JPS61141500A - 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] "Industrial Application Field" This invention calculates the degree of similarity between a time series of features of an input speech and a time series of a corresponding standard pattern speech, and calculates the degree of similarity between the elements of these time series. The present invention relates to a speech recognition device that calculates the distance between.

``Prior art'' In this type of conventional speech recognition device, the time series of the feature values of the input voice = 4 + aI2p ..... and the time series of the feature values of the standard pattern voice [3 = J + "2'...
... Tom, the distance measure between its element ai and , was determined using a spectrum in which the total power in each element was normalized to a constant value. For this reason, the element cho is a vowel and the element 1bj is a prelude, so the voice power is extremely 2.
The disadvantage is that it is possible for the distance between two different elements to become quite small.

In speech recognition, it is desirable to be able to recognize the voices of as many people as possible using one standard pattern, but many of the conventional methods use distances based on spectra (distances) that are susceptible to individual differences. Errors such as erroneous recognition of words and previews are likely to occur, and the recognition rate is particularly low when the input speech is not from the same speaker as the standard pattern.

To deal with this problem, attempts have been made to use a matching distance measure between elements of speech time series that combines spectral information and power information (e.g., References, Ayo, Kano,
Furui: Word speech recognition using distance weighted with bow information, Acoustical Society of Japan speech study group material, 581-59.198
(see 1). However, with this method, it is necessary to deal with fluctuations in power information due to the input level of the voice, and it is also necessary to check the maximum and minimum values of the power of the entire voice section and normalize the voice power so that this becomes a constant value. was there. For this reason, it is necessary to check the change in voice power until the end of the speech interval or for several seconds, and then start calculating the distance measure. This may cause a time delay before the recognition result is obtained, or There was a drawback that a high-speed element had to be used for calculation.

An object of the present invention is to provide a single word speech recognition device that can recognize speech of unspecified speakers more accurately and without delay time than conventional methods.

"Means for Solving the Problem" According to the present invention, the distance of the linear regression coefficient of the time waveform of voice power, which has little individual difference and is not affected by fluctuations in the voice input level, is defined as the spectral distance. Combined. In other words, the spectral distance Ds between the elements of the feature time series of the input speech and the standard pattern speech is determined, and the time waveform of the linear regression coefficient is calculated from the time waveform of the audio power for the input speech and the standard pattern speech, respectively. ', These linear regression coefficients are used to determine the power regression coefficient distance Ds between elements of the input voice and the standard pattern. Using the spectral distance Ds, the power regression coefficient distance, and the inter-element matching distance, the degree of similarity between the input speech and the standard pattern speech is calculated by time-normalized matching.

``Embodiment'' Figure 1 shows an embodiment of the present invention, in which the audio input terminal
The input speech signal applied to the child 1 is first subjected to a speech section detection circuit 2 in which silent (noise) sections are removed and only actual speech sections are extracted. This voice section can be detected by using several well-known methods, such as short-term power of the human voice signal, and time during which the power continues to exceed a certain value. The signal wave of the detected voice section is sent to the linear prediction analysis circuit 3, where it is converted into a linear prediction coefficient and a time waveform of power. This technique is already known (see, for example, literature, board book, Saito: Estimation of speech spectral density and formant frequency using statistical methods, Transactions of the Institute of Electronics and Communication Engineers, 53-A, 1゜P35.1970).
, details are omitted, but basically, first pass it through a low-pass filter, then sample and Europeanize it, multiply the waveform in a short period by a Hamming window etc. at regular intervals, cut it out, and calculate the sum of products. Calculate the power and correlation coefficient. For example, a value such as 3Qm3 is used as the length of the Hamming window, and a value such as 1Qms is used as the period for updating this window. A linear prediction coefficient is extracted from this correlation coefficient by solving an algebraic equation through repeated calculation processing. For example, values from the first order to the tenth order are calculated as the linear prediction coefficients.

The time waveform of the extracted linear prediction coefficient is converted into a linear prediction cepstrum coefficient by the cepstral conversion circuit 4C. This technique is already known (for example, literature, Saito, Nakata, Basics of Two-Simple Voice Information Processing, Ohmsha, Chapter 7, P.
2O3, 1981) Although the details are omitted, linear prediction cepstral coefficients (hereinafter referred to as cepstral coefficients for simplicity) can be easily obtained by recursive calculation using linear prediction coefficients. The extracted cepstral coefficients are stored in the feature parameter register 5 (knee register).

On the other hand, the time waveform of the power, which is the other characteristic extracted by the linear prediction analysis circuit 3, is logarithmically transformed into a waveform of a certain time length section at each extraction period (l Q m s in the above example). Thereafter, the data is stored in a power register 6, and the contents of this register 6 are sent to a regression coefficient calculation circuit 7, where a linear regression coefficient is calculated. For example, a value such as 5Qms is used as the length of the time waveform input by the register 6 and the regression coefficient calculation circuit 7. M)
If so, this linear regression coefficient a (under U, this is called a power regression coefficient) can be obtained by the following calculation.

The power regression coefficient is calculated according to the input of the regression coefficient calculation circuit 7, which is updated every cycle as described above, and is stored in the feature parameter register 5 together with the cepstrum coefficient.

The switch 8 is a switch for selecting the learning mode and the recognition mode. First, the switch 8 is connected to the terminal 8a, and then the user who inputs the voice to be recognized, or multiple users different from the user, A feature parameter waveform consisting of cepstrum coefficients and power regression coefficients is obtained from the speech for each word to be recognized, stored in the feature parameter register 5, and then input to the standard pattern storage section 9, where it is stored as a standard pattern for that word. .

For the subsequent speech to be recognized (2), switch 8 is connected to terminal 8b, and the contents of feature parameter register 5 are input to time normalization matching circuit 10. At the same time, standard patterns corresponding to each language are input. Each pattern is read out from the standard pattern storage section 9 and input to the time normalization matching circuit 10.The time normalization matching circuit 10 calculates the degree of similarity of the feature parameters between the standard pattern and the human voice.

The speaking rate of speech changes each time the same word is uttered repeatedly by the same speaker (bipartite and global), so to compare the two, it is necessary to make sure that the common sounds (phonemes) correspond. In order to do this, it is necessary to appropriately expand or contract one of the time axes in a non-linear manner to match the time axis of the ability, and compare the characteristic parameters at corresponding points in time.

As a technique for non-linearly expanding and contracting the time axis of the ability so that the two best match each other based on one of them (so that the degree of similarity between the two is maximized), dynamic programming 5 U is used to achieve the optimal . Using method 3. 1 (Reference, Sakoe, Chiba: Continuous word recognition based on temporal normalization of speech using dynamic programming, Journal of the Acoustical Society of Japan, 27, 9.
P2S5, 1971).

Also in the apparatus of the present invention, the time normalization matching circuit 10 performs dynamic programming calculations. Let the cepstral coefficient at a certain point k of the standard pattern be cl(1≦i≦p, p
(as mentioned above, use a value such as 10), the power regression coefficient aζ, and the point in time of the input audio! If the cepstral coefficient is 0 (1≦i≦p) and the power regression coefficient is
The following values are used as the distance from the input voice (a numerical value indicating that the smaller the degree of similarity is) DS(k, ri, Dp(k, j)).

D, (k, jri=(aζ-ri)2...
...(3) Next, calculate the weighted average of both as shown below. Performs dynamic programming operations.

1) (k, ri = -1iiD, -(-M-, li +
(1-W) Dp (k, j)・・・・・・・・・(
4) The weight W used in this equation has a value between 0 and 1, and this value is set to an appropriate value based on the results of preliminary experiments and stored in the weight register 11 so as to obtain relatively high recognition accuracy. I'll keep it.

The inter-element matching distance between the standard pattern and the input voice at corresponding points when the time axes are matched so that the degree of matching between the standard pattern and the human voice is the best is determined by dynamic programming calculations. Calculate the average value for. This value will be referred to as the overall distance between the standard pattern and the human voice.

The total distance between the standard pattern corresponding to each word work and the input speech is input to the comparison circuit 12, and the logic circuit determines the word plan with the smallest total distance among all these total distances. This determination result is output from the output terminal 13.

By the way, the temporal waveform of speech power has a basic property that it is high in the vowel part and low in the consonant part, and this property does not change even if the speaker is different. Figure 2 shows the time waveform of the logarithmic power of the word 1900 uttered twice by each of the four speakers, and the logarithmic power 1-hour waveform is normalized so that the maximum and minimum values are constant. It shows. This second
As can be understood from the figure, the power one-hour waveform does not differ much even if the speaker changes, and it changes relatively smoothly over time. If we calculate the linear regression coefficient of the time waveform, that is, the slope when linearly approximated, this value will not be affected by the overall constant I increase or decrease of the time waveform due to the principle of the linear regression coefficient. It is possible to extract stable word features that are common to the two and are not affected by fluctuations in the vocalization level.Therefore, as in this example, the power regression coefficient is combined with the cepstral coefficient to perform time-normalized matching of the standard pattern and the input speech. If this is done, parts that are similar in both spectrum (cepstrum) and power will be matched, and matching between vowels and consonants can be avoided, and the recognition rate can be improved.

As a result of this structure, it is possible to use stable features contained in the power time waveform without normalizing the power time waveform by examining the maximum and minimum power values in the entire speech interval. As a result, it is possible to realize a word speech recognition device that can start calculations for recognition immediately when speech is input, and can output recognition results with high precision for anyone's voice without any time delay. According to previous experiments (2), when 100 words of city names were recognized and the speech of one speaker different from the person was used as a standard pattern, the recognition rate of a conventional device using only cepstral coefficients was 85. .5%, whereas the device of this example achieved a recognition rate of 89.3%, confirming that the present invention is superior.

By calculating the linear regression coefficient b (referred to as the cepstrum regression coefficient) of the cepstrum coefficients, and performing time-normalized matching of the similarity between the input speech and the standard pattern speech using the cepstral coefficients, cepstral regression coefficients, and power regression coefficients. , an even higher recognition rate can be obtained.

FIG. 3 shows this example, and parts corresponding to those in FIG. 1 are given the same reference numerals. The cepstrum coefficient Cn calculated by the cepnutrum conversion circuit 4 is directly supplied to the feature parameter register 5, and the time waveform of this cepstrum coefficient Cn is temporarily stored in the cepstrum register 14 in sections of a certain time length at regular intervals. The contents of this register 14 are sent to a regression coefficient calculation circuit 15, where a linear regression coefficient (cepstral regression coefficient) is calculated. The length of the time waveform input to the cepstrum register 14 and the regression coefficient calculation circuit 151 is, for example, 50 m5, and the period for updating this is, for example, a value such as 10 m5. Time waveform of cepstrum coefficient 1yj (j=-M,
...Taku・M), this cepstral regression coefficient can be found by the following calculation.

The cepstrum regression coefficients are calculated for the cepstrum coefficients of each order according to the input of the regression coefficient calculation circuit 15, which is updated every lQ, ms, and together with the cepstrum coefficients, the cepstrum regression coefficients are calculated in 2P dimension. The feature parameters are sent to the feature parameter register 7 (= sent and stored. In the time normalization matching circuit 10, the cepstrum coefficients and cepstrum regression coefficients at a certain time point k of the standard pattern are rki (1≦i≦2p), Let xzi (1≦i≦2p) be the cepstrum coefficient and cepstrum regression coefficient in , then the following value is used as the distance between the two (a numerical value indicating that the smaller the degree of similarity is).

1 = up to 2p if the order of the cepstral coefficient is P,
The order of the cepstral regression coefficient is P, and the total of both is 2
This is because it becomes the order of P. Here, Wi is a numerical value indicating a predetermined weight for each coefficient, and this value is set to an appropriate value based on the results of preliminary experiments to obtain relatively high recognition accuracy, and is set in the weight register 16. Save it up. The distance d is calculated as the sum of squares of the difference between the input speech and the standard pattern for the P-th order cepstrum coefficient and the P-th order cepstrum regression coefficient at the same point in time, as shown in equation (6), that is, the cepstrum coefficient and cepstrum regression coefficients, which have different properties, are used, and Wl is weighted to balance them. Therefore, when calculating the cepstrum coefficient as a town value, At least two values are used: Wa used and Wb used when calculating the cepstral regression coefficient.

These weights Wa-Wb are stored in the weight register 16.

The time normalization matching circuit 10 further calculates the distance d between the standard pattern obtained from equation (6) and the input audio at the time point.
Equation (4) is calculated by using (k, ri) as Ds(k, ri) in equation (4). The operation of the pond is the same as in the case of FIG.

Note that although cepstral coefficients are used as voice feature quantities, linear prediction coefficients, formant frequencies, Percoll coefficients, logarithmic cross-sectional area ratios, number of zero crossings, etc. may also be used.

"Effects of the Invention" As explained above, according to the present invention, input speech and standard pattern speech are matched using distances consisting of power regression coefficients and spectral distances, so spectral distances alone may cause recognition errors. This method has the advantage that recognition ability can be improved in easy speaker-independent word speech recognition, and there is no need for normalization calculation of the absolute value of power, so there is no time delay in recognition calculation.

[Brief explanation of the drawing]

FIG. 1 is a block diagram functionally showing an embodiment of the word speech recognition device of the present invention, FIG. 2 is a diagram showing the time pattern of the speech logarithmic power of the ``Tungo'' word "Sapporo", and FIG. FIG. 2 is a block diagram functionally showing an embodiment of the present invention. 1: Voice input terminal, 2: Voice section detection circuit, 3: Linear prediction analysis circuit, 4: Kepnutram conversion circuit, 5: Feature parameter register, 6: Power register, 7: Regression coefficient calculation circuit, 8: Switch, 9: Standard pattern storage section, 10:
Time normalization matching circuit, 11: Weight register, 12
: Comparison circuit, 13: Output terminal. Patent applicant: Representative of Nippon Telegraph and Telephone Public Corporation: Takui Kusano 1.Ko 2. Former 5173 area

Claims (1)

[Claims] (1) Means for determining the spectral distance Ds between elements of feature time series of input speech and standard pattern speech, and linear regression coefficient time from the time waveform of speech power for each of the input speech and standard pattern speech. means for deriving a waveform; means for determining a power regression coefficient distance Dp between elements of an input voice and a standard pattern using the linear regression coefficient; and the spectral distance Ds and the power regression coefficient distance Dp.
A word speech recognition device comprising means for calculating the degree of similarity between the input speech and the standard pattern speech by time normalized matching using the inter-element matching distance determined from the above.