JPS6051898A - Continuous 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 continuous speech recognition device that can efficiently recognize continuously uttered input speech.

[Technical background of the invention and its problems] For Japanese word processors and voice typewriters that use voice as an information input means, it is important to efficiently recognize voices that are continuously uttered in a natural manner. It becomes a challenge. However, one of the conventional continuous speech recognition methods uses a recognition unit as a phoneme, and converts the time series of the feature parameters of the input Mμ into a sequence of phoneme labels or a so-called segment lattice. There is a method that extracts simple B"74 and sentences. However, in the case of input speech uttered in series '1jljj, even if the phoneme is the same, a so-called key g combination occurs depending on the phoneme environment before and after it, and as a result, There is a property that acoustic expression undergoes various transformations.For this reason,
It is difficult to convert to the above phoneme labels with high accuracy,
It lacked practicality.

In contrast, a method has been proposed in which the recognition unit is word-sized, the words are directly identified from the time series of feature parameters, and then the word string is recognized as a sentence. This method avoids the above-mentioned problem of articulatory combination by having a hypomorphic pattern as a word. However, the above word identification methods detect word boundary positions from the input speech and identify words for a partial interval of the input voice determined by the boundaries, and conversely, they detect word boundaries from the input speech and identify words for a subinterval of the input voice that is determined by the boundaries, and conversely, they detect all of the input speech without detecting the boundaries. If a word exists in a subinterval, it is roughly classified into those that identify the word profanely. The detection of the boundary is performed, for example, by extracting characteristic parameters such as the voice power and spectrum change of the input voice, and finding the extreme values in the time series. However, for example, the number 2" (/n
i/) and the number 1'' (/1tfi/) are uttered consecutively to become (/ni 2tfi/), there were problems such as the inability to detect the word boundary.

In this regard, the latter word identification method described above has been put into practical use in some cases. In other words, the basic old algorithm for word identification uses feature parameters that are analyzed at regular intervals for standard patterns for each word in a word (defined as a speech recognition unit, not a linguistic meaning). Prepare as a time series. Then, the distance from the standard pattern is determined for all partial sections of the input speech, and the word that provides the minimum distance is determined.

At this time, the distance between feature parameters (C, inter-frame distance) obtained at each predetermined analysis time is calculated and used for dynamic planning full-time normalization to determine the distance between time-series patterns. Then, the distance from the input speech as a word string is evaluated for all combinations of partial sections, the minimum cumulative distance is waited, and a word string corresponding to the entire input speech is obtained as a recognition result.

However, although this method works well when the speaker is specified, it causes the following problems when the speaker is unspecified. That is, when targeting unspecified speakers, the speech patterns of words vary greatly depending on the speaker, so it is necessary to prepare a huge amount of standard word patterns corresponding to each speaker. Therefore, for unspecified speakers, the same principle requires a limited number of quasi-patterns, which is extremely difficult to realize.

Recently, it has been considered to prepare only a finite number of standard patterns for each word and apply a clustering method to solve the problem of standard patterns for unspecified speakers. However, in this case, the recognition rate for word strings (branches) decreases significantly, which is unbearable in practical terms. Moreover, if all of this method is adopted, it will be necessary to calculate all the distances for all word categories and also for each of multiple time-series standard patterns, which has the disadvantage that the total amount of calculation processing will be extremely large. There was a drawback. For these reasons, it has been extremely difficult to efficiently and effectively recognize continuously uttered input speech.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to efficiently recognize input speech continuously uttered by an unspecified speaker with high precision and through real-time processing. Our goal is to provide a highly practical continuous speech recognition device.

[Summary of the invention]

That is, the present invention has a standard pattern as a feature vector sequence having a time series of feature vectors of a certain dimension in the frequency axis direction for a specific time point, and features that can be determined by analyzing input audio at fixed time intervals. The degree of similarity between the parameter vector and the feature vector of each of the standard patterns at each time is determined, and the degree of similarity of the standard pattern to a subsection of the input speech is determined from these degrees of similarity between the frequency and frequency range, and candidates for that subsection are determined. After determining the standard Noreen (candidate word) and its degree of similarity, calculate the candidate word string to be used for the entire subinterval string by calculating the sum of each unit similarity of the subinterval string that forms an interval equal to the input speech interval. It was designed to be evaluated.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings. Note that although the recognition unit of input speech is described here as a word, this t)1 word is not defined in a linguistic sense, but as a unit of speech handling in speech recognition processing. Saitako's simple and yori can be in syllables, phrases, or something similar.

Now, FIG. 1 is a schematic configuration diagram of an embodiment apparatus, and FIG. 2 is a diagram showing the main processing procedure of the same apparatus. Input speech that is continuously uttered by an unspecified speaker is input to the acoustic analysis section 1 and analyzed at predetermined analysis time intervals to determine its characteristics/speech.
-, converted to 6 rammeters. The acoustic analysis section 1 is constituted by a filter bank consisting of a plurality of band-pass filters that divides the audio band into about 16 to 30 bands and performs spectrum analysis. As a result, the features of the input voice, the f-lame-taka, and the growling feature vectors are determined for each pair of steps.

Therefore, the feature vector of this input voice is input to the frequency human similarity calculation unit 2, and the standard pattern (feature vector of each time point of the standard pattern registered in advance in the turn storage unit (memory) 3) The Frey-human similarity of j
The J1 reciprocal value is retained. By inputting this similarity value, the unit similarity calculation/judgment unit 4 calculates the total similarity for each word of 3゛1 for the partial interval in which the word is present in the input speech.
-I'm looking forward to it.

FIG. 2 schematically shows the processing steps II performed by each of these parts, and in this apparatus, similarity calculation is performed using, for example, the composite force 11-r method in turn recognition.

Here, however, a single voice pattern is expressed as an M-dimensional feature vector consisting of features/parameters at M points along the frequency axis, arranged at N points along the time axis. The N points in the time axis direction are determined by linearly dividing the duration of the word seito into N points, and the M points in the frequency axis direction are determined by dividing the duration of the word seito into M points in the filter bank. is determined by corresponding to each output of a bandpass filter. +'jf used in this composite similarity method
Multiple words (recognition unit) registered in advance in memory 3
The 6 standard patterns are, for example, those that are pre-processed from an unspecified number of utterances or 1 μ words. That is, each word's category l (1"1 #2 a...,
I) and its time point n (n-1, 2, -, N), calculate all the correlation matrices corresponding to the distribution on the M-dimensional space (in the frequency axis direction) independently, and calculate their eigenvectors with large eigenvalues. Arranged in order from top to bottom, it can be expressed as rrl ・rr2 1rij ・=−rIJ. As a result, the four standard turns of each word are expressed as mutually orthogonal feature vectors.

For such a standard pattern, the word similarity for word category i is calculated as follows. That is, when the feature vector Let J'=1 be the Frei human similarity 8i with each feature vector of the turn. Here, (x-rij") is a vector! krij". This amount of calculation is, for example, when the number of filter punctures M is 10, the number of sample points N in the time axis direction is 16, and the standard If the number J of eigenvectors of the pattern is 5, and the number of word categories is given as 10 using numbers as an example, then (MXNXJXI) = 8000 multiplications and additions are performed within a certain time of acoustic analysis. At this time, the analysis time interval of the audio signal only needs to be about 16m9cc, so each of the 8000 multiplication and addition processes described above only needs to be performed within 2μ (8), which is sufficient to perform real-time processing. ,
In the above example, the amount of inter-frame similarity to be stored at this time only needs to be stored for each time point (NXI=160 (words) t), and the amount of storage does not become a large burden.

In this way, all inter-frame similarities between the feature vector of the input speech and the feature vector at each time point of the standard pattern are determined.

The unit similarity calculation unit 4 calculates the inter-frame similarity for all sub-intervals in which there is a possibility that words formed up to this point in the input speech may exist from the inter-frame similarity Si determined in this way. 8i on the time axis, and from the resampled inter-frame similarity Sin, the similarity for the recognition unit (word, i.e., category i) in that subinterval is determined as Si″FC1n = 1. Then, each subinterval The word category name i?f that takes the maximum similarity value for i?f is taken as the recognition result for that subinterval, and is stored together with its similarity value and the position of the subinterval.This calculation involves addition for the number of sample points N. It only needs to be done a few times, and the amount of calculation is small.

Thereafter, the unit evaluation determining section 5 selects a sequence of partial sections having the same start and end ends as the voice input section from among all combinations of partial sections. Then, for the column of the subinterval, calculate the sum of the word similarities St determined for each subinterval, compare the values of the sums determined for each column, and determine the size of the word similarity St. A word sequence that makes up an interval sequence? , is being evaluated.

For example, the one for which the sum of the similarities of the subinterval sequences is the maximum is
It is evaluated that matching is achieved over the entire interval of continuously uttered input speech, and the entire sequence of word category i found in each subinterval constituting the subinterval sequence is output as a recognition result.

The above is the operation of continuous speech recognition processing by this device.

This will be explained in more detail with reference to FIGS. 3 to 6 as follows. That is, assuming that the time series of feature vectors analyzed at fixed time intervals of input speech is shown in the middle of Figure 3, the features at each time point of the standard pattern for each input speech feature vector at each sample time point are The interframe similarity with the vector is 81, and is expressed as B2 to BL. That is, for each sampling time, all combinations of I and N are evaluated for inter-frame similarity with respect to the characteristic parameter X of the input voice, and are stored and held as a table, for example. This partial similarity calculation is performed sequentially at fixed analysis time intervals as the audio input progresses over time.

Accordingly, the unit similarity calculation/judgment unit 4 has determined, as shown in FIG. 4, candidate sections in which words can exist in the input speech as partial sections from the start of speech input to the present time. In other words, the length of a subinterval in which a word can exist is mostly determined by the range, and for example, compared to the above analysis unit time, the shortest is 3 unit hours, and the longest is 11 unit hours. It is defined as.

Based on these voice input conditions, for example, based on the entire current time, a partial interval of 3 unit time, a partial interval of 4 unit time...
-Assume 11 unit subintervals, etc. and,
For each of these partial sections, the degree of similarity with respect to each of the four standard turns of the partial section is calculated from the partial similarity determined at the sample time point corresponding to that partial section. When performing this similarity calculation, in order to absorb the difference in the time length of the input speech word due to the difference in the length of each subinterval as described above, this is resampled and the time length of the word to be processed is It is necessary to absorb fluctuations. Therefore, for example, the fifth
As shown in the figure, the number may be determined to be the same for partial segments having different lengths based on the current point in time. Then, the position of the subscript in) used to calculate the inter-frame similarity 84 based on the sample points? Based on the inter-frame similarity determined and selected in this way, the similarity 8 for the partial interval is determined as shown in C in FIG.
All you have to do is put it in 2. As a result, the similarity to a plurality of standard patterns is determined for each subinterval, so the maximum similarity value among them is obtained, the similarity value exceeds a predetermined darkness value, and the second highest similarity value is obtained. If the word category is 91, although the difference with the similarity value is sufficiently wide, it is recognized as a candidate word for that partial interval.

In this way, the candidate words for each subsection and the degree of similarity obtained from the candidate words are arranged for each position of the subsection as shown in FIG. 6. Therefore, the unit string evaluation determining unit 5 selects a string of subintervals that forms an interval equal to the voice interval, and for example, in this example, (L, J, El)
, (L, G, C), (K; H, C), (I, B) are selected, and the sum of similarities of each subinterval sequence is calculated. Based on the value of this sum, it is evaluated whether the partial interval sequence matches well with all the intervals of the input speech. Regarding the evaluation of this subinterval sequence,
■C■ It is also possible to use a dynamic programming method known as continuous word speech recognition using syllables as a unit, or a parallel search method using a task domain. Further, at this time, intermediate results of word similarity obtained up to the point in time may be sequentially used. In this way, the recognition result can be obtained in real time as soon as the voice input ends.

Therefore, according to the present invention, the speech pattern of a word, which is a unit of recognition, is expressed by arranging feature vectors of a certain dimension in the frequency direction in time series, and feature parameters corresponding to the frequency axis of the input speech can be obtained. Since the similarity is calculated for each time point of each turn (standard of the word) and each turn, continuous speech can be processed in real time.

Moreover, since the standard pattern of words is statistically processed using characteristic vectors of a certain dimension in the frequency axis direction and arranged in time series for recognition processing, it is possible to absorb pattern fluctuations due to differences in the utterances of unspecified speakers. Simple and effective implementation 1
1 function, and its practical advantages are enormous.

Note that the present invention is not limited to the above embodiments. For example, word similarity calculation, Mahalanobis distance calculation,
This can also be done using a statistical discriminant function. In this case, the distance values and IJ'J values may be subjected to mapping processing to be used as the total similarity. It goes without saying that the recognition unit may be the back or the image iFi, and these may be combined. In short, the present invention can be modified and implemented without departing from the gist thereof.

[Effects of the Invention] As explained above, according to the present invention, a standard pattern is expressed as a feature vector sequence having constant dimension feature vectors of feature parameters in the frequency axis direction in time series.
It becomes possible to sufficiently deal with various variations in speech patterns caused by unspecified speakers and to recognize speech with high accuracy. Furthermore, by calculating the similarity for a subinterval as a function of the similarity with the feature vector at each time point and determining candidate words for the subinterval based on this, the similarity calculation can be decomposed into parts for each analysis time. Therefore, real-time processing becomes possible. Therefore, it becomes possible to perform highly accurate recognition processing in real time, and a great practical effect can be obtained.

[Brief explanation of the drawing]

The figure shows an embodiment of the present invention.
The diagrams showing @ and FIGS. 3 to 6 are diagrams each showing the processing concept in the recognition processing process. J...Acoustic analysis unit, 2...Inter-frame similarity calculation unit, 3...Standard / turn storage unit, 4...Unit similarity calculation judgment unit, 5...Unit sequence evaluation judgment unit . Applicant's representative Patent attorney Takehiko Suzue Is Figure 4 the sound of Figure 6? V interval A, Shuima section B") (WL, WJ, WB) D. 口 G. 1, + Procedural amendment Showa'5ε5ite 1・"Jchoζ'x L+ Commissioner of the Japan Patent Office Kazuo Wakasugi 1, Indication of the case Japanese Patent Application No. 159316/1989 2, Name of the invention Continuous sound P recognition device 3, Person making the amendment Relationship to the case Patent applicant (307) Tokyo Shibaura Electric Co., Ltd. 4, Agent

Claims (5)

[Claims] (1) A memory in which each of a plurality of recognition unit standard gloss turns is stored as a time series consisting of a predetermined number of time points in a feature vector of a constant dimension; Means for determining the degree of similarity between the feature vector of the input voice and the vector at each time point of the standard/four-turn feature vector sequence, and calculating the degree of similarity of each standard pattern to the partial interval of the input voice from these degrees of similarity. A standard pattern salt that calculates the maximum value, a means for determining its unit similarity, and a means for determining the unit similarity for each subinterval of the sequence of subintervals that forms an interval equal to the input speech interval. 1. A continuous speech recognition device comprising: means for evaluating a standard pattern name sequence constituting a sequence of lever subintervals. (2) The continuous speech recognition device according to claim 1, wherein the recognition unit is defined as a syllable, a word, or a phrase. (3) The continuous speech recognition device according to claim 1, wherein the partial section of the input speech is defined as a section of recognition units that can be taken up to the present point in the analysis processing process. (4) The continuous speech recognition device according to claim 1, wherein the degree of similarity is determined based on frame statistics. (5) The continuous speech recognition device according to claim 1, wherein the series of feature vectors representing the standard four-turn recognition unit is comprised of a number smaller than the number of acoustic analysis sections of the recognition unit.