastfr = ~0000;
printf('~That is the name of the output file?~n"); fflush(stdin);
gets(filename) -
if((fptag = fopén(filename,'W')) ~ 0) ~
printf("unable to open or create file. 7,s~n",filename);
exit~);
I
printf("Enter bias~n"~;
scanf(" %d",&bias); /~ detection bias (1 is good start) ~/ -
printf("Enter threshold~n")- fflush(stdout)
scanf(" %d",&thr); / adaptive threshold (3 is good start) ~/
printf("Enter positive clamp~n~l)
scanf(" %d",~cPclamp); /~ clamp value (12 is 68000 v~lv,e) /
printf("Enter negative clamp\n")
scanf(" %d",~Nclamp); / clampva~ue (--~is 6f~000va~ue~
printf("Enter options~n");
scarf(" %d",8coptions); /~ threshold offset (3 is good st~rt) ~/
printf("Enter dipdenom~n")
scanf(" 7Df",d~dipdenom); /~ deter~nines 7peak th~t de~nes reiative valley
for(i=O;i<DI~,i++) ~
printf("Enter coef[70d~n",i)-
scanf(" Xf",&coefli]);
'i . '
cabar = thr + options; /~ guessesfor center uf voicing ~/
cdsig = 2; /4 and average devi~tion /
printf("Enter coef scale factor~n"~;scanf(" 70f",~scale);
for~j=O;j<DI~;j++) ~
coef~ = scale;
~rintf("coefLXd~ = 7f\n",j,coef[~']); ~7
vyeso = 0; vyesol = 0; /~ initialize decision memorres ~/
vyeso2 = 0; vyeso3 = 0;
state = 0; oldstate = 0; /~ ini~lize voicing logic ~/
cnvcnt = 3;
sphflg = 0; glfflg = 0;
fromframe = -40; tofrarne =--40;
frameconnt = 0;
Poge 2 of ep~qne?l). c
-17 -
lZ18~8
epsqrlew. c epsqne~. c
n
for(x=C;x<100;x . +) stat~x] = 0; /~ clear histogram ~/
fprintf(fptag,"file processed = 70s~n\r,",argvll])
fprintf~fptag,"bias = 7d options = %d dipdenorl = 7f\n\n",bias,opticns,dipdenom);
fprintf(i`ptag,"cdsig = %d starting thr = 7Od\n\n",cdsig,th~)
fprintf(fptag,"coef array\n");
for(j=O;j<DI~;j++)
fprintf(fptag,"78.2f ",coef[j]);
fprintf(fptag,"~n\n");
fseek(fp,firstfr~s~zeof(buf),0); /*skip tofirstframe ~/
/~ make imusq ~rray ~/
c = log(256.0)/127.0;
for(i=O;i<120;i+.) ~
imu = (exp(c~ 1.0) i85.0;
imusq~i] = (~nt)( 32767.0~imu~imu/9.0 );
Start ~rame Loop ~/ ~
for (i=firstfr;i<=lastfr;i++) l
if(fread(buf20,s~zeof buf20,1,fp) <= 0) break;
/~,find neqlJ threshoLd once every l, OOOframes ~/
ncount = framecount 7 1000;
if(ncount == 999) ~
peak=0; /~ note: histogrGm inde2 = real value , 49 t /
for(x=O;x< 100;x++) /~ find peak ~/
if(stat[x] > peak) l
peak = stat[x];
thr = x;
x = thr; /t 2 = peak ~ocation ~
peak = peak/dipdenom; /~ take some ~of realpe~k ~/
lvhile(stat[x] > peak) x++; /~findcorrespondi77g ~r value ~/
thr = x--49 + bias; /* new threshold = appro~imate valley - offset + bias ~/
cdbar = thr + options;
~ diagnos~cs ~/ printf("\n stats ~n");
for(x-O;x<25;x+ ' )printf(" %d ",stat[x]);
for(x=25;x<50;x++)printf(" 70d ",stat[x]);
printf("\n");
for~x=50;x<74;x++)printf(" %d ",stat[x]);
printf("~n");
for(x=75;x<100;x++)printf(" %d ",stat[x]);
printf("~n thr= %d ~n",thr)-
~* end ~gnostics ~/ prir.tf~"cdbar = 70d cdsig = 70d ~n",cdbar,cdsig);
for(x=O;x< 100;x++) /~ shi~t histo~rams do~n ~/stat[x] = stat[x] 1;
/~ end statistics update------begi,n frame--by--frame ~ork */
for(j=1;j<19;j++) 1
buf20[j] = buf20[j] .1
if(buf20[j] < 0) buf20[j] = 0;
if(buf20[j] > 127) buf20t;] = 127;
buf[0]=(b~f20[1] buf20[2]+buf20[3])/4
buf[ 1]=(buf20[4]+buf20[~]) /2;
for~j=2;j<DI~;j++)
bufLj]=buf20[j~4];
cd = 0;
Pa~e 3 of EpSg~E~) C
-18 -
~LZ~L8~S~3
epsqnew.c S.E. Hutchins et al l~le~ e'' C
.. n~in
~or~j=C;j<DI~ ) cd += coer[j]*imusq~ buf[jl ];
/~ i,f((framecount ~ 50~ == 0) ~
~"intJ'("cd = Z~d\n",cd)
~/ 1 for~=oli<DIM~i++)printf~li7nusq[b?~f[7d]] = ~f~n",j,imusq~ buf~ ]);
cd = cd 16; /~ jvst keep most signi~cant word of result ~/
vyes = cd, /~ ranse:--25-->25for most data ~/
if(vyes < -49) vyes = -49- /~ but some noise can go far negative ~/
iE(vyes > 50) vyes = ~o; /~ vyes is most of the decision ~/
cd = vyes; /~ clamp vyes as vJell ~/
statLvyes + 49]++; /~ accumulate in the histogram ~/
vyes = (~yes--thr); ~ minv,s variable threshold here ~/
/~ weighting based on distance from center of voice--like data ~/
tmp = cd--cdbar;
tf = (~at)(tmp)/(float)cdsig; /~ d~mp down score if beLow mean t/
ampscale = 1.0; /~J'orvoice--like thir~gs ~/
if(tmp < 0) ampscale = 4.0/(tf~tf + 4.0); /~ weight is 0. 80 if off 1 std dev */vyes = vyes ~ ampscale; /* raw weishted voicins decision ~/
vyeso = 7 ~ ~eso ~ 8 + vyes; /o smooth the decision ~/
if(vyeso > Pclamp) ~ /~ clamp to help word endi77,ss drop of~ ~/
vyeso = Pclamp
/~ diagnositcs ~/ printf("+ );
iE(vyeso < Nclamp) ~ /- clamp to help word starts cor~e ~p ~/
vyeso = Nclamp;
/~ diasnostics ~/ printf("-");
oldstate = state;
decision = 0; /~ ~ook atfourframes to decide ~/
if(vyeso > 0) decision = decision + 1
if(vyesol > 0~ decision = decision + ;;
if(vyeso2 > 0~ decision = decision + 1
if(vyeso3 > 0) decision = decision + 1; /~ decision=4for 4 hits in a . ow ~/
vyeso3 = vyeso2;
vyeso2 = vyesol; /~ decision = 4 is the voicins start trigger */
vyesol = vyeso; /~ it means this is the 4th vo*edframe ~/
/~ once voicing sta~ts decision>0 contir~ues it ~/
state = 0;
if(oldstate == 0 d~& decision == ~) state = 1;
if(oldstate == 1~& decision > 0) state - 1;
i~(state == 1~; sphflg == 0)
sphflg = 1;
ctpc = -6;
iE(state == 1) ~
ctpc = ctpc + 1;
cnvcnt = 3;
I(state == 0 &~ sphfig > O && ctpc < 0) glfflg = 25~;
if(state == 0 && sphfig > 0 && ctpc >= 0)
cnvcnt = cnvcnt + l;
i~(cnvcnt >= 32) glf~lg = 1;
i~(glf~lg == 0 && sphflg == 1) ~
-19- Pa3e ~ of epSq?~E7~1.C
,, .
12~S~3
S . E . Hutchins et al 1-1-4-3 -3
epsql~ e~. r epsqne ~ . ~
. m~in
sphfig = 255;
fromframe = frar2ecc~r.t -2;
i~(glf~lg == 255)
spl~fig = O;
glfflg = O;
if(gl~lg == 1) ~
toframe = framecount--32
sphflg = O;
glf~lg = O;
fprintf (fptag,"%d 7Dd ~n",fromframe, toframe);
framecount+ ';
Pr~e ~of e?sqne~.c
--20 ~
~Z~8~
S . E . Hutchins et al l -1-4-3-3
agc~tats.c agcs~;s c
/~ Computes cost fur-tion for evaluating a set
o,f bpf coeficient u~eiJ~hts ard perforrri?lg ~ gra~lient search
J'or neighboring vectors ~h betterperformance
~put: files o~f sta~stics on speech ard noise and ini~l ~eight vectors
- Loa~via: cc agcstats.c--o agcstats -g -Im ~/
ritten by A. Higgins, Oct. 1981 ~/
odiffedfor tactical recog. agc by H.Koble, Aug. 19~Z ~/
/~ Completely gutted by Steven Ste?l)art",~larch lg83 /
/t commer~ts added 6/22/83 by S.E.Hutchins ~/
lude "stdio.h"
#include "math.h"
~define DIM 15 /~number offeatures,i.e. dimer~sion of feature vector~/
f3Dat covl[DIM][D~],cov2[Dr~]~D~a]; /~tspeech & no*e covariances~/
l~at meanl~DIM],~ean2~DIM~ speech & noise means~/
main(argc,argv) ma?,n
intargc; ~ argv~];
FILE Ifpinl,~fpin2,~fpwt;
~rLt fdout,i,j,k,n,ii,dim,Rag;
int buff[DlM];
flDat stp,best;
S~at temp[2];
fiDat weight[DLM],newwt[DIM],wt[DI~];
double cost();
char filename[80];
/~ operL requestedJiles of statistics~/ -
i~(argc < 4) ~
printf("Usage YOs stat_flle1, 1 stat_file7,2 ueight file\n",argv[0]);
exit();
~f((fpinl = fopen(argv[l],"r")) == NULL) I
printf("Unable to open input file: 7s~n",argv[1]);
exit();
if(~fpin2 = fopen(argv[2],"r")) == NULL) ~
printf("Unable to open input file: 7s\n",argv~2~);
exit();
/~ openfile of starting vectors ~/
if((fpwt = fopen(argv[3],"r")) == ~rULL) ~
printf("~1nable to open weights file: ,Os\n",argv[3]);
exit();
/~ printf~ hatis~henarrLe ofthe outputfile?");ffl~sh~stdout);
gets (filer ame);
if((fdout = creat(filer~ame,O644)) < O) ~
printf~'unable to open or create file: 7s\n",filenameJ;
e~it();
read speech statistics~ /
i~(fread(meanl,s~ meanl,l,fpinl) <= C) ~
printf("unab~e to read n~eanl data\n");
exit();
if(fread(covl,s~ze~f covl,l,fpir,l~ <= 0) ~
printf("~mable to read covl data\r");
exit();
Page 1 of a3cstats. c
-21 -
~Z~8~5~3
-- S . E . Hutchins et al 1-1~3 -3
a~cstats c agc ats.c
. . . ma~n
/~read noise statistics~ /
if(fread(mean2,siYeo~ mean2,1,fpin2) <= 0) ~
pr?ntf("unable to read mean2 data\n");
eXIt();
i~fread(cov2,sizeo~cov2,1,fpin2) ~= 0) 1
print~("unable to read cov2 data\n");
exit();
printf("Enter step size\n"); /~step size for vectorperturbation~/
scanf(" 7~,f",&stp);
/~read aweLght vector~/
for(j=O;j<DlM;j++) flag=fscanf(fpwt,"7Df",&weight[j]);
if(flag <= 0) goka endit;
fe~r(j=O;j<DlM;j++) ne~nrt[j] = weight[j];
best = cost(newwt); /~ best = cost(startin~ vector)$/
/~ie measure of separation giverl, by initial vector~/
iterate hl~enty times... each time perturb each vector elemer~/
for(n=O;nC20;n++)
for(k=l;k<DI~;k++) ~ /~for each dimension~/
for(ii=O;ii<2;ii++) ~ /~try T--step value~/
~(ii == O)
newwt[k] = weight[k]--stp;
cLse
ne-t~-t[k] = weight[k] + stp;
temp[iil = cost(newwt);
/~pick the chc~nge for no chGnge~
/~with the best separation ~/
i~(temp[~] > temp[ 1])
if(temp[0] > best) ~
new rt[k] = weight[k]--stp;
~eight[k] = weight[k]--stp;
best = temp[0];
eLse
ne~nrt[k] = weight[k];
~Ise
if(temp[ 1] > best) ~
new;Yttkl = weighttkl + stp;
weightLk~ = weightLkJ + stp;
best = temp[1~;
else
ne n~t[k] = ~Yeight[k];
/~ output st~s to fUe~/
endit dim= DIM-
rite ffctout, ~im, sizeof (dim));
write(fdout,weight,sizeof(weight));
write(fdov,t, best,sizeof(best));
close lfdout);/
fclose(fpinl~;
fclose(fpin2);
fclose(fpwt);
printf("~nDone~n");
png~ 2 of agcst ats. o
-22 -
~2~ 8
agcstats. c agcs'~a;~s. c
. . , ~L.a,in
~/~en3 mair~/
double ccs~(ne~-t) ~ou~e
/~ comp?~te the stal~s~al separat~ion of ~he t~o ~/
classes cl' dsta g~ve7~ the cun ent vector~/
~:at ~ne~
iQt i,j;
noat alpha
~at wt[DI~];
~at sqs,upper,x,y,su.n;
double ~abs(),sqrt();
sum= 0;
far(j=O;i<D~ + )
sum+= newwtLj]~new~t[j];
sqs= sqrt((double)sum);
for~j=O;j<DlM;j++~
wt~]= newwt[~/sqs;
upper=0
or(j=O;j<DIM;j++)
upper+= wt[j]~t(meaIll[j]-mean2[j]);
x=O;
fcr(i=O;i<DIM;i++)
sum=O
for(j=O;j<DIM;j++)
sum+= ~t[j]~covl[j][i]:-
x+= sum~r,t[i];
x= sqrt((doll~le)x);
y=O;
fc~r~i=O;i<DI~;i++)
surn=0
for(j=O;j<DIM;j++)
sum+= ~t~j]~cov2~j][i];
y+= sum~t[i];
y= sqrt((dou~le)y);
alpha = upper/(x+y);
printf("~n\n~3PF CoefIicient Weights:");
for(j=O;j<DIM;j++)
printf(" 7f ",lO.O~wt~j]);
printf("~n\nCost FuncticrL");
printf("79.4f",fabs(a~pha));
~tunl(fabs(alpha));
While the present invention has been disclosed in
connection with the preferred embodiment thereof, it
should be understood that there may be other embodiments
which fall within the spirit and scope of the invention
as defined and by the following claims.
- - 2 3 - Page 3 ~f agcstats c
