JPS61275799A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] The invention will be explained in the following order.

A. Field of industrial application B. Overview of the invention C. Prior art D. Problem to be solved by the invention E. Means for solving the problem F. Effect G. Example G. Description of acoustic analysis circuit (Fig. 1) 02 Time Explanation of normalization processing (Figures 1 and 2) G3 Explanation of pattern matching process (Figure 1) G4 Explanation of preliminary selection (Figures 3 to 5) G5 Explanation of other examples of preliminary selection ( Figures 6 to 8) Effects of the Invention A Industrial Field of Application This invention performs pattern matching between a standard pattern of words to be recognized, which has been created and stored in advance, and an input pattern of the word to be recognized. The present invention relates to a device for performing voice recognition.

B. Summary of the Invention The present invention is a pattern matching type speech recognition device that, at the time of registration, only stores the regular standard pattern of the recognition target word, and the binarization pattern of the acoustic parameter sequence of the recognition target word is stored. A standard pattern is also registered, and during recognition, the binarized pattern is obtained from the acoustic parameter series of the input word, the distance between this binarized pattern and the binarized standard pattern is calculated, and the pattern is compared to the input word pattern. This method performs preliminary selection to narrow down the standard patterns for distance calculation, which reduces the amount of calculation for distance calculation during pattern matching and shortens recognition response time.

C Conventional technologySpeech is a phenomenon that changes along the time axis, and is a phenomenon that changes along the time axis.
Unique words and words are created by uttering sounds with ever-changing patterns. Speech recognition is a technology that automatically recognizes words and phrases spoken by humans.
At present, it is extremely difficult to achieve speech recognition that is comparable to the human auditory function. For this reason, most speech recognition methods currently in practical use are performed by pattern matching a standard pattern of W recognition target words and an input pattern under certain conditions of use.

Figure 9 is a diagram for explaining the outline of this speech recognition device, in which voice input from the microphone (1) is input to the acoustic analysis circuit (
2). This acoustic analysis circuit (2) extracts acoustic parameters representing the characteristics of the input speech pattern. Various acoustic analysis methods can be considered to extract this acoustic parameter, but for example, a bandpass filter and a rectifier circuit are used as one channel, and a plurality of such channels are arranged with different passbands, and this bandpass filter A method is known in which a temporal change in a spectrum pattern is extracted as a group output. In this case, the acoustic parameters are their time series Pi(n) (i = 1. 2...
・III is, for example, the number of channels of the bandpass filter,
n=1.2...NUN can be expressed as the number of frames used for recognition in the section determined by voice section determination.

Acoustic parameter time series P from this acoustic analysis circuit (2)
i(n) is, for example, a mode switching circuit (
3). The switch of this circuit (3) is terminal A
When it is switched to the registration mode, the acoustic parameter time series Pi(n) is stored as a recognition parameter in the standard pattern memo January 4). That is, prior to speech recognition, the speaker's speech pattern is stored as a standard pattern in this memo. Note that during this registration, the number of frames for each registered standard pattern generally differs due to variations in speaking speed and differences in word length.

On the other hand, when this switch (3) is switched to the terminal B side, it is the recognition mode. In this recognition mode, the acoustic parameter time series of the input speech at that time from the acoustic analysis circuit (2) is supplied to the input speech pattern memory (5) and temporarily stored therein. Then, the distance calculation circuit (6) calculates the difference between this input pattern and each of the standard patterns of a plurality of recognition target words read out from the standard pattern memory (4). The recognition target word with the smallest difference from the minimum value determination circuit (7) is detected, and the input word is recognized by this.

In this way, the input speech is recognized by pattern matching processing between the registered standard pattern and the input pattern, but in this case, even if the same word is uttered in the same way, the spectrum pattern will be different from the time axis. It is necessary to take into consideration the possibility of deviation or expansion/contraction in the direction. In other words, for example, when recognizing the word "hai", if the standard pattern is registered as "hai", the input voice is "
If the word ``hi'' extends along the time axis, the distance will be very different and it will be treated as a completely different word, making it impossible to recognize it correctly. For this reason, in pattern matching for speech recognition, it is necessary to perform time normalization processing to correct this shift in the time axis direction, expansion and contraction, and this time normalization is an important process for improving recognition accuracy. It is.

One method of time normalization is D P (Dynami
cProgram+ning) There is a method called matching (see, for example, Japanese Patent Laid-Open No. 50-96104).

This DP matching does not prepare a large number of standard patterns that take time axis deviations into account, but instead generates a standard pattern that normalizes a large number of times using a distortion function, and calculates the distance between this and the input pattern. Speech recognition is performed by detecting the one with the minimum value.

By the way, when using this DP matching method, the number of frames of the standard pattern to be registered is undefined, and it is necessary to perform DP matching processing between all registered standard patterns and input patterns, and as the number of double voltages increases, the amount of calculation increases. The disadvantage is that the amount increases dramatically.

Furthermore, since DP matching is a matching method that emphasizes a stationary portion (a portion of a spectrum pattern that does not change over time), there is a possibility that erroneous recognition may occur between partially similar patterns.

The present applicant has previously proposed a time normalization method that does not cause such drawbacks (for example, Japanese Patent Application No. 106177/1982).
.

That is, the acoustic parameter time series Pi[nl draws a point sequence when considering its parameter space. For example vgl
! li When the target word is rHAIJ, the number of acoustic analysis band-pass filters is 2, and if Pi(nl= (Pz P2 ), then the acoustic parameter time series of the input speech is the 10th in the two-dimensional parameter space. Draw a sequence of points as shown in the figure.As is clear from this figure, the sequence of points in the non-stationary part of the voice is distributed coarsely, and the sequence of points in the quasi-stationary part is densely distributed.In this case, even if the voice is completely stationary, In this case, the parameters do not change, and in that case the point sequence will stay at one point in the parameter space.However, even if a human generates the same sound, it will not become completely stationary due to fluctuations in the sound, and the sequence of points will remain at one point in the parameter space. As a quasi-stationary part,
There is an effect of fluctuation.

From the above, it is considered that most of the deviation in the time axis direction due to variations in speech rate is caused by differences in point sequence density in the quasi-stationary part, and that the influence of the time length of the unsteady part is small. Therefore, if we estimate a trajectory drawn by a continuous curve that approximately passes through the entire point sequence as shown in Fig. 11 from the point sequence of this input parameter time series Pi(n), this trajectory will become the utterance of the voice. It can be seen that it remains almost unchanged with respect to speed fluctuations.

Based on this, the applicant proposed the following time axis normalization method. That is, first, the time series P of input parameters
A trajectory drawn by a continuous curve P i (8) from the start end Pi (1) to the end PiN of nl is estimated. In this case, the trajectory is estimated by, for example, linearly approximating the acoustic parameter time series as shown in FIG. The length S of the trajectory is determined from this estimated curve Pi(p). Then, resampling is performed at a predetermined length T along this trajectory as indicated by O in FIG.

For example, when resampling to M points, T-3/
(M-1)...(The trajectory is resampled using the length of 11 as a reference.

The parameter time series that depicts this resampled point sequence is Qih) (i=1.2...I, m=1.2...
・M), this parameter time series Qi (ml) has basic information on the trajectory, and is a parameter that is almost invariant to variations in the speech rate.

In other words, it is a recognition parameter time series whose time axis has been normalized.

Therefore, this parameter time series Qi ((b)) is registered as a standard pattern, and the input pattern is also obtained as this parameter time series QHIn), and the distance between both patterns is calculated using this parameter time series Ql (ffl). If the voice recognition is performed by detecting the one with the minimum distance, the voice recognition will always be performed in a state where the deviation in the time axis direction is normalized and removed.

According to this processing method, the number of frames in the recognition parameter time series old (E) is always M regardless of the speech rate fluctuation or word length difference at the time of registration, and in addition, the recognition parameter time series Qi (m ) has been time-normalized, so even the simplest calculation of the Chebyshev distance can be expected to yield good results when calculating the distance between the input pattern and the registered standard pattern.

Furthermore, the above method is a time normalization method that places more emphasis on non-stationary parts of speech, and reduces misrecognition between partially similar patterns such as in DP matching processing.

Furthermore, since the speech rate fluctuation information is not included in the normalized parameter time series old (E), it is easy to handle the global characteristics of the parameter transition structure arranged in the parameter space, and it is useful for speaker-independent recognition. It becomes possible to use various methods that are also effective against these conditions.

In addition, in the following, the above time normalization process is performed using N A
T (Nor+++alization＾long T
This process is called ``direction'' processing.

D. Problems to be Solved by the Invention Here, the difference in the amount of calculation between a speech recognition device that performs NAT processing and a speech recognition device that performs DP matching processing will be explained as follows.

Let α be the average amount of calculation in the DP matching distance calculation unit per standard pattern for the input pattern, and let γ be the average amount of calculation in the NAT processing unit. When the average amount of calculations in the DP matching process for the JIB standard pattern is set to β, the amount of calculations C1 in the DP matching process for the JIB standard pattern is C1-value α. Further, the amount of calculations C2 due to the NAT processing for the five standard patterns is C2-β·J+γ. Generally, the average amount of calculations α has the relationship α)β with respect to the average amount of calculations β. Therefore, as the number of words to be recognized increases, the amount of calculation C1 becomes larger than the amount of calculation C2 (, L>C2, and according to the speech recognition device that performs NAT processing, the amount of calculation can be significantly reduced.

In addition, since the recognition parameter time series 0f ((b) obtained from the NAT processing unit can be set to a constant number of parameters in the time series direction, the storage area of the standard pattern memory (4) can be used effectively, and its storage capacity can be reduced. It can be done relatively little.

In a speech recognition device that performs SAT processing in this way, compared to a speech recognition device that performs DP matching processing, the number of words to be recognized is lower due to the difference in the average amount of calculations per standard pattern for input patterns. The amount of calculation decreases as .

However, even in a speech recognition device that performs this NAT processing, it is necessary to calculate the distance between the input pattern and all standard patterns, as in the case of DP matching processing, and the absolute amount of calculation is still For this reason, the recognition response is relatively slow.

E. Means for Solving the Problems The present invention includes an acoustic analysis means (2) for obtaining an acoustic parameter sequence of an input audio signal;
) for estimating the trajectory drawn by the acoustic parameter series in the parameter space and calculating the trajectory length of this trajectory, and a standard pattern in which the recognition parameter series of the standard pattern of the recognition target word is stored. Memory (4)
and a first distance calculating means (6) for calculating the difference between the input pattern recognition parameter sequence formed based on the acoustic parameter sequence and the standard pattern recognition parameter sequence from the standard pattern memory; Means (6
); a minimum value determining means (7) for obtaining a recognition output by detecting the word of the minimum standard pattern of the values calculated in (7); and means (41) for forming a binarized pattern from the acoustic parameter series;
Distance calculation between the binarization standard pattern memory (43) in which the binarization standard pattern of the recognition target word is stored and the binarization standard pattern from the binarization standard pattern memory (43) A second distance calculation means (44
), and the standard pattern for calculating the distance to the input pattern in the first distance calculating means (6) based on the output of the second distance calculating means (44) is stored in the standard pattern memory (44).
).

F By calculating the distance between the binary pattern of the action input pattern and the binary standard pattern, among the regular standard patterns of all registered words, those that are close to the input pattern are preselected, and this preliminary selection The distance between only the standard pattern that has been created and the input pattern is calculated.

Therefore, at the stage of preliminary selection using this binarized pattern, the distance calculating means (6) narrows down the number of standard patterns to be calculated with the input pattern in advance, so the absolute amount of calculation is reduced accordingly. It is.

G. Embodiment FIG. 1 shows an embodiment of the speech recognition apparatus according to the present invention, in which a 16-channel band-pass filter group is used for acoustic analysis.

G1 Description of the acoustic analysis circuit (2) In other words, in the acoustic analysis circuit (2), the audio signal from the microphone (11) is passed through the amplifier (211) and the low-pass filter (212) for band limitation to the A/D converter ( 213) and is converted into a 12-bit digital audio signal at a sampling frequency of, for example, 12.5kHz.This digital audio signal is supplied to the digital bandpass filter (213) of each channel of the 16-channel bandpass filter bank (22). 221z),
(2212), ..., (221xs). This digital band pass filter (22h)
, (2212) ,..., (22ha)
is composed of, for example, a Butterworth fourth-order digital filter, and each band obtained by dividing the band from 250 Hz to 5.5 KHz at equal intervals on the logarithmic axis becomes the pass band of each filter. And each digital band pass filter (2211), (2212).

..., (22118) output signals are respectively rectified circuits (2221), (2222), ...
, (222ts) and these rectifier circuits (222ts).
221), (2222).

The outputs of (2221g) are respectively digital low-pass filters (2231), (2232),
..., (223ts) is supplied. These digital low-pass filters (2231), (2232)
, . . . (223ts) is composed of, for example, an FIR low-pass filter with a cutoff frequency of 52.8 Hz.

Each digital low-pass filter (2231), (2232), etc. which is the output of the acoustic analysis circuit (2)
.

The output signal of (22318) is supplied to a sampler (231') constituting the feature extraction circuit (23). This sampler (231) uses a digital low-pass filter (22
31) , (2232) ,..., (22
3ss) is sampled every frame period of 5.12m5ec. Therefore, from this, the sample time series Ai(n) (i = 1. 2. . . . 4
6; n is the frame number n-1+2+ ..., N
) is obtained.

The output from this sampler (231), that is, the sample time series At(n), is supplied to a sound source information normalization circuit (232), which removes differences in vocal cord sound source characteristics depending on the speaker of the speech to be recognized. Ru.

That is, for the sample time series Ai(n) supplied from the sampler (231) every frame period, At(n)=
log (Ai(n)+B)...
(2) A logarithmic transformation is performed. In this equation (1), B is set to a value such that the noise level is hidden by the bias.

Then, when the vocal cord sound source characteristics are approximated by the formula yi=a−i+b, the coefficients of a and b are determined by the following formula.

(1=16) ・ ・ ・(3) I (
1-1) (I-16) ・ ・ ・(4) And,
Letting the normalized parameters of the sound source be Pi (b),
a(mouth1<0, parameter Pi(nl is Pi(nl
=Al(nl-(a(n) ・i+ b(n))
...It is expressed as 151.

Also, when a (n)≧0, only level normalization is performed,
The parameter Pi(n) is expressed as...(6).

In this way, the sound source information normalization circuit (232) obtains an acoustic parameter time series P1(n) in which differences in vocal cord sound source characteristics are normalized and removed.

The acoustic parameter time series Pi (nl) from this sound source information normalization circuit (232) takes both positive and negative values.

The acoustic parameters Pi(n) from the sound source information normalization circuit (232) are supplied to the intra-speech interval parameter memory (8). The intra-speech-segment parameter memory (8) receives a speech-segment determination signal from the speech-segment determination circuit (24), and stores parameters Pi(n) for each determined voice period.

The voice section determination circuit (24) includes a zero cross counter (24).
1), a power calculation circuit (242), and a voice section determination circuit (
243), and the digital audio signal from the A/D converter (213) is supplied to the zero cross counter (241) and the power calculation circuit (242). For zero cross counter (24'l), one frame period is 5.12m5
For each ec, the number of zero crosses of the 64 samples of the digital audio signal within one frame period is counted, and the count value is supplied to the first input terminal of the audio section determining circuit (243). The power calculation circuit (242) calculates the power of the digital audio signal within this l frame period for each frame period, that is, the sum of squares, and the output power signal is sent to the second voice section determination circuit (243). Supplied to the input end. The voice section determination circuit (243) is further supplied with sound source normalization information from the sound source information normalization circuit (232) to its third input terminal. In this speech section determining circuit (243), the number of zero crossings, the power within the section, and the sound source normalization information are processed in a composite manner, and a process for determining silence, unvoiced sound, and voiced sound is performed, and a speech section is determined.

The voice interval determination signal indicating the determined voice interval from the voice interval determination circuit (243) is transmitted to the voice interval determination circuit (243).
) as the output of the voice interval parameter memory (200
).

In this way, within the judgment voice section, the memory (200)
The acoustic parameter time series P 1 (11 is N
The signal is supplied to the AT processing circuit (9).

02 Description of Time Normalization Process The NAT processing circuit (9) consists of a trajectory length calculation circuit (91), an interpolation interval calculation circuit (92), and an interpolation point extraction circuit (93).

Parameter time series Pi(nl (i = 1. 2. −, 16;
n-1,2.

..., N) are supplied to the trajectory length calculation circuit (91).

This trajectory length calculation circuit (91) calculates the length of the trajectory of the acoustic parameter time series Pi(n) by linear approximation drawn in the parameter space as shown in FIG. 11 described above.

In this case, the Euclidean distance D (a+, bt) between the ■-dimensional vectors al and b[ is (7). Therefore, ■-dimensional acoustic parameter time series Pi (
From nl, the distance S (n) between adjacent parameters in the time series direction when the trajectory is estimated by linear approximation is 5(n)
= D (Pi (n + 1), Pi (n))
(n=1...,N)...(8). Then, the first parameter Pi(1) to the nth parameter Pitn in the time series direction
The distance 5L (nl) to l is expressed as 5L(1)-〇.

Then, the total trajectory length SL is expressed as. The trajectory length calculation circuit (91) is this (11)
The signal processing shown in equations (12) and (13) is performed.

The trajectory length SL obtained by this trajectory length calculation circuit (91)
A signal indicating the interpolation interval calculation circuit (92) is supplied to the interpolation interval calculation circuit (92).

This interpolation interval calculation circuit (92) calculates the resampling interval T when resampling is performed along the locus.

In this case, if resampling is performed at M points, the resampling interval T is T = SL/ (M-1) ... (
11.

A signal indicating the resampling interval T from this interpolation interval calculation circuit (92) is supplied to an interpolation point extraction circuit (93). In addition, the acoustic parameter time series Pi(n) from the parameter memory (8) is also stored in this interpolation point extraction circuit (93
). This interpolation point extraction circuit (93) follows the locus of the acoustic parameter time series Pi(n) in its parameter space, for example, the locus of linear approximation between the parameters, at resampling intervals T as shown by O in FIG. The recognition parameter time series Qt(e) is formed from the new point sequence obtained by this sampling.

Here, the interpolation point extraction circuit (93) performs processing according to the flowchart shown in FIG. 2 to form a recognition parameter time series Qi.

First, in step (1013), the value 1 is set to the variable J indicating the number of the resampling point in the time series direction, and the value 1 is set to the variable IC indicating the frame number of the acoustic parameter time series Pi[n), Ini) Charized. Next, in step (102), the variable J is incremented, and in step (103), the variable J at that time is incremented.
By determining whether or not is less than or equal to (M-1), it is determined whether the number of the resampling point at that time in the time series direction is the last number that needs to be sampled. If it is the last number, the process advances to step (104) and resampling ends.

If it is not the last number, go to step (105) and select the first resampling point (this is always a silent part).
A resampling distance DL from to the third resampling point is calculated. Next, the process proceeds to step (106), where the variable IC is incremented. Next step (10
7), the resampling distance DL is the acoustic parameter time series P
L(nl's first parameter Pi(1) to first C
Depending on whether the distance to the th parameter Pi ((6) is smaller than the distance SL (Y), it is determined whether the resampling point at that time is located closer to the starting point of the trajectory than the parameter Pi + Yu at that time on the trajectory. and if it is not located on the starting point side, return to step (106) and change the variable IC
After incrementing , the positions of the resampling point and the parameter piaa on the trajectory are compared again in step (107), and when it is determined that the resampling point is located closer to the starting point than the parameter memory o on the trajectory. , proceed to step (10B) and recognize the recognition parameter Q
ig) is formed.

That is, from the resampling distance MDL by the third resampling point, the (IC-1)th parameter P located closer to the starting point than this third resampling point
i ac-1) by subtracting the distance S L (IC-11) to obtain the (IC-1)th parameter P 1 (Ha-1>
Find the distance SS from to the third resampling point. Next, the distance 5(n) between the parameter P i [IC-13 and the parameter Piaa located on both sides of this third resampling point on the trajectory (this distance %ll
Divide this distance SS by S (nl is obtained by signal processing shown in equation (11)), and the division result S
S/S (Ic-11 has parameters Pi(10
and P foc-1+ (P too -P i++
c −1+ ) to calculate the (IC−1)th parameter P located adjacent to the starting point side of the third resampling point on the trajectory.
The interpolation amount from ir+c-t+ is calculated, and this interpolation amount and the (IC-1)th parameter P t (IG
-1+ is added to form a new recognition parameter Qiσ) along the trajectory.

In this way, the starting point and the ending point (when these are silent, respectively, Qi (1) = Pi (o) = 0,
Qicx=Pi+g+=0. ) except < (
By resampling M-2) points, an m-identified parameter time series Qi (1111) is formed.

G3 Description of Pattern Matching Process The tIm parameter time series from the NAT processing circuit (9) is supplied to the mode switching circuit (3) and also to the preliminary selection circuit (40). Further, a signal indicating the calculated trajectory length from the trajectory length calculation circuit (91) is supplied to the preliminary selection circuit (40).

At the time of registration, the recognition parameter time series is stored in the standard pattern memory (4).

The recognition parameter time series Q1 (b) supplied to the preliminary selection circuit (40) is converted into a binarized pattern as described later,
This is registered in memory. In addition, a preliminary selection circuit (40
) is used to create a trajectory length dictionary for preliminary selection, as will be described later.

Next, at the time of recognition, M from the NAT processing circuit (9)
The 11ia pattern time series Qi (E) is supplied to the distance calculation circuit (6) via the mode changeover switch (3), and the distance between it and the parameter time series of the standard pattern from the regular standard pattern memory (4) is calculated. It will be done. In this case, the distance calculation circuit (6) performs distance calculations on the input parameters based on the output of the preliminary selection means (40), as will be described later. It has been selected and narrowed down.

The distance calculated by the distance calculation circuit (6) is calculated as, for example, a simple Chebyshev distance. The distance calculation circuit (6) calculates the distance between each standard pattern and the input pattern to determine the standard pattern that has the minimum value, and based on this determination result, the recognition result of the input audio signal is
) can be obtained.

G4 Preliminary Selection Description FIG. 3 is an example block diagram of the preliminary selection circuit (40).
All of these can also be achieved through computer processing.

First, the interpolation point extraction circuit (93) of the SAT processing circuit (9)
The recognition parameter time series Qi ((b)) is supplied to the binarization pattern generation circuit (41). As mentioned above, the recognition parameter time series Qi (e) is supplied to the sound source information normalization circuit (41).
232), it has positive and negative values, so the binarization pattern generation circuit (41) uses, for example, "l" for positive values and "l" for negative values. 0'' are associated with each other to generate a binarized pattern.

In the registration mode, this binarized pattern is transferred to the binarized standard pattern memory via the mode switching circuit (42).
(43) is written.

Further, in this example, in this registration mode, the trajectory length output SL from the trajectory length calculation circuit (91) is stored in the trajectory length memory (47) via the mode switching circuit (46).

In this example, when the registration of all regular standard patterns, binarized standard patterns, and trajectory lengths of the recognition target word is completed, the trajectory length dictionary creation means (48) stores the trajectory length in the memory (47). Based on this, a trajectory length dictionary is created that indicates which word standard pattern is to be read out from the binarized standard pattern memory (43) for trajectory lengths in a predetermined length range.

In this example, the dictionary is created as follows.

That is, the maximum and minimum values are determined from all the trajectory lengths stored in the memory (47), and the distance between the maximum and minimum values is divided into n equal parts to form n trajectory length ranges. , Create a dictionary by registering the words of trajectory length belonging to each range.0 For example, if the maximum value is 600, the minimum value is 200, and the range length is 50, as shown in Figure 4, ■ ■
...The trajectory length range of @ is determined, and the data of each range and the data of words A, B, C, D, etc. belonging to each range (for example, each word A.

The address data of the binary standard pattern memory (43) of B, C, D, ...) is the trajectory length memory (47)
Stored in

Next, at the time of voice tlk, preliminary selection using a trajectory length dictionary is performed as described below, further preliminary selection is performed using a binarized pattern, and then pattern matching processing is performed.

That is, FIG. 5 is a flowchart of preliminary selection of a binarization pattern based on the trajectory length, in which the trajectory length SL of the acoustic parameter of the input word from the trajectory length calculation circuit (91) is preselected via the mode switching circuit (46). It is supplied to means (49) (step (201)). Next, 0 from memory (47)
The trajectory length range of the number is read out (step (202)
), that is, at the beginning, the trajectory length range of ■ is read out. It is determined whether the input trajectory length SL is within the range of ■ (step (203)), and if it is not within the range, the trajectory length range is read out. It is determined whether or not it is the number 0, that is, the last one (step (204)), and if it is not within the range of number 0, the process returns to step (202) and the next number 0, for example, within the range of number 0 is determined. Then, it is determined whether the trajectory length SL is within the range.

Then, when the trajectory length SL is determined to be within the trajectory length range of number 0 in step (203), the address data of the registered word in the trajectory length range of number 0 is read out from the memory (47), and this is used as the preliminary selection. It is supplied to the binarized standard pattern memory (43) via the circuit (49), and from this
Only the binarized pattern of the registered word in the trajectory length range of number is read out (step (205)), and the distance calculation circuit (44)
supplied to On the other hand, the binarization pattern generation circuit (41)
The binarized pattern of the input pattern is supplied to the distance calculation circuit (44) via the mode switching circuit (42),
The distance to the read binarized standard pattern is calculated.

In this case, the distance can be calculated by exclusive ORing the binarized Tievishiev distance. For example, the normalization pattern, that is, the recognition parameter Qi (E) has a precision of 16 bits, and is 16×1.
An 8-word normalized pattern can be compressed into an 18-word binarized pattern, and when the exclusive OR is performed, distance calculation is possible with the sum of the count values of the "0" outputs. The amount of calculation is approximately 1/1 of that when using a 16-bit normalized pattern.
A calculation amount of 0 to 1/16 is sufficient.

That is, the distance T is determined by the following formula.

T-Σ bit count (IIj■Sj) step (2
03), if the input trajectory length SL does not fall within all trajectory length ranges from 0 to 0, step (2)
04), and the input word is not among the registered words.
Alternatively, it is determined that they are far apart, and a reject signal is obtained at the terminal (50), and the difference in distance between the binarized standard pattern and the input pattern is not calculated, and, for example, an unrecognizable display is displayed. Of course, at this time, the distance calculation circuit (6) does not calculate the distance due to the output of the preliminary selection means (45), which will be described later.

As described above, the binarized standard pattern that has been preselected based on the trajectory length, narrowed down and read out from all the binarized standard patterns can be used to calculate the distance between the input pattern and the binarized pattern as described above. This is performed in a distance calculation circuit (44), and the distance calculation output is supplied to a binarization pattern preliminary selection circuit (45). This preliminary selection circuit (45) selects only the top regular standard patterns (k is a number smaller than all the registered patterns) from those with the smallest distance calculation output value, and inputs them to the distance calculation circuit (6). A signal SR for calculating the distance to the pattern is obtained. In this example, this signal SE is supplied to the distance calculation circuit (6), and a standard pattern to be used for distance calculation with the input pattern from the mode switching circuit (3) is selected.

In addition, the memory address of the standard pattern whose distance from the input pattern is to be calculated is obtained as an output from the preliminary selection circuit (45) by preliminary selection using the binarized pattern, and this output controls the reading from the standard pattern memo (January 4). Preliminary selection may also be performed.

Further, when obtaining a signal for selecting standard patterns from the shortest distance to the highest in the preliminary selection circuit (45), only standard patterns whose output values from the circuit (44) are within a certain distance are selected. It's okay.

In this case, if the distance differences of the binarized patterns do not fall within a certain distance range and all of them are large, the terminal (51
) can get a reject signal.

In the above example, the preliminary selection based on the trajectory length was performed when calculating the distance using the binarized pattern, but when the distance calculation circuit (6) calculates the distance using the regular standard pattern, the preliminary selection based on the binary pattern is performed. They may be performed in parallel. In this case, the standard pattern that is calculated with the input pattern in the distance calculation circuit (6) is a logical difference between the one selected by the preliminary selection based on the trajectory length and the one selected by the preliminary selection based on the binary pattern. It may be Japanese.

In this example, preliminary selection based on the trajectory length is performed in addition to preliminary selection using the binarization pattern, so the number of standard patterns that are reduced by the preliminary selection using the binarization pattern and whose distance must be calculated is further reduced. A significant improvement in response speed can be expected.

In addition, at the time of registration, if the trajectory length range is determined, the trajectory length range of the standard pattern to be registered is determined, and the writing address of the registered standard pattern is determined according to the determined range. A trajectory length dictionary can be created along with the registration of standard patterns.

G5 Explanation of Other Examples of Preliminary Selection FIGS. 6 to 8 are diagrams for explaining other examples of preliminary selection of the apparatus of the present invention, in which the method of preliminary selection based on trajectory length is different from the previous example.

FIG. 6 is a block diagram of the main part. During recognition, the distance calculation circuit calculates the distance between the recognition parameter time series from the SAT processing circuit (9) and the parameter time series of the standard pattern preselected as described below. The distance is calculated in step (6), the minimum value of the distance is determined in the minimum value determining circuit (7), and the recognition output is obtained at the output terminal (70), as in the previous example.

In this example, when registering, one word is entered multiple times,
As a standard pattern, an integrated pattern obtained by taking the OR (logical sum) is registered. Of course, you can register everything.

Then, the calculated trajectory length SL from the trajectory length calculation circuit (91)
is supplied to the maximum value/minimum value detection circuit (52) via the mode switching circuit (46), and the maximum value Max and minimum value Min of the trajectory length when one word is input multiple times is detected. The maximum value Max and minimum value Min are written in the memory (53) in association with the address of each word registered in the standard pattern memo (January 4) as shown in FIG.

Lol! , a preliminary selection is made according to the flowchart of FIG. 8.

That is, first, the address of the memory (53) is initialized (step (301)). Then, the calculated trajectory length SL of the input word from the trajectory length calculation circuit (91) is sent to the preliminary selection circuit (49) via the mode switching circuit (46).
(step (302)).

On the other hand, the address of the memory (53) is incremented (step (303)), and first, the first address is specified, and the stored maximum value Max and minimum value Min of the trajectory length are read out, and these are used as the preliminary selection. Circuit (49
). Then, the calculated trajectory length SL is the minimum value Mi
It is determined whether it is larger than n (step (305)
), if it is larger, the process proceeds to step (306), where it is determined whether or not the calculated trajectory length SL is smaller than the maximum value Max. The binarized standard pattern of registered words having values is read out from the memory (43) and supplied to the distance calculation circuit (44).

The address is then incremented (step (308)
) ) , a standard pattern with maximum and minimum values such that Min≦SLY Max is detected and stored in memory (4) until the address is the last one in memory (53).
3) is read out (step (3053 to step [307)).

Then, when it is determined that the address is at the end of the memory (53) (step (309)), the preliminary selection ends.

On the other hand, in step (305), the calculated trajectory length SL is the minimum value M
Even if it is determined that the trajectory length SL is smaller than in, or even if SL≧Min, the trajectory length SL is set to the maximum value Ma in step (30B).
When it is determined that it is larger than x, step (303
), the maximum value Max and minimum value side n of the trajectory length of the registered standard pattern at the next address are read out, and it is determined whether Max≧SL≧Min as described above, and if so, the standard The pattern is read out from the memory (43) (steps (305) to (309)), otherwise the process returns to step (303) and is repeated.

If the maximum value Max and the minimum value Min are read from all addresses but Min≦SL5 Max is not satisfied, this is determined in step (304) and pattern matching is not performed. , a reject signal indicating this is output from the preliminary selection circuit (49) to the terminal (50).

In this way, by preliminary selection, only the binarized standard patterns of words that can be taken as trajectory lengths are stored in memory (43
), and the amount of calculation is smaller than when calculating the distance with all the binarized standard patterns.

The distance calculation circuit (6
), the input pattern and the standard pattern whose distance is to be calculated are narrowed down, as in the previous example.

Further, in this example as well, the preliminary selection based on the trajectory length may be the preliminary selection of a regular standard pattern from the memory (4) instead of the preliminary selection of the standard pattern of the binarization pattern, as in the previous example.

It goes without saying that the same effects as in the previous example can be obtained in this example as well.

In addition, in the above embodiment, the acoustic parameter time series P
Although the case has been described in which the trajectory length of the trajectory in the parameter space is calculated from i(n), the trajectory length of the trajectory in the parameter space may be calculated from the acoustic parameter frequency series.

Further, in the above embodiment, the length of the trajectory is calculated by linear approximation, but the length of the trajectory may be calculated by arc approximation, spline approximation, or the like.

Furthermore, in the above embodiment, the acoustic parameter time series Pi(n) of the acoustic analysis section (2) is transmitted to the NAT processing circuit (9).
As described above, preliminary selection is made using the trajectory length calculated by the trajectory length calculation circuit (91) of the NAT processing circuit (9). A trajectory length calculation circuit is separately provided, and the new recognition parameter time series Qi (E) from the NAT processing circuit (9) is supplied to the other trajectory length calculation circuit to calculate the trajectory length of the trajectory in the parameter space. The trajectory length may be calculated and preliminary selection may be made based on this trajectory length.

Of course, it is not necessary to perform preliminary selection based on trajectory length.

It goes without saying that the present invention can also be applied to a speech recognition device that does not include a NAT processing circuit. For example, in a speech recognition device that performs DP matching processing, the acoustic parameter series of the acoustic analysis section (2) is supplied to a binarization pattern generation circuit, and the binarization pattern generated by the binarization pattern generation circuit is The amount of calculation for the DP matching process can also be reduced by selecting the standard pattern based on the distance calculation output between the standard pattern and the binarized standard pattern.

In the above example, the readout of the binarized standard pattern from the binarized standard pattern memory is controlled by the preliminary selection based on the trajectory length, and the two are supplied to the distance calculation means (44).
Although the value conversion standard pattern is narrowed down, the distance calculation means (4) is similar to the process in the distance calculation circuit (6).
In the input stage of 4), the standard pattern may be gated by the preliminary selection output, or the standard pattern that requires distance calculation may be excluded from the distance calculation by preliminary selection when calculating the distance from the input pattern. good.

Conversely, reading of the memory (4) may be controlled to preliminarily select a standard pattern to be supplied to the distance calculation circuit (6).

In the above example, in binarizing the acoustic parameter time series, we used the fact that the output of the sound source information normalization circuit takes positive and negative values. Alternatively, a slice level may be determined, and when the slice level is greater than the slice level, "1" is assigned, and when it is smaller than the slice level, "0" is assigned, and the binarization is performed.

H Effects of the Invention As described above, according to this invention, a binarized pattern of the acoustic parameter sequence of an input word is generated and registered, and during recognition, the binarized pattern of the input word and the registered binarized standard pattern are The distance between the input pattern and the regular standard pattern to be calculated is preliminarily selected based on the calculation output, and the standard pattern narrowed down from all the registered standard patterns by this preliminary selection is supplied to the distance calculation circuit and input. Since the distance to the pattern is calculated, the amount of calculation when calculating the distance can be reduced. In this case, the distance calculation for the binarized pattern is added, but the amount of calculation is very small as described above, so the overall amount of calculation can be significantly reduced.

Therefore, the response during recognition becomes shorter and faster by the reduction in the amount of calculation.

In addition, when the input word is not a recognition target word, the preliminary selection stage can detect it and obtain a reject signal,
It also has the advantage of being able to respond quickly.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the device of the present invention, FIG. 2 is a flowchart for explaining the operation of a part of the device, FIG. 3 is a block diagram of an example of the main part, and FIG. 4 5 is a diagram for explaining the example shown in FIG. 3, FIG. 6 is a block diagram of another example of the main part of this invention device, FIGS. 7 and 8 are diagrams for explaining the same, and FIG. 9 is a diagram for explaining the example shown in FIG. The figure is audio m! FIGS. 10 to 12, which are block diagrams showing the basic configuration of the II device, are diagrams for explaining NAT processing. (2) is an acoustic analysis circuit, (4) is a standard pattern memory,
(6) is a distance calculation circuit between a standard pattern and an input pattern, (7) is a minimum value judgment circuit, (9) is a NAT processing circuit,
(41) is a binarization pattern generation circuit, (43) is a binarization standard pattern memory, and (45) is a binarization preliminary selection circuit. Precise review I! 11 J^ Naka Jf
t ho 1 ■ B 10 Figure 1 and the future of the future

Claims (1)

[Scope of Claims] (a) Acoustic analysis means for obtaining an acoustic parameter sequence of an input speech signal; (b) Standard pattern memory means in which an acoustic parameter sequence of a standard pattern of a recognition target word is stored; (c) a first distance calculation means for calculating the difference between the acoustic parameter series of the input pattern and the acoustic parameter series of the standard pattern read from the standard pattern memory means; (d) the value calculated by the first distance calculation means; (e) means for forming a binarization pattern from the acoustic parameter series; (f) a binarization standard for the recognition target word; a binarized standard pattern memory means in which the pattern is stored; (g) a second distance calculation for calculating the distance between the binarized pattern and the binarized standard pattern from the binarized standard pattern memory means; (h) selecting a standard pattern from all of the standard pattern memory means for calculating the distance to the input pattern in the first distance calculating means based on the output of the second distance calculating means; A speech recognition device comprising preliminary selection means.