JPS6075898A - Word 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 [Technical Field of the Invention] The present invention relates to a word voice recognition device, and particularly to a word voice recognition device for controlling various devices or inputting data by voice, for example. The present invention relates to a word speech recognition device, and more particularly to a word speech recognition device that recognizes voiced sounds using the frequency spectrum of voiced sounds as one of the characteristics of speech.

[Prior Art] FIG. 1 is a schematic block diagram showing an example of a conventional word speech recognition device (hereinafter simply referred to as a recognition device). In the figure, an audio input section 1 includes a microphone, an amplifier, a low-pass filter, etc. (though not shown), and converts audio into electrical signals for input. The output of the audio input section 1 is given to the feature extraction section 2 and also to the start n-end detection circuit 6. The feature extraction unit 2 analyzes the input audio signal and extracts audio feature parameters. The voice feature parameters extracted by the feature extraction section 2 are given to the recognition processing section 5. The start and end detection circuit 6 is a circuit that detects the start and end of word speech.

The detection result of the start/end detection circuit 6 is given to the recognition processing section 5. The recognition processing section 5 is composed of a microprocessor, a microcomputer, etc., and performs speech recognition processing. An input pattern memory 3 and a registered pattern memory 4 are connected to the recognition processing section 5.

In the above-mentioned recognition device, the audio waveform is divided into frames of a fixed time, and the frequency spectrum of each frame is extracted as a characteristic parameter. and,
In the registration mode, the recognition processing section 5 writes the extracted feature parameters of the registered word or feature parameters of the standard speech into the registered pattern memory 4. That is, the registered pattern memory 4 stores in advance feature parameters of multiple gB voices. Furthermore, in the acoustic recognition mode, the recognition processing unit 5 writes the feature parameters of the extracted word sounds into the input pattern memory 3. Then, the degree of similarity between the feature parameters stored in the input pattern memory 3 and the feature parameters of multiple words stored in the registered pattern memory 4 is calculated in sequence, and word sounds are recognized based on the calculated results. Do the following.

FIG. 2 is a circuit diagram showing details of the feature extraction section 2 shown in FIG. 1. In the figure, the audio signal from the audio input section 1 is
Bandpass filter 201-1, 201°-2...20
1-N. These bandpass filters pass specific frequency components of the audio signal waveform.

The output of each bandpass filter 201-1 to 201-N is
The signals are respectively applied to smoothing circuits 202-1 to 202-N. The output of each smoothing circuit 202-1 to 202-N is given to an analog multiplexer 203. This analog multiplexer 203 includes each smoothing circuit 202-1 to 202-
This is a circuit that passes the output of N in a time-division manner. The output of the analog multiplexer 203 is A/D conversion circuit 20
3 and is converted into digital data and output.゛Figure 3 shows the bandpass filters 201-1 and 201-2 shown in Figure 2.
It is a figure which shows the frequency characteristic of 01-N.

As shown in FIG. 3, the N filters are set to extract almost equally all frequency components of the audio waveform. In this case, the characteristics of the voice are expressed by the magnitude pattern of the N11l values of the frequency components extracted by the N11l filter. The above N is usually 8~
At 16, there is noise in the audio waveform.

If the voice is not mixed in, relatively good voice characteristic parameters can be obtained. Therefore, the recognition performance was also sufficiently satisfactory. However, if the audio is mixed with noise such as factory noise or other people's voices,
At the same time as the voice, the frequency components of the noise also pass through the bandpass filter, which affects the values of the characteristic parameters.

If the extraction accuracy of feature parameters is evaluated based on spectral distortion, in conventional recognition devices, spectral distortion due to noise in the input waveform will appear as it is in the feature parameters. Therefore, when the conventional recognition device is used in a noisy environment, the recognition performance deteriorates significantly.

[Summary of the Invention] This invention was made to eliminate the drawbacks of the conventional recognition device as described above, and by configuring the feature extraction section using a digital filter adapted to the bits of the voice, The purpose of this invention is to provide a speech recognition device that has excellent recognition performance even in a noisy environment.

Hereinafter, this invention will be described in more detail with reference to embodiments shown in the drawings.

[Embodiment of the Invention] FIG. 4 is a schematic block diagram showing an embodiment of the invention. In the figure, the audio input section 10 includes a microphone 11
, microphone amplifier 12, AGC circuit 13, and A
, /D conversion circuit 14 and a waveform memory 15.

The output of the audio input section 10 is given to the level calculation circuit 7 and also to the feature extraction section 20. The output of the level calculation circuit 7 is given to the start/end detection circuit 6 as well as to the recognition processing section 50. The output of the start/end detection circuit 6 is given to the recognition processing section 50. On the other hand, feature extraction section 2
0 includes a pitch period extraction circuit 21, a filter coefficient setting circuit 22, and a digital filter 23. The pitch period extraction circuit 21 and the digital filter 23 include:
The output of the aforementioned audio input section 10 is given. The output of the pitch period extraction circuit 21 is given to the recognition processing section 50 and also to the filter coefficient setting circuit 22. The output of the filter coefficient setting circuit 22 is a digital filter 23
given to. The output of the digital filter 23 is given to the recognition processing section 50. An input pattern memory 3 and a registered pattern memory 4 similar to the circuit shown in FIG. 1 are connected to the recognition processing section 50.

Next, the operation of the embodiment shown in FIG. 4 will be explained. The input waveform of the voice captured by the microphone 11 is amplified by the microphone amplifier 12, adjusted by the AGC circuit 13 so that the highest value of the waveform is at a certain level, and then sent to the A/D conversion circuit 1.
4, each sampling point is converted into a digital value. The sampling data for one frame is stored in the waveform memory 15.
is temporarily stored.

The level calculation circuit 7I3 and the feature extraction unit 20 calculate the data X (+ >, (i = 1.2...■) in the waveform memory 15.
f) to perform the following processing.

First, the level calculation circuit 7 calculates, as shown in the following equation (1),
Calculate the square P of the sampling data and find a value corresponding to the power of that frame.

F P=Σ x (1)'' (1) i·1 This numerical value P is given to the recognition processing unit 50 and used to determine whether the input waveform signal is a voiced sound.

Next, the pitch period extraction circuit 21 calculates the autocorrelation function value C of the waveform data in the waveform memory 15, as shown in the following equation (2).
OR(τ) is calculated, and the pitch period is set as τ that gives the maximum self-function value within the pitch period search range.

(2) The filter coefficient setting circuit 22 generates a filter coefficient such that an integral multiple of the pitch frequency (reciprocal of the pitch period) becomes the resonance frequency of the digital filter 23, and sets the generated filter coefficient to the digital filter 23. Set to . Note that this filter coefficient setting circuit 22 is realized by means of configuring a filter coefficient table in a ROM or the like and searching the contents of the ROM in correspondence with the pitch frequency and its integral multiple.

FIG. 5 is a block diagram showing an example of the configuration of the digital filter 23 shown in FIG. 4. In the figure, the output x (1) of the waveform memory 15 in FIG. This first-order difference circuit 231 performs, for example, subtraction!
It is composed of 1, etc., and is used to strongly suppress high frequencies. The output of the first-order difference circuit 231 is sent to a two-stage lattice filter 232. This two-stage lattice filter 232 includes three adders and subtracters 2321 to 2323, and three
It is constructed by including multipliers 2324 to 2326 and delay circuits 2327 to 2328 of 211I. The output of the two-stage lattice filter 232 is given to a square circuit 233.

The output of this square circuit 233 is given to an integration circuit 234. The output 5(n) of this integration circuit 234 is given to the recognition processing section 50 as a filter output.

Next, the operation of the digital filter 23 described above will be explained. The summing data x(i) stored in the waveform memory 15 of FIG. 4 is input to the primary valve circuit 23'I of the digital filter, where t1 of the following equation (3) is performed.

Δ>+ (i)=x (i)-x (i-1)...
-<3) The output Δχ (1) of the first-order difference circuit 231 is given to the two-stage lattice filter 232. This two-stage lattice filter 232 sequentially calculates the following equations (4) to (7).

Vz<i)-Δ×(i)+)<2(n)-112(1-1)...(4) V+(+>=y:(i)+KI(n>・b,(i 1) ) ...(5) 112 (+ >=+1. (i -1)i<, (n>
・y, (i) ... (6) lz <1>-V
, (+ > ... (7) Two-stage high-order filter 23
The output y+(i) of 2 is the square circuit 233 and the integration circuit 2.
At 34, the following equation (8) is calculated.

f s(n)−Σ y + (i) 'y + (+ >
-' (8) -Å As described above, the filter output 5(n) at 1° C. is derived. However, the initial values of 1 and b2 (1), b, and (1) are 0. Further, 11 means the nllth count 1 shift set by the filter coefficient setting circuit 22, which is the nllth count of the pitch frequency.
Counts corresponding to harmonics (also direct).

The two-stage lattice filter 232 has a resonance frequency f. .

Band width B. It has a resonance characteristic of filter coefficient Kl (
11), the following equation (9
), (10).

K+(11)←cos 2π<ra/fs) ・・・(
9) K2 (n) J=F-exp (-2π80/f
s)...(10) However, f is the sampling frequency.

When K2(n)→-1t, that is, the band width B0 is extremely small, the characteristic shown in FIG. 6 with a high Q value is obtained. The arithmetic processing of the digital filter 23 is performed using the same waveform data x
(i) is executed N times, and N outputs Js (n
) (n = 1.2. -N) is obtained. As already mentioned, the filter coefficients 1(n) and K2(11> are set by the filter coefficient setting circuit 22 so that the frequency component of the pitch frequency matches the resonance frequency of the filter, so the filter output 5(n) ) corresponds to a value obtained by extracting only the pitch frequency harmonic component included in the waveform data x (i).This filter output 5(n) is given to the zero-identity processing unit 50 shown in FIG. , used as the main data for recognition processing.

151 The processing unit 50 normalizes the amplitude 5 time axis of the feature parameter 5(n) given from the digital filter 23, and then writes the normalized spectral time series pattern to the registered pattern memory 4 in the registration mode, and
In the l mode, the data is written to the input pattern memory 3. Further, in the recognition mode, the recognition processing section 5 calculates the degree of similarity between the contents of the registered pattern memory 4 and the contents of the input pattern memory 3 by pattern matching to obtain a recognition result. Note that the start/end detection circuit 6 detects the start/end of the audio signal based on the power calculated by the level calculation circuit 7. These operations are substantially similar to those of the circuit in FIG.

Next, the features of the embodiment shown in FIG. 4 will be explained. One of the features of this embodiment is that the level of the waveform is calculated for each frame by digital processing, and significant parts of the speech waveform, that is, frames of vowels, are detected. Another feature is that the pitch period for such a vowel frame is determined by means such as an autocorrelation method. Yet another feature is that only the harmonic components of the pitch frequency are extracted by providing a resonant digital filter. As a general feature of speech waveforms, voiced sounds such as vowels have high power, and the proportion of speech information that is masked by high noise contamination is small. Furthermore, if a pitch extraction method such as an autocorrelation method is used, the pitch period can be extracted with high accuracy even if white noise is mixed in. In addition, voiced sounds such as vowels have an integer pitch frequency (Il1, which has a component only at 8).
This spectrum pattern is effective information for vowel identification. therefore,
The output of the digital filter 23 that resonates at the frequency of the integer f8 of the pitch frequency extracted by the feature extraction unit 20 of the recognition device shown in FIG. 4 is a feature parameter that directly expresses the features of the vowel. Furthermore, even if high noise is mixed in, most frequency components of the noise are blocked by the digital filter 23 and are not output. Therefore, even if the input speech waveform has large spectral distortion due to noise, the feature parameters are less likely to be distorted and can be effective feature parameters for recognition. therefore.

In the recognition device shown in FIG. 4, deterioration in recognition performance due to noise can be minimized, and recognition performance can be improved.

FIG. 7 shows J of the digital filter 23 shown in FIG.
7 is a block diagram showing another configuration example. Note that this FIG. 7 has the same configuration as the digital filter shown in FIG. 5 except for the following points, and corresponding parts are given the same reference numerals and their explanations will be omitted. . In this example, 1
A feature is that a multiplier 235 is provided before the next difference circuit 231, and the audio waveform can be enlarged or reduced. Multiplier 235
By inserting , the above equation (3) becomes the following equation (11
).

Δ× (1)-α(y, (i)-x (i-1))・
(11) Here, α is a waveform multiplication coefficient whose value can be set arbitrarily. If the power of the audio waveform is too large, the fAt'i result 5 (TI> shown in equation (8) above) may overflow. Therefore, if the power of the waveform is large, α should be small; For example, if α is increased, the dynamic range of filter operation is improved.The power value calculated by the level calculation circuit 7 can be used.

Let P be the power that guarantees that the filter expansion will not diverge when α − α and α are constant, then when the power is P and below, α = α, (12) When the power exceeds PH By setting α=U+Cr17v (13), divergence in spectrum calculation can be prevented. In addition, the above-mentioned No. (
In formula 13), P represents power. The relationship between α and power is shown in FIG. When the recognition processing unit 5o receives power from the level calculation circuit 7, it calculates α and sets this α in the multiplier 235.

FIG. 9 is a block diagram showing still another example of the configuration of the digital filter 23 shown in FIG. 4. Note that the same parts as in the digital filter shown in FIG. 5 are given the same reference numerals, and the explanation thereof will be omitted. This practical example is characterized in that a multiplier 236 is inserted after the integration circuit 234 to adjust the frequency characteristics of the filter. The gain GN (n '') of the two-stage lattice filter 232 is expressed by the following equation <14
).

Note that in the above equation (14), B7. B2 is the following formula (
15) and (16).

B+ = to + (n) = to + (II > ・K2
(n)...(15) 82 = 2 (n L - (16) The filter coefficient setting circuit 22 calculates K+(n>, 2(n)
In addition, a gain correction coefficient expressed by the following equation (-17>) is set in the f! calculator 236.

G(n) = 1/(GN2 (n)) - (17) In this way, by providing the multiplier 236 and multiplying the spectrum fls(n) obtained by the filter y4 by G(N), the gain is constant. It is possible to obtain the result of filter calculation.

As a result, recognition performance can be improved.

In the above embodiments, the recognition device was described as a specific speaker registration type for convenience of explanation. It goes without saying that this can also be achieved with equipment.

Furthermore, in the above embodiments, the digital filter was mainly explained as a two-stage lattice filter, but if it is a horizontal digital filter with a high Q that can change the characteristics for each frame, a two-stage lattice filter can be used. You don't have to.

Furthermore, in the above embodiment, the filter coefficients are set and used in one digital filter in a time-sharing manner, but it is also possible to provide multiple digital filters in parallel and set different filter coefficients to each digital filter at the same time. Set,
Thereby, the resonance frequency of each digital filter may be an integral multiple of the pitch frequency.

Furthermore, in the embodiment shown in FIG. 7, the waveform multiplication coefficient α is determined by the power of the audio signal, but in addition to the power, other quantities (for example, level) corresponding to the waveform size of the audio signal are calculated, and α may be set by the calculated amount.

Furthermore, in the embodiment shown in FIG. 9, the gain correction coefficient is determined so that the gain of the filter output is completely constant; however, the gain of the filter output does not have to be completely constant, and some variation may occur. There may be.

[Effects of the Invention] As described above, according to the present invention, only the harmonic components of the pitch frequency of the audio signal are extracted as feature parameters, so that the input audio signal has large spectral distortion due to noise. Even if the noise is different, the feature parameters are less likely to be distorted, and therefore an excellent recognition device with very little deterioration of recognition performance due to noise can be obtained.

[Brief explanation of drawings]

FIG. 1 is a schematic block diagram showing an example of a conventional recognition device. FIG. 2 is a block diagram showing details of the feature extraction section 2 shown in FIG. 1. Figure 3 is shown in Figure 2? It is a figure which shows the frequency characteristic of tl pass filters 201-1-201-N. FIG. 4 is a schematic block diagram showing one embodiment of the present invention. FIG. 5 is a block diagram showing an example of the digital filter 23 shown in FIG. 4. , l FIG. 6 is a diagram showing the frequency characteristics of the two-stage lattice filter 232 shown in FIG. FIG. 7 is a block diagram showing another example of the configuration of the digital filter 23 shown in FIG. 4. FIG. 8 shows the power of the audio signal and the multiplier 235 shown in FIG.
FIG. 3 is a diagram showing the relationship between the waveform multiplication coefficient α and the waveform multiplication coefficient α set to . FIG. 9 is a block diagram showing still another example of the configuration of the digital filter 23 shown in FIG. 4. In the figure, 3 is an input pattern memory, 4 is a registered pattern memory, 7 is a level calculation circuit, 10 is an audio input section, 2
0 is a feature extraction unit, 21 is a pitch period extraction circuit, 22 is a filter coefficient setting circuit, 23 is a digital filter, 50
2 shows a recognition processing unit, and 232 shows a two-stage lattice filter. Agent Masu Oiwa Figure 1 Figure 3 ■ Figure 6 Procedural amendment (self-initiated) 1. Indication of the incident Japanese Patent Application No. 183842/1982 2 Name of the invention Word speech recognition device 3 Representative of the person making the amendment Part 4 to Hitoshi Katayama, Agent 5, Subject of amendment Column 6 of detailed explanation of the invention in book IB, Contents of amendment (1) "Higher order form" on page 12, line 13 of the specification has been changed to "lattice" (2) Amend “Do” on page 16, line 15 of the specification to “do.” (3) Add + (n to rB+- on page 18, line 16 of the specification).
) − to + (n) − to 2 (n) J
[B, − to + (n) − Kl (n) − to 2
(n) Corrected to J. Written amendment to the above procedure (voluntary) Mr. Commissioner of the Japan Patent Office 1. Indication of the case: Japanese Patent Application No. 58-183842 2. Name of the invention: Word speech recognition device 3. Person making the amendment 5. Figure 6 of the drawing to be amended. Contents of the amendment Figure 6 of the drawings will be amended as shown in the attached sheet. that's all

Claims (1)

[Scope of Claims] 1) Audio signal input means for converting audio into an electrical signal and inputting it; feature extraction means for extracting feature parameters of the audio signal waveform input from the audio signal input means; and the feature extraction means. an input back storage means for recording the feature parameters of the native language speech to be recognized extracted by the feature extraction means; a registration pattern for storing in advance the feature parameters of the plurality of word sounds extracted by the feature extraction means storage means,
and a speech recognition process that calculates the degree of similarity between the feature parameters of the input speech stored in the input cover pattern storage means and the feature parameters of the plurality of word sounds stored in the registered pattern storage means, and performs speech recognition processing. The feature extracting means includes means for detecting the pitch frequency of the audio signal inputted from the audio signal inputting means, and a resonance frequency of which changes according to a set filter coefficient, and a digital filter for extracting the spectrum data of No. 18 using the characteristic parameters; and means for setting the number of filters of the digital filter so that the resonance frequency of the digital filter is Mvi times the pitch frequency. Including, word speech recognition device. (2) The word speech 12 according to claim 1, wherein the digital filter is digitized by one, and the filter coefficient setting means sets the filter coefficient to the digital filter in a time-sharing manner. recognition device. (3) A plurality of the digital filters are arranged in parallel, and the filter coefficient setting means sets a different filter coefficient to each of the digital filters arranged in parallel. The word speech recognition device according to item 1. (4) The digital filter according to any one of claims 1 to 3, wherein the digital filter further includes means for adjusting the level of the input audio signal according to the level of the input audio signal. ! iSpeech recognition device. (5) The digital filter further includes means for adjusting the level of the output signal according to the resonance frequency so that the level of the output signal at each resonance frequency is constant. The word speech recognition device according to any one of Item 3.