JPH0424717B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field) The present invention relates to a speech recognition device, and particularly to speech section detection for detecting speech sections with high accuracy.

(Background Art) A block diagram of a conventional speech recognition device is shown in FIG. In FIG. 1, 1 is a signal input terminal, 2 is a frequency analysis section, 3 is an audio capture control section, 4 is a capture start signal, 5 is a voice section detection section, 6 is a capture end signal,
7 is a starting end/end end information, 8 is a recognition section, and 9 is an output terminal. Each section will be explained below.

The frequency analysis unit 2 is configured as shown in FIG.
is analyzed by N band-pass filter groups 13, full-wave rectifier groups 14, and low-pass filter groups 15, which are divided at equal intervals on a logarithmic scale. The data are quantized by the AD converter 17 while sequentially switching the multiplexer 16 at each sample period, and N analysis results 18 are output at each sample period.

After receiving the capture start signal 4, the voice capture control unit 3 transmits the analysis result 18 of the frequency analysis unit 2 to the voice section detection unit 5 and the recognition unit 8 for a certain period of time, or when the voice input has certainly finished. Output until a decision is made. As a method for determining the end of audio input, for example, the average value of N pieces of data for each sample period (hereinafter referred to as frame power) is used, and if the frame power exceeds a preset threshold, There is a method of determining that the voice input has ended when a certain number of values do not exceed the threshold value and a certain number of consecutive values do not exceed the threshold value.

A block diagram of the voice section detection section 5 is shown in FIG. In Figure 3, 18 is the analysis result, 21
Reference numeral numeral 6 indicates a parameter calculation unit, numeral 6 indicates a capture end signal, numeral 22 indicates a blocking unit, numeral 23 indicates a voice section determination unit, and numeral 7 indicates start/end information, which will be described in detail below.

The parameter calculation unit 21 is a part that calculates parameters defined by equation (1) used for voice section detection from the analysis result 18.

P _j = a _j・_j …(1) where a _j ; Spectral slope _j of the jth analysis result; Average value of the jth analysis result Also, the spectral slope a _j, that is, the slope of the least squares approximation straight line is , the j-th N analysis results are x _ij
Assuming that (i; the number assigned in order from the lowest frequency to the group of N-divided bandpass filters),
a _j is determined by equation (2).

If N is fixed in equation (2), _N 〓 ⁱ⁼¹ i and _N 〓 ⁱ⁼¹ i ² become constants, and C ₁ = _N 〓 ⁱ⁼¹ i and C ₂ = N・_N 〓 ⁱ⁼¹ i ² −{ _N 〓 ⁱ⁼¹ i} ² can be replaced, and equation (2) is transformed into equation (3).

Therefore, a _j can be found by finding _N 〓 ⁱ⁼¹ i・x _ij and _N 〓 ⁱ⁼¹ x _ij .

Also, _j can be obtained by dividing _N 〓 ⁱ⁼¹ x _ij by N. FIG. 4 is a block diagram for calculating P _j , which will be explained below with reference to the figure.

j-th N analysis results x _ij (i=1, 2,...
N) are output in order, adder 1
01 and register 102, the cumulative value of x _ij is _N 〓 ⁱ⁼¹
x _ij can be set in register 102,
The result is output to multiplier 103 and divider 106. The multiplier 103 multiplies _N 〓 ⁱ⁼¹ x _ij by C ₁ (= _N 〓 ⁱ⁼¹ i), and the complementer 104 multiplies -
The value of C ₁ · _N 〓 ⁱ⁼¹ x _ij is determined and input to one side of the adder 105 . Further, the product i x _ij of the output of the counter 107 that operates in synchronization with the data output of x _ij and x _ij is obtained by the multiplier 108, and the product is calculated by the adder 109 connected to the output of the multiplier 108. , furthermore, by the register 110 connected to it, _N 〓 ⁱ⁼¹ i・
x _ij can be found. The output _N 〓 ⁱ⁼¹ i・x _ij of the register 110 is connected to one input of the multiplier 111, and the other input of the multiplier 111 is set to _N. 〓 ⁱ⁼¹
i·x _ij is calculated and input to the other side of the adder 105. In the adder 105, N・_N 〓 ⁱ⁼¹ i・x _ij
is calculated and connected to the divider 112. In the divider 112, N・_N 〓 ⁱ⁼¹ i・x _ij −C ₁・_N 〓 ⁱ⁼¹ i・x _ij
By dividing C ₂ by C 2 , the spectral slope a _j of the j-th sample data is obtained, and the result becomes one input of the multiplier 113 . Further, in the divider 106, _j is obtained by dividing _N 〓 ⁱ⁼¹ x _ij by N, and the result becomes the other input of the multiplier 113 _{. j}・_j ) can be obtained. By performing the above calculation for each sampling period, all the values of P _j at each sampling time can be calculated.

The blocking unit 22 receives the result P _j of the parameter calculation unit 21 until it detects the capture end signal 6, and after detecting the capture end signal 6, blocks the audio (blocks the parts that are considered to be audio). detection)
FIG. 5 shows a block diagram of the portion where this is carried out, and the explanation will be given with reference to FIG.

P _j for each sample period of the parameter calculation unit 21
are sequentially stored in the P parameter memory 200, so they are read in order and the absolute value circuit 201
|P _j | is input to one side of the comparator 202. At the other input of the comparator 202, a threshold value P _TH of |P _j | is set. The comparator 202 outputs α when |P _j |≧P _TH ;
When P _j |<P _TH , a significant signal is output to each β output. The counter 203 counts up when |P _j |≧P _TH , and is cleared when |P _j |<P _TH , and counts the continuous amount that satisfies |P _j |≧P _TH . Further, the output of the counter 203 is always set in the register 204. The value set in the register 204 (|P _j |≧
P _TH (consecutive number) is input to the comparator 205,
K is set to the other input of the comparator 205, and when the continuous amount (hereinafter referred to as block length) where |P _j |≧P _TH is greater than or equal to K, the output C of the comparator 205
A significant signal is output.

The block length is K (a natural number of K≧2) or more (when outputting C signal), and the β output of the comparator 202 (|
The timing at which P _j |<P _TH ) appears is determined by AND circuit 2.
Captured by 06. The counter 207 is AND
It counts the amount of P _j read from the output of the circuit 206, and by subtracting the result of the register 204 (block length) from the output of the counter 7 by the subtracter 208, the distance (time) between blocks is calculated. can be found. Also counter 2
09 is counted in synchronization with the reading of _Pj , and by subtracting the output (block length) of the register 204 from the result of the counter 209 by the subtracter 210, the head of the block can be determined. By the adder 211 and the register 212, |P _j |
≧P _TH is calculated, S _B representing the block size is calculated, and when the signal of the AND circuit 206 is detected, it is set in the register 213 and at the same time the output of the register 213 (hereinafter referred to as block amount) is calculated. ), the output of the subtracter 210 (block head information), the output of the register 204 (block length), and the output of the subtractor 208 (interblock distance) are registered in the block table 214. In this way, the entire amount taken in can be blocked.

The speech section determining section 23 judges the speech section from the block table 214 obtained by the blocking section 22 in the following manner. In other words, the block with the maximum block amount is detected, and if the distance between the blocks before and after the detected block is the center of the speech section, and the distance between the blocks is less than a certain value, the block is included in the speech section. was making a judgment.

The recognition unit 8 sends a capture start signal to the voice capture control unit 3, stores the analysis results from the voice capture control unit 3, and further receives start and end information 7 from the voice section detection unit 5. , and a standard pattern whose content is known and has been prepared in advance, and determines that a voice with the same content as the standard pattern with the highest degree of similarity has been input, and outputs the result.

However, in the speech interval detection using the above-mentioned conventional technology, (1) the spectral slope a _j changes depending on the strength of the input speech, so an unstable parameter, that is, P _j
is an unstable parameter.

(2) The spectral slope a _j varies depending on the phonology, speaker characteristics, microphone characteristics, etc.
The audio portion also takes a value close to 0, and as a result, P _j also takes a value close to 0, resulting in incorrect blocking.

(3) When the noise is loud, it is difficult to distinguish it from the noise (especially consonants).

There was a drawback.

(Problem to be solved by the invention) An object of the present invention is to provide a speech recognition device that can eliminate misrecognition and improve the recognition rate.The feature is that a noise pattern is subtracted from a speech pattern when detecting a speech section. This provides a means for detecting voice segments with higher accuracy and increasing the recognition rate, which will be described in detail below.

(Structure and operation of the invention) FIG. 6 is a block diagram of the invention.
200 is an input terminal, 200 is a frequency analysis section, 300 is a logarithmic conversion section, 400 is a spectrum conversion section, 500 is a speech interval determination section, and a logarithmically transformed data section 50
1. Noise pattern detection section 502, subtraction circuit 50
3. Multiplication circuit 504, addition circuit 505, division circuit 506, P parameter memory 507, comparator 1
508, FLAG509, smoothing 1 51
0, smoothing 2 511, blocking 51
2. Comparator 2 Consists of 513, block determination 514, speech interval determination 515, and MAXBLK table 516, 600 is a resampling section, 700 is a distance calculation section, 800 is a standard pattern memory, 900 is a judgment section, and 1000 is a recognition result output It is a terminal.

In such a configuration, the input audio signal input from the input terminal 100 is processed by the frequency analysis section 200.
is input into the quantized signal U _{(i,j) corresponding to multiple frequency bands, and is frequency-analyzed as a quantized signal U (i,j) corresponding to} a plurality of frequency bands.
Sent to 0.

The data sent to the logarithmic conversion section 300 becomes spectral information, power information, etc., the spectral information is sent to the spectral converter 400, and the spectral information and power information are sent to the speech interval determination section 500.

The logarithmic conversion unit 300 calculates equation (4). Let the frequency analysis data be U _(i,j) .

U _(i,j) i=1~19 j=1~∞ 0≦U _(i,j) ≦2047 Let the logarithmically transformed data be V _(i,j) .

V _(i,j) i=1 to 19 j=1 to ∞ Here, i indicates frequency (1ch to 19ch), and j indicates time (1 frame to ∞ frame). Also, input data from the preprocessing section is assumed to be U _(i,j) . U _(i,j) i=
1 to 19 j=1 to ∞ 0≦U _(i,j) ≦2047 Let NB be the number of logarithmic conversion bits. Here, NB=8.

Here, the calculation formula for the input pattern power POW (j) and the input pattern 10 frame power is expressed as (5)
It is defined by Equation (6).

POW(j)=1/19 ₁₉ 〓 ⁱ⁼¹ V(i,j) j=1~∞ (5) POW10(k)= ₁₀ 〓 ^l=1 POW _(j+l-1) (6) k= (j-1)/10+1 However, j=(k-1)*10+1.

The noise pattern is defined by equation (7).

When the noise pattern measurement interval is k=k ₁ to _{k 2} , NLEVEL=1/k ₂ −k ₁ +1 _k2 〓 ^k=k1 POW10(k) …(7) However, here, k ₂ = k ₁ + 2. When the slice level L ₁ is set to L ₁ = NLEVEL + LO, POW10 (k ₃ ) is greater than L1 and POW10 (k ₃ +1) is greater than L1 from the point k ₃ .
Let the frame j ₁ with 40 frames reversed be j ₁ = (k ₃ -1)*10+1-40, and the temporary voice start frame STFR1 be STFR1 = MAX (j, 1).

Termination detection is performed if k ₄ is greater than k ₂ +1, and
When POW10 (k ₄ ) becomes smaller than or equal to L1, a temporary audio end frame EDFR1 is set as EDFR1=(k ₄ -1)*10-1+9.

Now, the logarithmically transformed data V(i,j) calculated by the logarithmically transformed section 300 is transferred to the logarithmically transformed data section 50.
1, the noise pattern detection unit 502 calculates the noise pattern NPAT(i) in order to obtain the noise pattern NPAT(i). However, when the noise level measurement interval is set to k= _k1 to _k2 , the values of _j2 and _j3 are calculated using equation (8).

j ₂ = (k ₁ -1) *10 + 1 j ₃ = (k ₂ -1) *10 + 1 + 9 ... (8) The formula for determining the noise pattern NPAT (i) is expressed as (9)
As shown in the formula.

J=STFR1~EDFR1NPAT(i)=1/ _j3 - _j2 +1 ( _j3 〓 ^j=j2 (i,j)+ _j3 -j2 ₊₁ /2)...(9) Next, the subtraction circuit 503 and the multiplication circuit 504, addition circuit 505, and division circuit 506, V _{(i, j)} stored in logarithmically transformed data section 501 and noise pattern detection section 502, NPAT(i) found from equation (9). The power after subtracting the noise pattern is calculated using equation (10).

P(j)=1/19 ₁₉ 〓 ⁱ⁼¹ ((V(i,j)−NPAT(i)/4) ² +9…(10) P(j) found from equation (10) is P The signal is stored in the parameter memory 507, and the comparator 1508 performs a comparison according to the following equation (11).

FLAG(j)=0 Pp(j)<L2 1 P(j)≧L2 …(11) In equation (11), if slice level L2 is greater than P(j), FLAG(j)=0 shall be. Also
FLAG if L2 is less than or equal to P(j)
Let (j)=1. Determined in equation (11)
The value of FLAG(j) is stored in FLAG509,
Smoothing 1 5 depending on the value of FLAG(j)
10 or smoothing 2 511. Smoothing 1 For 510, FLAG(j) = 0
When FLAG (j-1) = 0 and FLAG (j + 1) = 0, FLAG (j)
=0. Also, in smoothing 2 511
Perform the operation when FLAG(j) = 1 and set FLAG(j
-1)=1 and FLAG(j+1)=1, then FLAG(j)=1.

Next, in blocking 512, FLAG(j)=1
is continuous for 4 or more frames, POW1 of that section
(l)=1/8〓P(j) is defined as a block if POW1(l)≧L3, that is, POW1(l) is greater than or equal to the slice level L3.

The number of blocks is BLKS, the first frame of block l is S(l), and the last frame of block l is E.
Let it be (l). The added value of the power P(j) after subtracting the noise pattern of block 1 is obtained by equation (12).

POW1(l)=1/8 _E(l) 〓 ^j=s(l) P(j)...(12) The number of frames of block 1 is determined by equation (13).

FR1(l)=E(l)-S(l)+1...(13) Also, the distance from the previous block (l-1) is the (14th)
It is determined by the formula.

FR2(l)=S(l)-E(l-1)...(14) Here, _l1 is the first audio block and _l2 is the final audio block.In the comparator 2 513, regarding the audio first block _l1 , As long as the condition of equation (15) is satisfied, l ₁ =l ₁ -1.

FR2(l ₁ )≦MIN(POW1(l ₁ −1)/SC1+SC2,
SC3) ...(15) Also, regarding the final audio block _l2 , as long as the condition of equation (16) is satisfied, _l2 = _l2 +1.

FR(l ₂ +1)≦MIN(POW1(l ₂ +1)/SC1+
SC2, SC3) …(16) Here, SC1 to SC3 are constants, SC1=16, SC2=
8, SC3=30.

Based on the above equation, it is determined whether or not the blocks before and after the largest block are to be taken in as speech section blocks, and are adopted as speech section block candidates.

In general, the sounds in the Japanese 50 sounds such as ka, ta, pa, ba, da, and ga are called plosives. Such plosive sounds are produced by once holding the breath and then opening the vocal cords all at once to vibrate them. In general, a word containing a plosive may have a silent period of about 10 to 30 ms during the utterance of the word, and the length varies from person to person.

Therefore, in this invention, as mentioned above, during the utterance period of a word, the power is equal to or greater than a predetermined value,
If the audio portion continues for a predetermined number of frames or more, the audio portion is defined as a block.

The number of blocks in one word varies depending on how many plosives it contains, but is usually 1 or more. However, when recognizing telephone voices, depending on the noise level superimposed on the received voice,
There is a possibility that the noise will be recognized as a block of speech. Therefore, since two or more consecutive blocks can include not only speech blocks but also noise blocks, such blocks are called speech section block candidates. In other words, a speech section block candidate may be a speech block as it is, or it may be a noise block. The values of the voice section block candidates _L1 and _L2 , which are voice section block candidates determined in this way, are sent to block determination 514.

Next, MAXBLK 516, which is the maximum number of blocks of recognition words used in speech section determination 515, will be explained.

Generally, one word can be divided into multiple blots if it contains a plosive. However, there are individual differences in vocalization and the timing of frame sampling (one frame is about 10ms).
Therefore, even if the word contains a plosive, it is not always divided into multiple blocks. However, even if you ignore the plosive at the beginning of a word, the number of plosives in the word will not exceed the number of plosives + 1.

The number of blocks included in one word is determined based on the number of plosives included in the word, and the maximum value is called the maximum number of blocks. Even if the actual number of blocks may be less than the maximum number of blocks,
It will never exceed this. For example, the maximum number of blocks for "1" is 2. However, in reality,
For the reasons mentioned above, the number of "first" blocks may be one.

A maximum block (MAXBLK) table is a table in which such block numbers are associated with each of a plurality of words. The MAXBLK table 516 is a maximum block (MAXBLK) table stored in memory.

An example of the maximum block number MAXBLK is shown in FIG. The left side shows the categories (16 words), and the right side shows the maximum number of blocks for each category calculated in advance from the utterance data. The largest of these recognition word sets is
Select MAXBLK. For example, if the recognized words include "Moichido", MAXBLK=3. Generalizing, the maximum number of blocks (MAXBLK) is, in principle, the number of plosives included in a word + 1.
However, the initial plosive sound shall not be included.
As an exception, for example, the ``ri'' in the word ``owari'' is not a plosive, but the number of blocks is 2, which is experienced by the utterer.
The number of blocks is 2.

In the voice section determining section 515, when BLKS≦MAXBLK, that is, when the number of blocks BLKS is smaller than or equal to the maximum number of blocks MAXBLK, all blocks are determined as voice sections. Conversely, when BLKS>MAXBLK, that is, when the number of blocks BLK is larger than the maximum number of blocks MAXBLK, for example, the seventh
In the figure, if the number of blocks BLKS = 3 and the maximum number of blocks MAXBLK = 2, a combination of or is considered, and after calculating the power PP (l) of each combination of blocks, the power PP of the block is calculated by comparing the PP. The combination of blocks for which (l) is the maximum is defined as a voice section. The power PP(l) of the block is obtained from equation (17).

PP(l)=1/E(l+MAXBLK-1)-S(l)+1 _{E(l+MAXBLK-4)} 〓 ^j=s(l) P(j)...(17) l=1~BLKS-MAXBLK+1 th S(l ₁ ) obtained from equation (17) is the audio start block, E(l ₂ ) is the audio final block, audio start frame STFR is STFR=S(l ₁ ), and audio end frame EDFR is EDFR =E(l ₂ ). Further, the number of input pattern frames IFR is expressed by the following equation (18).

IFR=EDFR-STFR+1 (18) The process is determined to be finished when the final audio block _l2 satisfies all of the conditions in equation (19) below.

POW10(K ₄ )≦L1 POW10(k ₄ +1)≦L1 POW10(K _{4 +2)≦L1 POW10(k 4 +3} )≦L1 POW10(k ₄ +4)≦L1 …(19) That is, L1 is k ₄ , k ₄ +1, k ₄ +2, k ₄ +3,
If it is greater than or equal to any of k ₄ +4, the process is terminated.

If the condition of equation (19) is not satisfied, recognition is stopped and the next value of k ₄ where POW10(k ₄ )≦L1, that is, L1 is greater than or equal to, is determined.

The voice interval STFR determined in this way and
EDFR is sent to the resampling unit 600 at the same time as W(i,j) sent from the spectrum conversion unit 400. The resampling unit 600 normalizes the time axis of the audio. The time axis normalization method is a conventionally known technique, and the linear matching method
This is a method in which a voice section is divided into a fixed number determined by the conditions of the recognition device at equal intervals in time and resampled. Then, in the distance calculation section 700,
The determination unit 900 calculates the distance from the output of the standard pattern memory 800 created in the same way and uses the result.
send to

The determination unit 900 compares the distance value with the total distance, and outputs the category name with the smallest total distance as the recognition result to the recognition result output terminal 100.
Output from 0.

As described above, in the present invention, by subtracting the noise pattern from the speech pattern when detecting the speech section, the speech section can be detected more accurately and the recognition rate can be increased.

(Effects of the Invention) The present invention makes it possible to detect speech segments more accurately by subtracting information about speech noise patterns from speech pattern information when detecting speech segments, thereby improving the recognition performance of the speech recognition device. It is effective in improving.

[Brief explanation of drawings]

Figure 1 is a block diagram of a conventional speech recognition device, Figure 2 is a detailed block diagram of the frequency analysis section in Figure 1,
FIG. 3 is a block diagram of the voice section detection section in FIG.
4 is a detailed block diagram of the parameter calculation section in FIG. 3, FIG. 5 is a detailed diagram of the blocking section in FIG. 3,
FIG. 6 is a block diagram of the speech recognition apparatus of the present invention, FIG. 7 is a diagram showing combinations of blocks in speech sections, and FIG. 8 is a diagram showing the maximum number of speech blocks. DESCRIPTION OF SYMBOLS 1... Input terminal, 2... Frequency analysis section, 3... Audio capture control section, 4... Capture start signal, 5... Audio section detection section, 6... Capture end signal, 7... Start/end information, 8
... recognition unit, 9 ... output terminal, 11 ... input audio signal,
12... Preamplifier, 13... Band pass filter group, 14... Full wave rectifier group, 15... Low pass filter group, 16... Multiplexer, 17... AD converter,
18... Analysis result, 21... Parameter calculation section, 22
...Blocking unit, 23...Voice section determining unit, 10
1,105,109...adder, 102,110...
Register, 103, 108, 111, 113... Multiplier, 104... Complementer, 106, 112... Divider, 107... Counter, 200... P parameter memory, 201... Absolute value circuit, 202, 205... Comparator, 203, 207, 209...Counter, 20
4,212,213...Register, 206...AND
Circuit, 208, 210... Subtractor, 211... Adder, 214... Block table, 100... Input terminal, 200... Frequency analysis section, 300... Logarithmic conversion section, 400... Spectrum conversion section, 500... Speech interval determination section, 501 ... Logarithmic conversion section, 502 ... Noise pattern detection section, 503 ... Subtraction circuit, 504 ... Multiplication circuit, 505 ... Addition circuit, 506 ... Division circuit,
507...P parameter memory, 508...Comparator 1, 509...FLAG, 510 Smoothing 1, 5
11...Smoothing 2, 512...Blocking, 5
13... Comparator 2, 512, 514... Block determination, 515... Voice section determination, 516...
MAXBLK, 600...Resample section, 700...Distance calculation section, 800...Standard pattern memory, 900...
Judgment unit, 1000... recognition result output terminal.

Claims (1)

[Claims] 1. In a speech recognition device that determines the speech interval of an input speech signal, noise contained in logarithmically transformed data obtained by frequency-analyzing the speech signal and logarithmically transforming the result. means for calculating a pattern; means for calculating power information included in the voice pattern by subtracting the noise pattern from the logarithmically transformed data; and comparing the power information with a predetermined first reference level to determine the magnitude between the two. means for determining a theoretical level speech interval flag indicating a relationship; and determining the logic level of the specific speech interval flag when the logic level of the specific speech interval flag is the same as the logic level of the speech flags before and after it; means for performing smoothing by performing smoothing; and when the voice section flags smoothed by the means for performing smoothing are continuous for a predetermined period and the power information exceeds a second reference level that is greater than the first reference level; , means for determining the power information as an audio block candidate; and means for determining the audio signal based on the result of comparing the maximum number of blocks of the audio block candidate with the maximum number of blocks in the maximum block table corresponding to the audio block candidate. and means for determining a voice interval, and the maximum block table can be continued by the means for determining a voice interval by associating a plurality of words with a maximum number of blocks determined in advance for the utterance data of those words. A voice recognition device characterized by being stored in a voice recognition device.