JPS62160499A - 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 [Field of Industrial Application] The present invention relates to a speech recognition device that can recognize continuously uttered input speech with high precision.

[Conventional technology]

Automatic speech recognition is extremely important as an interface technology for direct information input from humans to machines. This automatic recognition of speech is understood as the process of converting a speech pattern made by continuously uttering a series of linguistic symbols such as phonemes and syllables into discrete linguistic symbols, and recognizing each of these linguistic symbols individually. be able to.

Research on speech recognition using phonemes as the unit of recognition has been conducted for a long time. For example, the input speech wave is divided into phonological units (
Segmentation) and methods for recognizing human speech by determining (labeling) the content of each phoneme have been proposed. This method can be said to be a method for detecting locations where the speech content changes significantly as phoneme boundaries based on temporal changes in speech power, spectrum, etc. However, the detection accuracy of the phoneme boundaries tends to be affected by the speaking speed and speaking attitude, and no improvement in recognition accuracy can be expected.

On the other hand, a method has been proposed in which phoneme labels are sequentially assigned to input speech at fixed time intervals (frames of speech analysis) and the input speech is recognized from the connected relationships of the phoneme labels. However, a major problem is how to organize the phonetic symbols that are sequentially obtained. For this reason, it has generally been the case that the problem has been dealt with by composing ad hoc rules.

Therefore, the present inventor previously proposed in Japanese Patent Application No. 80-59880 a recognition method capable of dividing and recognizing linguistic symbols of phonemes and syllables with high precision and improving the recognition accuracy.

This recognition method sets restrictions on the linguistic symbols of phonemes and syllables, such as the duration of the phoneme and the conditions for connection between phonemes, and finds the phoneme symbol sequence that has the maximum similarity sum under the restrictions. It is something.

That is, as shown in FIG. 5, the input voice is first guided to the acoustic analysis section l, where it is acoustically analyzed at predetermined analysis time intervals, and, for example, characteristic parameters consisting of spectral data are sequentially output as the output of a filter bank. demand.

Then, the degree of similarity between the feature parameters (spectral data) sequentially obtained at each predetermined analysis time and the phoneme symbols consisting of phonemes and syllables registered in advance in the phoneme dictionary 2 is calculated using, for example, a composite similarity that has a high ability to absorb pattern deformation. The similarity calculation unit 3 calculates each using the method.

Then, the sequence of similarities calculated for each feature parameter is stored in the similarity storage unit 4 in correspondence with its sequence of phonetic symbols, and the sequence of phonetic symbols that has the maximum similarity among them is selected to have the maximum similarity. It is calculated by the degree series calculation section 5. This maximum similarity series is calculated by referring to the connectable information between phoneme symbols stored in the connection information table 6 and the information regarding the duration length of each phoneme stored in the duration length table 7, and taking into account these limitations. This is done by applying dynamic programming to extract only phoneme symbol strings that satisfy the information.

That is, the phonetic symbol p obtained at frame t (l≦p≦P)
Let S (t, p) be the similarity to S (t, p), and let Lmr1xs be the maximum duration of the phonetic symbol p, and La1n be the shortest duration, then t -L max≦t'≦t-L
For every frame t' LIlin, calculate . Here, q is a phonetic symbol that can precede p, and T (t'; q) is the cumulative similarity.

This process is applied to all phonetic symbols p and all frames t.
Then, find the sequence of phonetic symbols for which the sum of similarities B (t'; q) is maximum.

Figure 6 shows the calculation of X from multiple 0 points using this calculation process.
An example of a route for calculating the sum of similarities leading to a point is shown.

This process is repeated over the interval of the input speech, and a sequence of phonetic symbols with the maximum similarity sum is determined. As a result, it becomes possible to recognize input speech with considerably high accuracy.

[Problem that the invention seeks to solve]

However, in the speech recognition method proposed in this patent application No. 80-59880, since phonemes or syllables are considered as phonetic symbols as recognition units, it is difficult to recognize, for example, speech uttered at high speed in conversation. It has been found that in such cases, a problem arises in that an incorrect phoneme symbol may be assigned to a transitional part between phonemes.

For example, when one frame of spectral data of a transitional part between vowels is captured, there may be cases where the spectral data coincidentally resembles another phoneme. Specifically, rhaei is used in response to the utterance of “yes”.
A case may arise in which a phoneme symbol sequence J is required.

The present invention has been made in consideration of the above circumstances, and its purpose is to effectively prevent incorrect allocation of phoneme symbols in transitional parts between phonemes, and to improve continuously uttered input speech. An object of the present invention is to provide a speech recognition device that can recognize speech with high precision.

[Means for solving problems]

The present invention analyzes input speech and calculates the degree of similarity between feature parameters obtained at regular intervals and phonetic symbols, and calculates the degree of similarity between the phonetic symbols and the connectability information between each phonetic symbol and the duration length thereof. When recognizing the input speech by finding candidates for a phoneme symbol sequence that is allowed as a phoneme according to the degree of similarity determined by the phoneme, and finding the phoneme symbol sequence for which the sum of the degrees of similarity is maximum among them, as the phoneme symbol, Phonetic symbol L indicating a phoneme or syllable
This method is characterized in that the above-mentioned processing is executed by preparing a phoneme symbol Lt indicating a transitional connection between phonemes or syllables.

The recognition result is obtained as a sequence consisting only of phonetic symbols Lp, which is obtained by removing the phonetic symbol Lt from the phonetic symbol sequence that gives the maximum similarity sum.

[Action and its effects]

Thus, according to the present invention, each phoneme in continuously uttered speech changes sequentially from frame data having its typical spectral pattern to subsequent phonemes, so that the degree of similarity to that phoneme is directed toward the center. It gradually increases in size, and then decreases as it moves away from the center. Therefore, by finding the phonetic symbol string with the maximum similarity sum as described above from among these similarities,
It becomes possible to find the optimal phoneme segmentation for the input speech.

Furthermore, although some transitional parts between phonemes are far from the typical phonetic spectrum, one phoneme symbol is set for the transitional part between phonemes, and the sum of the maximum similarity among the phoneme symbol series is set. Since the user is asked what to take, for example, it is possible to prevent incorrect phoneme labels from appearing between phonemes. Then, the recognition result for the input speech is obtained by removing the phonetic symbols that express the transitional part between phonemes, so that continuous utterances can be recognized as a series of phonetic symbols without incorrect labeling of the characteristic parameters of the transitional part between phonemes. This provides great practical effects, such as making it possible to recognize input speech with high precision.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of an embodiment device. This device differs from the structure of the device shown in Japanese Patent Application No. 60-59880, which was previously proposed, in that it includes an Lp dictionary 2as that stores phonetic symbols Lp for phonemes or syllables and their standard patterns, and Or L that memorizes the phonetic symbol Lt for the transitional part between syllables and its standard pattern.
t dictionaries 2b, and these Lp dictionaries 2b.
The present invention is comprised of a result determination unit 8 that comprehensively determines the phoneme symbol sequence determined by the maximum similarity sequence calculation unit 5 according to the information registered in the a and Lt dictionaries 2b.

That is, the acoustic analysis unit 1 acoustically analyzes the continuously inputted voice at predetermined analysis time intervals, for example, every 10 m5ec,
The feature parameters are obtained as a frame. Specifically, a filter bank consisting of a plurality of band-pass filters, which is known as a method of spectroscopy, is used to obtain characteristic parameters as a spectrum consisting of the output energy of each channel. This feature parameter is stored in the input pattern memory in the acoustic analysis section 1 and used for calculating the degree of similarity with the phonetic symbols Lp and Lt.

The similarity calculation unit 3 calculates the spectrum data output at each predetermined analysis time interval and the Lp dictionary 2a.

and each phonetic symbol L registered in the Lt dictionary 2b.
The similarity between p and Lt with the standard pattern is calculated by, for example, a composite similarity method.

It goes without saying that when this composite similarity method is employed, several types of mutually orthogonal vectors must be prepared as standard patterns for each phoneme class.

Here, the Lp dictionary "!2a" is the eigenvector obtained by collecting many frames of spectral data at the phoneme center position for each phoneme, and expanding the covariance matrix with KL, for example, when targeting a phoneme as a phoneme. It is constructed by having the eigenvalues as a dictionary.In addition, the phonology is
For example, when using consonant-vowel combination (CV) syllables as a basis, the 07 part can be extracted over several frames and expressed as a one-dimensional vector by combining it with several channels in the frequency direction. The Lp dictionary 2a can be constructed in the same way as in the case of phonemes.

On the other hand, the Lt dictionary 2b is defined as a special phonetic symbol that expresses the transitional part between phonemes or syllables, and for example, the number of dimensions that is the optimal time length to express the characteristics of the transitional part. Represented by a vector. Specifically, for the transition part between phonemes, from the center of the A plate part to the n before and after it.
All you have to do is take out the frame data and create the dictionary in the same way as the Lp dictionary. This Lt dictionary may have a different number of dimensions from the Lp dictionary.

Further, the Lt dictionary does not need to be prepared for all combinations between phonemes, and may be prepared only for combinations between vowels and between voiced sounds, which have an important influence on recognition performance. In addition, since the information is concentrated in the center of the transition area,
For example, as (n-1), items such as ja tj and rial, which have opposite connections, may be prepared as one group.

The similarity calculation unit 3 calculates each phonetic symbol L p registered in the Lp dictionary 42a and the Lt dictionary 2b in this way.
, L j and the spectral data obtained for each predetermined analysis time are calculated.

However, this similarity information is stored in the similarity storage unit 4 in correspondence with the phonetic symbol for each frame. Then, the maximum similarity sequence calculation unit 5 refers to the connection information table 6 and the duration table 7, and extracts the phoneme symbol sequence that has the maximum similarity sum among the phoneme symbol sequences that are established as phonemes.

For example, as shown in FIG. 2, this table 8.7 contains information on the duration of each phonetic symbol Lp, t, 1, and
It stores information on phoneme symbols that can precede the phoneme. Specifically, the duration length of the phonetic symbol (h) is 32 to 1 Bofllsec, and the phonetic symbol (an l+ un e* Or Nrai
, au, ...).

Also, the duration of phonetic symbol (a) is 32 to 400 m5
ec, and the phonetic symbol (1, u, e, O, N+ pr
” indicates that it can be connected to ..., ia, ua, ea, ...).

Similarly, the phonological symbol (at) for the transitional part between phonemes
, the duration length is 32 to 100fflse
e, which indicates that the phonetic symbol that can precede is (a).

However, the maximum similarity series calculation unit 5 calculates the
Similar to the method proposed in No. 9880, for phonetic symbol strings that satisfy the duration conditions for each phonetic symbol and also satisfy the connection conditions between phonetic symbols, find the phonetic symbol string that has the maximum similarity sum among them. ing.

Here, maximizing the similarity can be expressed by an additive relationship, and is carried out by employing, for example, dynamic programming.

That is, the phonetic symbol p (1≦
Let S (t:I)) be the similarity for p≦P). Then, the calculation unit 5 has a cumulative similarity storage area T (t),
A storage area L(t) for the label name and a pointer storage area F(t) used to obtain the optimal result are prepared. Note that each of these storage areas is zero (0).
is initialized to .

Now, considering the t-th frame, the maximum time length LIIlaXs and the minimum time length Lm1n are given to each phonetic symbol p, and the following processing is executed. That is, t-Lmax
≦t′≦t−Lm1n 15p≦P For the frame t′ and the phonetic symbol p, the following is calculated. This process is executed for all frames i/ and each phonetic symbol p to obtain B(t';p), and the largest one among them is selected.

And the maximum value of B(t':p) B(t';p)IIa
x and give it t'ma't sp anax for each storage area T (t).

Stored in L(t) and F(t), respectively.

By repeating this process over the human voice section,
At the end of the input speech (frame tend), the maximum similarity sum T (t end; I) for each phonetic symbol p
will be required.

The result determination unit 8 calculates the results obtained by the maximum similarity sequence calculation unit 5 in this way, that is, the storage areas T (t), L
According to the results obtained for (t) and F (t), respectively,
We are looking for a correspondence between the phonetic symbols. Furthermore, since the phonetic symbol string obtained here includes the phonetic symbol Lt that indicates the transitional part described above, it is necessary to remove that phonetic symbol Lt as the final recognition result. .

FIG. 3 is a diagram showing the flow of result determination processing for a phoneme symbol sequence that takes the maximum similarity sum, including the removal of such phoneme symbols Lt.

That is, in this result determination process, first, the maximum similarity sum T (
The phonetic symbol p that maximizes t end; p) is the phonetic symbol L
Search p from p and set it as p a+ax (step a).

Next, initialize the processing variables TO as ten end and P as p max in step b), execute the process FROM ← F (To:P)+1 in step C), and set the phonetic symbol, its starting position, and ending position. (Step d).

Subsequently, the following process is performed: P - L (TO; P) To 4 - FROM - 1 FROM ← F (To; p) + 1 to obtain the next symbol (step e). This process is repeated until the TO becomes (0) in step f), and data (P
, FROM.

TO) are determined in reverse order.

Thereafter, for all the phonetic symbols p found, it is determined whether the phonetic symbol p is the phonetic symbol Lt defined for the transitional part between phonemes (step h), and if it is the phonetic symbol Lt, then , the frame with the maximum similarity s (tap) is detected as the center position of the transition part (step i). Then, the phoneme symbol Lt is divided into two at the center position of the transition part, the information is integrated into the phoneme symbols Lp before and after it, and the phoneme symbol Lt is removed (step j). This process is executed for all phoneme symbols p in step g), and a sequence of phoneme symbols Lp with the maximum similarity sum is determined.

In addition, if it is not necessary to accurately determine the phonological boundaries,
It is sufficient to simply delete the phonetic symbol Lt.

According to such processing, for example, FIG. 4(a) shows the speech power of the input speech "high", and the series of similarities of phonetic symbol strings for this and the phonetic symbol string that takes the sum of the maximum similarities are calculated. As shown in (b) and (c), the duration of the phoneme, based on the similarity to the phoneme symbol calculated at each predetermined time,
It becomes possible to obtain the optimal phoneme segmentation by introducing knowledge of the information and its connection relationships.

However, Fig. 4(b) shows an example of using only the phonetic symbol Lp proposed in Japanese Patent Application No. 60-59880.
c) shows a phoneme symbol string when phoneme symbols Lp and Lt according to the present invention are used.

As shown in comparison with FIG. 4, in the present invention, a phoneme symbol Lt is defined for a transitional part between phonemes and syllables, and a phoneme symbol string of input speech is created taking into consideration this transitional part. This prevents the connected parts of phonemes from being mistakenly recognized as other phonetic symbols. Therefore, unwanted phoneme symbols are not generated in transitional parts of phonemes or syllables, and input speech can be recognized with high precision.

Note that although the explanation has been given using phonemes as examples of phonetic symbols, the device can be constructed in the same way even with syllables, and the same effects can be achieved. Furthermore, the process of finding a sequence of phonetic symbols that yields the maximum similarity sum can also be implemented in a modified manner. In addition, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of drawings]

Fig. 1 is a schematic configuration diagram of a PIW location in which the present invention is implemented;
The figure shows an example of the structure of a table that stores information on phoneme duration and possible preceding phonemes, Figure 3 shows the flow of result determination processing, and Figure 4 shows the detection of phoneme symbol sequences for input speech. FIG. 5 is a diagram showing the results, and FIG. 5 is a schematic block diagram of the apparatus proposed earlier by the present inventor. FIG. 6 is a diagram showing the processing concept when calculating the maximum similarity sum.゛ 1... Acoustic analysis department, 2... Phonological dictionary, 2a.
...Lp dictionary, 2b...Lt dictionary, 3...Similarity calculation unit, 4...Similarity storage unit, 5...Maximum similarity series calculation unit, G...Connection information table, 7. ...Duration length table, death...Result judgment section. Applicant's agent Patent attorney Takehiko Suzue 2nd prisoner Figure 3? /-〆“ee h a e 1h
a 1 Fourth cause

Claims (3)

[Claims] (1) Means for analyzing input speech to obtain feature parameters of the input speech at regular intervals; A means for calculating the similarity of each feature parameter, and a phoneme symbol string that is allowed as a phoneme according to the connectability information between the phoneme symbols, information regarding the duration length of each phoneme, and the similarity of each feature parameter to the phoneme symbol. A speech recognition device comprising: means for obtaining candidates; and means for obtaining a phoneme symbol sequence having the maximum sum of similarities among these phoneme symbol sequences. (2) The recognition result for the input speech is obtained as a phonetic symbol sequence obtained by removing phonetic symbols at transitional connections between phonemes or syllables from the phonetic symbol sequence that gives the maximum similarity sum. The speech recognition device according to item 1. (3) The speech recognition device according to claim 1, wherein the maximum similarity sum of phoneme symbol sequences is calculated using dynamic programming.