JP2577891B2 - Word voice preliminary selection device - Google Patents

Description

The present invention is preferably applied to word speech recognition in which a large vocabulary is to be recognized. In order to reduce the number of words to be subjected to time-normalized matching with a large amount of information. The present invention relates to a word voice preliminary selection device to be used.

[Prior Art] In a word speech recognition method, all of the words to be recognized are stored in advance as standard feature parameters, feature parameters of an unknown input speech word are detected, and the feature parameters and the standard of each of the previously stored words are used. Recognition of an unknown input speech word is performed by performing time-normalized matching with a feature parameter. However, when the number of words to be recognized is large, that is, when a large vocabulary is to be recognized, the amount of calculation required for time-normalized matching is significantly increased.

Therefore, it has been proposed to first easily recognize an unknown input speech word, select some words, and perform time-normalized matching between the standard feature parameter of each selected word and the unknown input speech word. Selecting the few words is called word pre-selection. By relatively easily performing the word preliminary selection, speech word recognition can be efficiently performed.

The first method of the conventional word preliminary selection method is to divide the features of the word voice into equal intervals on the time axis, and use a value obtained by averaging each feature parameter at each sampling point in each section. Make a selection. In the second method, the change in the power level of the word voice is represented by a binary pattern, or the overall characteristics of the phoneme sequence, for example, the presence of / a / in a word, is represented by a binary pattern. This is a method of performing preliminary selection by simple comparison of series. Further, the third method clusters changes in the spectrum of the word speech for each short time, creates a set of representative spectra constituting each word as a template set, and uses this template to transform the spectrum of the input speech into a vector. This is a method of performing quantization and performing preliminary selection based on the distortion amount at this time.

In the first pre-selection method, the temporal cutoff of the uttered word voice and the extension of the word during the utterance do not match the previously stored word, and the position of the break is not the same. If there is a time lag,
Pre-selection ability decreases. Regarding the second preselection method, the power level pattern is often similar between words, and it is difficult to extract a phoneme sequence with sufficient accuracy.
There is a disadvantage that the preliminary selection ability is low. The third method is
Since time information such as the temporal flow and duration of the spectrum for each short time is not taken into account, a relatively long section of a certain word is associated with an instantaneous section of a different word, which often causes selection errors. There was a problem.

If the preselection is bad, many words will be selected and the benefit of the preselection will not be obtained. Alternatively, the correct word is dropped from the preliminary selection, resulting in erroneous or unrecognizable recognition.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned drawbacks of the conventional preselection method, by extracting feature parameters that can be extracted with sufficient accuracy, include time information, and are not easily affected by shifts in speech section cutout or temporal expansion and contraction. Using,
It is an object of the present invention to provide a word-speech preselection device which improves the performance of preselection and thereby makes it possible to reduce the number of subsequent matching operations.

"Means for Solving the Problem" The present invention provides means for calculating a linear regression coefficient indicating a relatively local temporal change derived from a time waveform of a spectral parameter every short time. The combination with the spectral parameter is clustered by the clustering means for each vocabulary to be recognized, and this is stored as a template for each vocabulary to be recognized. Then, the distortion amount at that time is calculated, and a plurality of words having relatively small values are selected.

According to the present invention, a conventional preselection apparatus uses a template obtained by combining clusters of spectral parameters and their linear regression coefficients at the same point in time as if they were a single type of parameter. In that vector quantization is performed.

"Embodiment" FIG. 1 shows an embodiment of a word voice preliminary selection device according to the present invention. The word voice input from the word voice input terminal 1 is subjected to spectrum and (logarithmic) power analysis via the voice section detection circuit 2 and the spectrum analysis unit 3, and the analysis result is stored in the spectrum parameter storage unit 4. .
As a parameter for expressing a spectrum, a cepstrum, an LSP (line spectrum pair) parameter, or the like that can express a spectrum with a relatively small number of parameters is used. Hereinafter, a case where a cepstrum is used will be described. The same applies to the case where LSP parameters and the like are used.

The cepstral and power time waveforms stored in the spectrum parameter storage unit 4 are input to a regression coefficient calculation circuit 5 to extract linear regression coefficients. The absolute value of the power is liable to fluctuate depending on the utterance level. This is excluded, and the cepstrum and the time waveform of the linear regression coefficient of the cepstrum and the power (these are collectively referred to as a feature parameter waveform) are temporarily stored in the feature parameter register 6. The learning mode and the recognition mode are switched by the switch 7, and in the case of the learning mode, the feature parameter waveform is input to the clustering unit 8 to perform clustering, and a plurality of templates created for each vocabulary are stored in the template accumulation unit 9. accumulate. In the case of the recognition mode, the feature parameter waveform and a plurality of templates of each vocabulary are input to the vector quantization circuit 10, and the amount of distortion when performing vector quantization using the plurality of templates of each vocabulary is calculated. The result is stored in the quantization distortion storage unit 11. The distortion amounts for all the vocabularies are input to the candidate word selection circuit 12, and a predetermined number of vocabularies having a relatively small distortion amount or smaller than the threshold value stored in the threshold value storage unit 13 Data indicating a vocabulary having a distortion amount is provided to the output terminal 14 as data indicating a candidate word.

The operation will be described in more detail. First, a speech wave used for word recognition is input from the word speech input terminal 1. Since the input voice wave usually includes an actual voice section and a silent (noise) section, the input voice wave is input to the voice section detection circuit 2 to detect the voice section. Do. For this detection, it is possible to use some well-known methods, for example, short-time power of an input signal wave, time during which power exceeding a certain value continues, and the like. The signal wave of the detected voice section is sent to the spectrum analyzer 3 and converted into a cepstral and power time waveform. Since this technology is already known (for example, see Literature, Furui: Digital Speech Processing, Tokai University Press, pp.44-48, 1985),
Although details are omitted, basically, after passing through a low-pass filter, sampling and quantization are performed, a short section waveform is cut out at regular time intervals, multiplied by a Hamming window, etc., and two Fourier transforms are performed. Cepstrum is extracted by logarithmic transformation or linear prediction analysis and iterative operation. As the length of the Hamming window, for example, a value such as 10 ms is used as a cycle for updating it, for example, 30 ms. The cepstrum calculates values up to a predetermined p-th order, for example, the first to tenth orders.

In the extracted cepstrum and power time waveforms, a section having a fixed time length is temporarily stored in the spectrum parameter storage unit 4 at regular intervals, and the contents of the storage unit 4 are sent to the regression coefficient calculation circuit 5 and are linearly converted. A regression coefficient is calculated.
The length of the time waveform input to the spectrum parameter storage unit 4 and the regression coefficient calculation circuit 5 is, for example, 50 m
s, as a cycle for updating this, a value such as 10 ms is used, for example. When the time waveform is represented by x _j (j = −M,..., M), the linear regression coefficient a can be obtained by the following calculation.

The linear regression coefficient is calculated according to the input of the regression coefficient calculation circuit 5 updated every 10 ms for each order cepstrum and power, and the linear regression coefficient is combined with the cepstrum as a (2p + 1) -dimensional feature parameter. It is sent to the characteristic parameter register 6 and stored.

The switch 7 is a switch for selecting a learning mode and a recognition mode.
Is connected to the terminal 7a, and a characteristic parameter waveform is obtained from voices of a person who inputs voice to be recognized later or a plurality of voices different from the person.
Clustering is performed on the feature parameters for each short time.
This clustering is to group a large number of parameter sets into a predetermined fixed number of representative sets.
For example, assuming that (2p + 1) -dimensional feature parameters (cepstrum and regression coefficients of cepstrum and power) are extracted every 10 ms from a certain word voice uttered by four speakers, the lengths of the word voices are averaged. If it is 500 ms, a total of 50 × 4 = 200 kinds of (2p + 1) -dimensional feature parameters are given. For example, 32 types (2p
+1) In order to summarize into the dimensional feature parameters, a known method (literature, Y. Linde, A. Buzo, and RMGray: An algorithm f
or vector quanti-zation, IEEE Trans.Commun., vol.CO
M-28, pp. 84-95, 1980) can be used. In this method, similar characteristic parameters are collectively represented by one average value, and all the original 200 characteristic parameters are replaced by the closest one of the 32 representative values. The representative value is determined so that the error is minimized as a whole. In this way, these 32 types of (2p + 1) -dimensional feature parameter representative values determined for each recognition target vocabulary are stored in the template storage unit 9 as templates.

Then switch to terminal 7b for voice to be recognized
To the contents of the feature parameter register 6,
That is, the feature parameters of the input voice for each short time are input to the vector quantization circuit 10. In the vector quantization circuit, templates of each vocabulary to be recognized are sequentially read from the template storage unit 9 to perform vector quantization. This process
For each feature parameter of the input speech for each short time, this is performed by selecting the closest template, and a value obtained by averaging the error between the template and the feature parameter at that time for the entire input speech, that is, quantization distortion is calculated.

The special parameter of the l-th template of a certain word k is r _kli ( _{consisting of} 1 ≦ i ≦ 2p + 1, p-order cepstrum, p-order cepstrum linear regression coefficient, and power linear regression coefficient), and input speech When the feature parameter at the time point m is represented by x _mi (1 ≦ i ≦ 2p + 1), the following value is used as the distance between them (a numerical value indicating that the smaller the distance, the greater the similarity).

Here, w _i is a numerical value indicating a predetermined weight for each feature parameter, and this value is set to an appropriate value based on the result of the preliminary experiment so as to obtain relatively high accuracy. The distance d, which is stored in the equation 15, is calculated by the difference between the input speech and the template for the p-order cepstrum, the p-order cepstrum linear regression coefficient, and the power regression coefficient at the same time, as shown in equation (2). the indicators are calculated as the square sum, that is used together properties different from each other between the spectral parameters and the linear regression coefficients, which performs weighting w _i to take these equilibrium, therefore w _i As the value of, at least three values corresponding to the cepstrum, the linear regression coefficient of the cepstrum, and the linear regression coefficient of the power are used.

In this way, the quantization distortion is obtained between the input speech and the templates corresponding to all the vocabularies to be recognized, and stored in the quantization distortion storage unit 11.Then, these are input to the candidate word selection circuit 12. The recognition target word name whose quantization distortion is smaller than the threshold value stored in the threshold value storage unit 13 is output from the terminal 14 as a candidate word of the input voice. Alternatively, compare the quantization distortion for all the recognition target words, and output a fixed number of word names with relatively small quantization distortion (for example, 1/10 of the whole, 100 words if 1000 words are targeted) as candidate words I do. If the threshold value is reduced, the number of candidate words is reduced, and the subsequent time-normalized matching processing can be significantly reduced.However, if the threshold value is too small, the correct word is cut off by preliminary selection, and in the subsequent processing, It will not be recoverable. The setting of the threshold value should be set according to the method of use, and is determined in relation to the recognition rate and the processing amount. The same applies to the number of candidates to be output when a certain number of candidate words are output.

In the related art, for example, clustering is performed only on the cepstrum, and vector quantization of the input speech is performed to calculate the quantization distortion.In this embodiment, the linear regression coefficients of the cepstrum and the power are the same as those of the cepstrum. Clustering and vector quantization are performed collectively at the time. This linear regression is a linear approximation of the cepstrum and power time waveforms, and the slope of this approximation is the linear regression coefficient, that is, the degree of similarity between the input speech and the template is determined for the tendency of the change in cepstrum and power. . As a result, the stationary part of the spectrum of the input voice does not correspond to the template of the excessive part of the different word, and the transient part of the spectrum of the input voice does not correspond to the template of the stationary part of the different word. A word voice preliminary selection system having high accuracy for voice can be realized.

According to previous experiments, in recognition of 100 words of city names uttered by unspecified speakers, a template for each word was created using the voices of four men, and the threshold value for quantization distortion was set. If properly set, candidate words can be narrowed down to an average of 4.5 words for 20 male voices different from the above four speakers, and the correct word is included in the candidate words at 99.9% It has been confirmed that
According to the conventional method using only cepstrum, an average of 4
It is understood that the present invention is superior to the fact that it was not possible to narrow down to 2.5 words, and that the ratio of correct words contained therein was only 99.0% accurate.

[Effects of the Invention] As described above, according to the preliminary selection device of the present invention, a representative pattern is selected by including its local time change characteristic in the spectral parameter, and vector quantization is performed by this. It is less susceptible to the effects of target segmentation and slow expansion and contraction in words, and moreover, a proper association is made between the representative pattern and the word voice, taking into account the temporal flow. There is the advantage that a preliminary selection can be made.

[Brief description of the drawings]

FIG. 1 is a block diagram of a word-speech preliminary selection device showing an embodiment of the present invention. 1: Word voice input terminal, 2: Voice section detection circuit, 3: Spectrum analysis unit, 4: Spectrum parameter storage unit, 5: Linear regression coefficient calculation circuit, 6: Feature parameter register, 7: Switch, 8: Clustering unit , 9: template storage, 1: vector quantization circuit, 11: quantization distortion storage, 12: candidate word selection circuit, 13: threshold storage, 14: output terminal, 15: weight register.

Claims (1)

(57) [Claims] An apparatus for recognizing an unknown input speech word by performing time-normalized matching between a feature parameter of an unknown input speech word and a standard feature parameter of each vocabulary to be recognized should perform the time-normalized matching. A device for pre-selecting standard feature parameters, means for calculating and storing parameters indicating a temporal change of a frequency spectrum and power of a sound wave, and a linear regression coefficient for each short time from a time waveform of those parameters. Means for calculating, means for clustering feature parameters composed of a combination of the above parameters and linear regression coefficients for each vocabulary to be recognized, means for accumulating a plurality of templates as a result of this clustering for each vocabulary to be recognized, The time waveform of the parameter and the linear regression coefficient of the speech wave to be recognized is used as the template Means for performing vector quantization using the method; means for calculating the amount of distortion for each vocabulary to be recognized at the time of quantization; and time-normalized matching of a standard feature pattern of a plurality of vocabulary to be recognized having a relatively small amount of distortion. A word voice preliminary selection device having means for selecting as a standard feature pattern to be performed.