JPS6273298A - Voice recognition system - 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] (Industrial application field) The present invention relates to a speech recognition method with high recognition accuracy.

(Prior Art) Research and development regarding speech recognition has been progressing for the purpose of increasing the efficiency of inputting information and communication equipment and improving system functions. A common method for performing this speech recognition is the pattern matching method.

First, prior to explaining the present invention, a conventional pattern matching method will be explained with reference to FIG.

In FIG. 7, lO is a voice input terminal, 11 is a voice analysis section, 12 is a section detection section, 13 is an input memory section, 14 is a comparison pattern memory section, 15 is a similarity calculation section, 16 is a judgment section, and 17 is a It is an output terminal.

In this conventional recognition method, the input voice input to the voice input terminal IO is input to the voice analysis unit! Time-series pattern of vectors representing characteristics in ■ (hereinafter referred to as audio pattern)
Convert to This audio pattern is generally created by sampling (hereinafter referred to as sampling) in-band frequency components extracted by two band-pass filter groups with different center frequencies at time intervals To (for example, 8 milliseconds). obtained by. On the other hand, in this Ondo analysis section 11.

Calculate the audio power at a time point corresponding to the audio pattern. The voice patterns calculated in the voice analysis section 11 are sequentially stored in the input memory section 13, and the voice power is outputted to the section detection section 12.

The section detecting section 12 determines a voice section, that is, the start and end of the voice, based on the voice power from the voice analysis section 11. The algorithm for determining the start and end of the voice based on the voice power is a complex algorithm as disclosed in Japanese Patent Application No. 108668/1982, and the time when the voice power exceeds the threshold is determined as the start and end of the voice. There are simple algorithms and other algorithms that consider the point at which the sound ends as the end of the voice, and any suitable algorithm is used to detect the section. The audio pattern between the start and end points determined by the section detection section 12 is read out from the input memory section and sent to the similarity calculation section 15. On the other hand, a comparison pattern is separately inputted to the similarity calculation unit 15 from the comparison pattern memory 14. This comparison pattern is a time-series pattern of vectors obtained by subjecting words to be recognized (hereinafter referred to as categories) to the same speech analysis process as speech patterns, and is stored in the comparison pattern memory section 14 in advance.

For this storage, a comparison pattern is created, but the creation method differs depending on the purpose of recognition. For example, in the case of a recognition method that limits speakers, the comparison pattern is a voice pattern obtained by using the frequency analysis unit 11 or an equivalent voice analysis process on the voice uttered by the limited speakers. 4 in the comparison pattern memory section]4.

The similarity calculation unit 15 calculates the similarity between the speech pattern and the comparison pattern. This similarity calculation may be carried out using, for example, a method called DP matching disclosed in Japanese Patent Publication No. 50-23941, a method that optimizes distortion in the time axis direction due to fluctuations in speech rate, or other suitable methods. A method is used.

Using the similarity for each category outputted from the similarity calculating section 15, the determining section 16 outputs the category name given to the comparison pattern that can achieve the maximum similarity as a recognition result.

The above is an outline of the conventional speech recognition method using the pattern matching method.

(Problems to be Solved by the Invention) The conventional recognition method described above evaluates the difference between a speech pattern that gives the shape of the speech spectrum and a comparison pattern calculated in advance by the same analysis process using a measure of similarity. , the recognition result was the category name of the comparison pattern that gave the greatest degree of similarity. Therefore, when the voice pattern category and the comparison pattern category are the same, the degree of similarity is large, and when they are different, the degree of similarity is small.

However, if the shape of the speech spectrum is distorted by factors other than speech, such as external noise, the degree of similarity between the two cannot be said to be large, even if they are in the same category.

Further, in the conventional recognition method, the calculation process takes time and requires a large storage capacity, so there is a problem that the structure of the device implementing this method becomes large.

SUMMARY OF THE INVENTION In view of these conventional problems, it is an object of the present invention to provide a speech recognition method with good recognition accuracy even in a noisy environment.

Another object of the present invention is to provide a speech recognition system which, when configured as a device, has a simple and compact structure, has a high calculation processing speed, and requires a small storage capacity.

(Means for Solving the Problems) In order to achieve the above-mentioned object, the speech recognition system of the present invention takes the following measures.

(a) First, the frequency components of the input audio are extracted by multiple bandpass filters, and the output is
(referred to as audio frames) and calculates a feature vector.

(b) Also, calculate a noise pattern obtained by time-averaging feature vectors in a predetermined noise section that is known in advance to be only noise.

(C) After this noise pattern extraction, a speech feature vector is calculated by subtracting the noise pattern from the feature vector.

(d) Calculate the least squares approximation straight line from the above audio feature vector for each audio frame, and convert this least squares approximation straight line (
A local peak vector is calculated by setting the component corresponding to the maximum channel in the frequency axis direction to 1.

(e) Calculate the frame power in the audio frame from this audio feature vector, and calculate the start and end of this frame power.

(f) On the other hand, for each voice of the recognition target word, a comparison pattern is created by performing processes corresponding to each of the above-mentioned processes (a) to (e) performed on the input voice (registration process and ).

(g) The above (a) for the voice uttered after recognition processing.
A non-linear matching process is performed between the input pattern obtained through the processes up to (e) and the comparison pattern to calculate the degree of pattern similarity between the comparison pattern and the input pattern.

(h) A process is performed in which the category name added to the comparison pattern that provides the maximum similarity among the pattern similarities calculated for each comparison pattern is used as a recognition result.

In the manner described above, the result of recognizing input speech is obtained.

The above-mentioned processes (a), (b), and (C) are processes for extracting only speech from an input with high noise, and are considered difficult at high noise f (e) This is a process that facilitates the process of detecting the voice section of a term.

Furthermore, recognition performance in a high noise environment is improved by using the local peak vector calculated in section (d) to calculate the similarity between sections (g) and (h).

This is because a local peak vector calculated based on a position giving a peak of the audio spectrum is used for similarity calculation, instead of using a vector giving the shape of the spectrum as in the conventional method. Therefore, it is based on the fact that when noise is mixed, the shape of the spectrum changes significantly, but the position of the peak of the spectrum does not change.

(effect) Next, the operation of this invention will be explained.

The functions for achieving the speech recognition method of the present invention are constituted by each processing section shown in FIG.

The detailed processing will be explained below.

The sound is converted into an electrical signal through a microphone, passed through an amplifier (not shown), a low-pass filter (not shown), and sent to an A/D converter (not shown), where it is sampled every 83 microseconds. ), then input to the input terminal 21.

Each of the above-mentioned items will be explained below.

[Feature vector calculation process in section (a)] Frequency analysis of audio data input to the input terminal 21 is performed by the feature vector calculation unit 22, and the data is converted into audio frame time series feature vectors.

The feature vector calculation unit 22 includes a plurality of band-pass filters each having characteristics with different center frequencies as shown in FIG. 2 for frequency analysis, a low-pass filter, and sampling for each audio frame. sampling means (each not shown in the figure).

Each band filter extracts only the center frequency component from the voice. The data series separated by each band filter in this way is called a channel. After performing an absolute value calculation on the bandpass output of each channel, the output is input to a low-pass filter. The low-pass filter output for each channel is resampled by a sampling means every audio frame period to obtain the components of the feature vector.

Now, if the output of the low-pass filter of the channel in the i-th audio frame is a', the feature vector a1 in the first audio frame is... It can be expressed as Here, K is the number of channels.

[Noise pattern calculation process in section (b)] This process is performed by the noise pattern calculation unit 23. For example, 10 consecutive audio frames (the number of audio frames) are is not woody), and this is called the noise period.

The feature vector of the noise section represents the spectral shape of the noise, and is especially called a noise vector and expressed as a town.

By the way, the average value of the noise spectrum within the noise section is calculated by , and this average value is called a noise pattern.

If the component of the noise pattern is Nk, then Z = (Nl
, N2. ... Hk, ... NK becomes 1.

[Section (C) A voice feature vector calculation process] This process is performed by the voice feature vector calculation unit 24.

After the noise section, that is, after the noise pattern calculation, the noise pattern image from the noise pattern calculation section 23 is subtracted from the feature vector aj output from the feature vector calculation section 22, and the speech feature vector is calculated using the following equation.

This processing in the processing unit 24 is a method for improving speech recognition performance in a high-noise environment, and is effective when noise continues relatively steadily.

[(d) Local peak vector calculation process for 'ff] This process is performed by the local peak calculation unit 25.

The audio feature vector Ibi sent from the audio feature vector calculator 24 is calculated by the local peak vector calculator 25.
is converted into a local peak vector fi at .

This conversion process will be explained with reference to FIGS. 3(A) to 3(C).

Each component b of the audio feature vector Toi is logarithmically transformed using the following equation.

Figure 3 (A) shows the logarithmic transformation x of this voice feature vector component.
An example of i(k) is shown, and the channel number k is plotted on the horizontal axis and Xi(k) is plotted on the vertical axis. This figure represents the shape of the logarithmic spectrum of the audio in the i-th audio frame. has been done.

Next, the least squares approximation straight line obtained by the linear equation Perform normalization using .

2 k (k) = x □ (k) - y k (k) = x・(
k) −ui(k) −k −v, (k)
(51 This normalized speech feature vector component ZL (
An example of k) is shown in FIG. 3(B), in which the channel number is plotted on the horizontal axis and 21(k) is plotted on the vertical axis.

Next, this z4(
k) to calculate the local peak vector lr4.

For k that satisfies the judgment condition of equation (8), r4=1
．． For k that is not satisfied, a vector ri having the value r□=0 as a component is calculated. This vector tri is called a local peak vector. This local peak vector tri
An example of this is shown in FIG. 3(C).

[Voice section detection processing in section (e)] This process is performed by the voice section detection section 26.

Using the audio feature vector lbi calculated by the audio feature vector calculation unit 24 for each audio frame, the frame power P1 of the audio frame is calculated.

In the speech section detection unit 2B, the speech feature vector lb
Voice section detection is performed using the frame power Pi obtained from i.

As mentioned above, various algorithms have been proposed for detecting speech intervals, but the purpose of this invention is not to use the algorithms themselves, but to detect speech intervals by subtracting the noise pattern N from the feature vector Therefore, for convenience of explanation, here, for convenience of explanation, the voice frame in which the frame power Pi is equal to or higher than the predetermined threshold Ps is defined as the start of the voice, and the frame power Pi from the start of the voice is defined as the threshold value. A voice frame that is less than Ps is considered to be the end of voice.

In Figure 4 (A) and (B), the input voice is [Sarapolo], and car noise is added as noise to this to reduce the S/N to 10.
Figure 4 (A
) is the frame power Pi calculated from the voice feature vector Toi in a no-noise environment, and (B) is the frame power P1' calculated from the feature vector ai by the same method in a noisy environment. The graph shows time plotted on the horizontal axis and frame power plotted on the vertical axis.

As can be understood from FIGS. 4(A) and (B), the change in frame power P1 obtained from the speech feature vector lbi that reduces the noise pattern varies depending on the period in which the speech is produced and the period in which the sound is produced. It has a clear distinction from areas where it is not. Therefore, voice section detection can be easily performed even in a noisy environment.

[Comparison pattern calculation and storage processing in section (f)] This processing is performed in the comparison pattern storage section 27.

In the specific speaker recognition method that limits speakers, words to be recognized (hereinafter referred to as categories) are uttered in advance,
It is necessary to store in advance a pattern (referred to as a comparison pattern) for expressing the word. Comparison pattern storage section 2
7, such comparison patterns are stored. The method for creating this comparison pattern will be explained below. The process of creating this comparison pattern is called registration process.

Here, for the sake of explanation, the number of categories is assumed to be M.

There is also a method of creating a comparison pattern by uttering the same category several times and taking the average of each pattern, but in this invention, a comparison pattern is created for one utterance of the category. .

The voice used to create the comparison pattern is called a learning voice.

Now, the m-th learning voice that has been digitized is input to the input terminal T-
21 to a feature vector calculation unit 22 to calculate a feature vector of the learning speech. On the other hand, noise, <Tan calculation unit 23
The noise pattern when no training speech has been input is extracted in advance. Therefore, the speech feature vector calculation section 24 subtracts the noise pattern from the noise pattern calculation section 23 from the feature vector from the feature vector calculation section 22 to calculate the speech feature vector of the learning speech.

Furthermore, the speech section detection unit 2B calculates the power of the learning speech and determines the start and end of the learning speech. Here, to make the explanation easier, the starting point of the learning voice is set as 1,
Let the terminal point be Jl.

The audio feature vector of the learning audio is determined by the local peak vector calculation unit 25 as a local peak vector! Sj
is converted to

Sj = (mSa + mSa +... 3,...
- sJmJ mj This local peak vector of the learning speech is particularly referred to as a comparison local peak vector.

Furthermore, the pattern represented by the time series of comparison local peak vectors from the start point 1 to the end point J1 is referred to as a comparison pattern, and S+? It is expressed as l.

Comparison pattern S□ for each category obtained in this way
its length Jl and the corresponding category name C, fI
It is also stored in the storage area of the comparison pattern storage section 27.

[Pattern similarity calculation process in section (g)] This process is performed by the pattern similarity calculation unit 28.

In contrast to the registration process of creating a comparison pattern as described above, the process of performing a recognition operation is called a recognition process. Therefore, the voice input during recognition processing is referred to as input voice.

The starting point calculated by the speech section detecting section 2B for this input voice is set to 1, and the ending point is set to ■.

In addition, the above-mentioned items (a) to (d) are also applied to the input audio.
The same or similar processing is performed to obtain the local peak vector rr4 (referred to as input local peak vector).
seek.

In this way, the input speech pattern expressed by the time series of input local peak vectors from the start point l to the end point ■ is called an input pattern, and is expressed by R.

Moreover, as already explained, the m-th comparison pattern Syn
is expressed as a time series from the start point 1 to the end point Jy11, and is stored in the comparison pattern storage section 27.

Next, a process for calculating the similarity between the input pattern R and the comparison patterns S, , l will be explained.

FIG. 5 is an explanatory diagram for establishing temporal correspondence between an input pattern and a comparison pattern when performing similarity calculation processing, where the horizontal axis represents the frame time point I of the input pattern, and the vertical axis represents the frame time point I of the input pattern. The frame time point j of the comparison pattern is taken and shown.

Therefore, when discussing the i-th frame time point of the '-1 input pattern and the fifth frame time point of the comparison pattern, we simply use the expression 'at grid point (i, j)'.

Input local peak vector lr1 at grid point (i, j) and comparison local peak vector, ll
The degree of similarity dr1 (il J) with 5j is defined as follows. Here, the right-hand subscript t represents the transposition of the vector.

Normally, complex calculations are required to calculate the degree of similarity expressed by equation (8), but since each vector ri, 5j in this invention is a local peak vector, its elements are O or l. Therefore, similarity calculation becomes extremely simple. In this sense, the method of expressing speech patterns in local peak vectors has important significance.

In addition, to simplify the calculation, we limited the calculation of the similarity in equation (8) to the range shown in Figure 5, and forced d + n (i, j) = O in other ranges. However, since it is not woody, its explanation will be omitted. Here, dyn
The range in which (i, j) is not forcibly set to 0 is called within the matching window.

Next, the cumulative similarity 111 at the grid point (i, j)
Let m(i, j) and path length Q, (i, j) be .

Here, the cumulative similarity Dm at the grid point (i, j)
In order to find (i l J ), lattice points (i ,
j), (i-1, j), (i, j-1), and grid points (i-2, j-1), (+-1, j-1
), (+-1, j-2) is required. In this sense, equation (8) is a recurrence equation. Therefore, an initial value is required.

The initial values required in this case are as follows.

becomes.

Equation (9) is used to calculate the cumulative similarity Dyn (i, j).

There are various calculation methods other than the calculation method given by the formula (lO). Other typical methods are shown below.

Recurrence Formula For this method, the path length Q, (i, j) is Qm (i, j
)=i, and there is no need to specially calculate Qyl(i, j).

Among the cumulative similarities at the grid point (i, j) obtained in this way, within the matching window and (i=I or j=
J) is the maximum cumulative similarity D among the cumulative similarities that are
max (II) is calculated, and this is called pattern similarity.

Perform the above calculations on all M comparison speech patterns,
Find M pattern similarities DITfax.

The fifth example of the prediction of the final path obtained in this way is
This is indicated by a curve X in the figure.

[Determination process in section (h)] The maximum value is determined again based on the cumulative similarity of M patterns.

mmax = tsrg max D, nax (m)
(12) l≦m≦M The category name Cm corresponding to the number m□d× of the comparison pattern whose maximum value is achieved, the vinegar outputs the recognition result to the terminal 30.
Output from.

As is clear from the above explanation, in the speech recognition method of the present invention, frame power is calculated using a speech feature vector obtained by removing noise patterns from input speech, and speech section detection is performed. Figure 4 (A) and (
As is clear from the comparison of the frame power Pi calculated using the voice feature vector and the frame power Pl' calculated using the unprocessed feature vector shown in B), there are fewer voice section detection errors, and in a noisy environment. It is possible to recognize the input voice with high accuracy even at the bottom.

Further, since the pattern similarity calculation process is performed using a local peak vector whose vector component is only O or I, which is calculated from the audio feature vector, the calculation process is extremely simple.

Furthermore, since the comparison local peak vector is used for the comparison pattern, the storage capacity thereof can be extremely reduced, and the speech recognition system can be downsized.

(Example) Hereinafter, an example of the present invention will be described with reference to FIG. 6.

FIG. 6 is a block diagram showing a specific circuit configuration for implementing an embodiment of the speech recognition method of the present invention.

In FIG. 6, 41 is a microphone, 42 is an amplifier for amplifying the audio signal, 43 is a low-pass filter,
44 is an A/D converter that converts audio into a digital signal;
45 is a signal processing processor that calculates a feature vector;
B is a processor; 47 is a program memory in which a processor program is stored; 48 is a comparison pattern memory for storing comparison patterns; 49 is a working memory;
is a noise pattern memory for storing noise patterns, 51
is an interface for outputting recognition results to the outside. However, although in a strict sense an interface circuit is required between each component, this is omitted here.

Next, an example of the speech recognition method of the present invention will be explained with reference to FIG.

After input audio from a microphone 41 is amplified by an amplifier 42, a low-pass filter (LPF) 43 removes its low frequency components.

Next, the input audio from which low frequency components have been removed is sampled into 12 bits by the A/D converter 44 at a sampling frequency of, for example, 12 kHz. The processing in the low-pass filter 43 described above is necessary for this sampling, and therefore, for example, a low-pass filter with a cutoff frequency of 5 kHz and an attenuation of 48 dB/oct is used.

The audio digital data sampled by the A/D converter 44 is converted into a feature vector by the signal processor 45. As this signal processing processor 45, for example, 32010 manufactured by TI can be used.

The processor 46 performs processing using the feature vector output from the signal processing processor 45 for each audio frame period, and the contents of the processing are divided into (1) registration processing (2) and recognition processing. Each of these processes will be explained below.

[registration process] This process is divided into the following processes.

Noise pattern calculation process Audio feature vector calculation process Comparison local peak vector calculation process Voice section detection processing Comparison pattern storage processing Each of these processes will be explained below.

(Noise pattern calculation process) For the registration process, for example, 10 audio frames are determined as a noise section. At this time, the speaker does not speak, and only ambient noise is input from the microphone 41.

This noise input is sent via signal paths (42, 43, 44) to a signal processing processor 45 from which a noise vector is generated which is sequentially stored in working memory 48. When noise vectors for 10 audio frames are stored in this memory 49, these noise vectors are averaged and the average value is stored in the noise pattern memory 50.

(Voice feature vector calculation process) After the end of the noise section, a speech feature vector is calculated by subtracting the noise pattern in the noise pattern memory 50 from the feature vector input from the signal processing processor 45,
This is stored in working memory 48.

Although this processing is performed for each audio frame period, the audio feature vectors until the start point is detected by the audio section detection processing are unnecessary, and therefore, in order to use the working memory 49 effectively, they can be discarded appropriately. To go.

(Comparison local peak vector calculation process) The audio feature vectors stored in the working memory 49 are
Through the processing in section d), it is converted into a comparison local peak vector and stored in the working memory 48. This process is also performed every audio frame period. In addition, comparison local peak vectors before the start edge detection are also discarded as appropriate.

(Voice section detection process) Frame power is calculated from the voice feature vector stored in the working memory 48.

The start and end of audio are determined by comparing this frame power with a threshold value.

(Comparison pattern storage process) Among the comparison local peak vectors stored in the working memory 49, the comparison local peak vectors from the start end to the end end are stored in the comparison pattern memory 48 as a comparison pattern.

[Recognition processing] This process is further divided into the following processes.

Noise pattern calculation processing Speech feature vector calculation processing Input local peak vector calculation processing Speech section detection processing Pattern similarity calculation processing Judgment processing (Noise pattern calculation processing) The noise situation has changed between the time of registration and the time of certification. Therefore, the noise pattern calculation is performed again.

It is best to calculate this noise pattern at σ before inputting a word, but it may slow down the input speed for inputting words, or
From the viewpoint of ease of utterance during the measurement of a single sound, it would be more practical to appropriately set up a special noise section and measure the noise pattern in that section.

As in the case of registration, certain 10 audio frames are defined as a noise section, and the speaker is prevented from speaking at this time. In this state, only ambient noise is input from the microphone 41 and sent to the signal processing processor 45 in the same manner as described above.
The noise vectors generated from this are sequentially stored in the working memory 49. When the noise vectors for 10 audio frames are stored, the average of these noise vectors is taken and this average noise vector is stored in the noise pattern memory 50.

(Speech feature vector extraction process) After the end of the noise section, the speech feature vector is calculated using a new noise pattern.

A voice feature vector is calculated by subtracting the noise pattern stored in the noise pattern memory 50 from the feature vector input from the signal processor 45, and is stored in the working memory 49. This process is performed every audio frame period. Furthermore, since the speech feature vectors before the start edge detection, which will be described later, are unnecessary, they are discarded as appropriate.

(Input local peak vector calculation process) Working memory 4
(d)
By processing the terms, the input local peak vector is converted into an input local peak vector and stored in the working memory 48.

This process is also performed every audio frame period. In addition, input local peak vectors before the start edge detection are also discarded as appropriate.

(Voice section detection process) Frame power P1 is calculated from the voice feature vector stored in the working memory 48. This frame power Pi
The start and end of the audio are determined by comparing the threshold and the threshold.

(Pattern similarity calculation process) Among the input local peak vectors stored in the working memory 49, the input local peak vectors from the start end to the end end are used as an input pattern, and this input pattern is compared with the input local peak vectors stored in the comparison pattern memory 48. Perform the pattern similarity calculation process in the above-mentioned (g) with the pattern,
As a result, D□Δ,(m) is stored in the working memory 49.

(Determination process) Pattern similarity degree D wa stored in the working memory 49
x (m), perform the determination process in the above-mentioned section (h), and the category name Cm obtained as a result.
, IIaχ are outputted to the outside through the interface 51.

Next, the results of an experiment to confirm the effects of this invention will be shown.

This experiment was conducted through a simulation using audio data obtained when a male speaker uttered 100 city names.

The simulation experiment used input audio in which vehicle noise was added to the audio data of 100 city names so that S/N I OdB was achieved, and training audio created by different utterances of 100 city names by the same speaker. We conducted the following four cases.

(Experiment I) Speech recognition method according to this invention The input pattern is created from the input local peak vector.

(Experiment 2) The input pattern was created from local peak vectors obtained directly from the feature vectors of the input speech.

(Experiment 3) The input pattern uses the audio feature vector of the input audio.

(Experiment 4) The input pattern uses the feature vector of the input voice. The comparison pattern is created by the same process as the input pattern for learning, and the pattern similarity calculation process is performed by the above-mentioned item (g) of this invention, and the determination process is performed by the process of the above-mentioned item (h) of this invention. I used it.

The above results are shown in the table below.

Experiment 1 (this invention): 95% Experiment 2 85% Experiment 3 87% Experiment 4 35% From the results of this experiment, it is clear that the recognition method of this invention, which uses local peak vectors calculated from voice feature vectors, is effective in noisy environments. However, it was confirmed that the present invention is effective in that voice recognition can be performed more accurately than before.

(Effects of the Invention) As is clear from the above description, the present invention provides the following effects.

■Since the frame power is calculated using the speech feature vector from which the noise pattern has been removed from the input and the speech section is detected, there are fewer errors in speech section detection, and therefore the recognition accuracy of the input speech is improved even in noisy environments. Improved than before.

■Since the pattern similarity calculation process is performed using a local peak vector that is calculated from the voice feature vector and whose components are only O or l, the calculation process when implementing the voice recognition method of this invention is extremely simple. Become.

■Since the comparison local peak vector is also used for the comparison pattern, its storage capacity is extremely small. Therefore, in addition to the above-mentioned effect (■), the structure of the device for implementing the recognition method of this invention is simple and compact. becomes.

[Brief explanation of drawings]

Fig. 1 is a block diagram for explaining the recognition processing of the speech recognition method of the present invention, Fig. 2 is a diagram showing the characteristics of the bandpass filter used for speech analysis processing, and Fig. 3 is for explaining local peak vector calculation. An explanatory diagram for. Fig. 4 is a diagram showing the state of frame power, Fig. 5 is an explanatory diagram for explaining the algorithm for calculating cumulative similarity, Fig. 6 is a block diagram showing an embodiment of the present invention, and Fig. 7 is a diagram showing the conventional To explain the speech recognition method. It is a block diagram. 21... Input terminal 22... Feature vector calculation unit 23... Noise pattern calculation unit 24... Voice feature vector calculation unit 25... Local peak vector calculation unit 2B... Voice section detection unit 27 ... Comparison pattern storage section 28 ... Pattern similarity calculation section 29 ... Judgment section, 30 ... Output terminal 41 ... Microphone, 42 ... Amplifier 43 ... Low pass filter, 44 ...・A/D converter 45...signal processing processor 46...processor, 47...program memory 48...comparison pattern memory, 49...working memory 50...noise pattern memory 51... ·interface. Patent applicant: Oki Electric Industry Co., Ltd. Procedural amendment June 25, 1986

Claims (1)

[Claims] (1) (a) A process of frequency-analyzing input speech and calculating feature vectors, which are vectors of frequency components of the input speech, at fixed time intervals called speech frames, and (b) A predetermined process that is known in advance to be only noise. (c) a process of calculating a voice feature vector by subtracting the noise pattern from the feature vector calculated for each voice frame; (d) For each audio frame, calculate a least squares approximation straight line from the audio feature vector, set the component corresponding to the channel that is maximum in the frequency direction based on the least squares approximation straight line to 1, and set the other components to 1. (e) calculating a frame power in the audio frame from the audio feature vector and detecting the start and end of the audio using the frame power; , (f) Calculate a comparison pattern in advance using the same or similar processing as the above-mentioned (a) to (e) for the training speech uttered once or multiple times for each recognition target word, and perform the comparison. (g) Non-linear matching processing between the input pattern obtained by the processing of the above (a) to (e) for the input speech to be recognized and the comparison pattern. (h) calculating the pattern similarity between the input pattern and the comparison pattern by performing A speech recognition method characterized by comprising a process of outputting a category name as a result.