JPH07104783A - Voice recognition device and method for generating standard pattern of recognition basic unit - Google Patents

Abstract

PURPOSE:To make articulation coupling to be adaptable by a speaker to deal with the deformation of a voice pattern by generating a standard pattern while putting similar speakers together with a cluster ring, and generating a standard pattern of a large unit while utilizing the correspondence relation between a phoneme and the standard pattern. CONSTITUTION:An input voice is converted by a voice input means 1 into an electric signal, which is analyzed by a voice analyzing means 2, so that a time series of feature vectors is outputted. A standard pattern connecting means 33, on the other hand, connects standard patterns of recognition basic units stored previously in a standard pattern storage 34 according to phoneme-recognition basic unit corresponding relation information stored in a phoneme-recognition basic unit correspondence relation storage means 34 and word information 35 to constitute a word standard pattern. A collation part 31 collates the standard patterns of respective words generated by the standard pattern connecting means 33 with the feature vector time series of the input voice to find the similarity between the input voice and respective words to be recognized. A decision part 32 outputs the recognition result on the basis of the similarity.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus that uses a speech section whose speech section has a length shorter than a phoneme as a standard pattern, and a method for creating the standard pattern.

[0002]

2. Description of the Related Art As a unit of a standard pattern in a speech recognition apparatus, several units such as a word unit, a syllable unit, a phonological unit (a vowel / consonant subdivided unit) can be considered. What kind of standard pattern is used in the voice recognition device depends on the size of the target vocabulary to be recognized, the versatility of the voice recognition device to have, the speech phenomenon expression ability given to the standard pattern, It is determined in consideration of several points such as the amount of learning voice data that can be used for learning the standard pattern.

First, consider from the viewpoint of the number of standard patterns. For example, in a small vocabulary voice recognition device that recognizes only 10 numbers, a standard pattern in word units is often used. On the other hand, in a speech recognition device that recognizes arbitrary sentences in Japanese, an enormous number of standard patterns (for example, 100,000 words) are required to prepare a standard pattern for each word, which is practically realized. Becomes impossible. Therefore,
A syllable or phoneme, which is a smaller unit (shorter in time length), is used as a unit of a standard pattern, and a method of integrating recognition results of syllables or phoneme units into a sentence recognition result is used. Alternatively, the phonological standard patterns are connected to create a standard pattern for each word, and the created standard pattern for each word is used for recognition. In the case of Japanese, since the number of syllables is about 120 and the number of phonemes is about 40, it is possible to cover all Japanese by preparing standard patterns of this number. That is, if a small voice unit is used, a wide range of recognition targets can be covered by a small number of standard patterns. On the other hand, if a large voice unit is used, many standard patterns are required.

Next, let us consider the speech pattern expression capability of the standard pattern. As described above, Japanese speech can be expressed in principle with about 120 syllables, but even in the same syllable, the preceding syllable and the following syllables greatly change the voice pattern (this is called articulatory combination). Also, even with the same syllable, the voice pattern varies greatly depending on the speaker. When one standard pattern is set for each syllable / phoneme unit, a single standard pattern represents different voice patterns depending on the surrounding environment and different voice patterns depending on the speaker. It is required to represent various transformations with one representative pattern,
There is a natural limit to their expressive ability. On the other hand, when a standard pattern is registered in a large unit such as a word, the standard pattern is created by including the deformation of the phonetic pattern based on the context of the syllables that make up the word. Will be taken into. That is,
From the viewpoint of the ability to express voice phenomena (particularly articulation), it is preferable to use a large voice unit for the standard pattern.

Consider next from the viewpoint of the amount of available learning voice data. As mentioned above, the total number of standard patterns can be kept small by using small voice units, and the total number of standard patterns by large units. Therefore, if the total amount of learning voice data that can be used for learning the standard pattern is constant, the smaller the unit, the larger the learning data amount for each standard pattern, and the more reliable the standard pattern can be created. In addition, when compared on the assumption that standard patterns of similar reliability are created, the larger the unit, the larger the required amount of training voice data.

From the above situation, it is common sense to use a standard pattern in word units when the number of vocabularies to be recognized is small, and to use a syllabic / phonological unit or a standard pattern in units equivalent thereto when the number of vocabularies to be recognized is large. Among the conventional techniques, as an improvement, a standard pattern of different sizes is used in a mixed manner, a standard pattern of a larger unit is created by connecting standard patterns of smaller units, a standard pattern with sufficient learning data and learning. There is an example in which a standard pattern with insufficient data is interpolated and used. Each conventional technique will be briefly described below.

An example of using a standard pattern in word units is The Bell System Technica.
l Journal Vol. 62, No. 4, Apr
il1983 “On the Application
of Vector Quantization a
nd Hidden Markov Modelsto
Speaker-Independent, Isol
There is an example of number recognition as described in "Arated Word Recognition". In this case, a standard pattern of each word is used for each number (in the case of the conventional example, Hidden Ma is used).
rkovModel: hereinafter referred to as HMM) is registered. Such a method is suitable for small vocabularies as described above, but registration becomes difficult as the number of vocabularies increases.

As an example of using the standard pattern of the phoneme unit, IEICE technical research report, SP89-94,
(1989-12) As described in "Performance evaluation of HMM-LR speech recognition system", a phoneme is cut out from learning speech data uttered word by word, and this is a standard pattern (HMM in the case of the conventional example). There is something used for learning.
In this conventional example, the standard pattern for each phoneme is prepared independently of the surrounding environment, and no consideration is given to articulatory coupling or transformation by the speaker.

As an example of using a standard pattern in syllable units, IEICE technical report, SP89-96,
(1989-12) "Syntax / Semantic Analysis Driven Japanese Continuous Speech Understanding System-SPOJUS-SYNO / SEMO"
As described in “Evaluation of −”, there is a method of using learning voice data uttered in syllable units for learning of a standard pattern (HMM in the case of the conventional example). It has been prepared independently of, and no consideration is given to articulatory coupling, and no consideration is given to deformation by the speaker.

As an example of using a standard pattern of a plurality of types of voice units having different sizes, an IEICE technical report SP88-96 (1988-12) "CV, continuous voice recognition system based on word spotting" As described in “”, there is a combination of standard patterns in syllable units and standard patterns in word units. In this conventional example, standard patterns in syllable units and standard patterns in word units are prepared for frequently-used words. At the time of recognition, both standard patterns are used to obtain the recognition result in syllable units and the recognition result in word units, and these are integrated into the final recognition result. In this conventional example, the articulation combination is dealt with for the frequent words, but other words are not taken into consideration. Also, there is no consideration for deformation by the speaker.

As an example of connecting standard patterns of small units to create a standard pattern of large units, a collection of papers on acoustics of Japan, 2-2-10, (1988-3) "Large pattern based on connection of syllable patterns" is given. As described in "Study on Learning Word Selection Method in Vocabulary Word Speech Recognition", there is a method of creating a standard pattern in syllable units as an initial standard pattern, and concatenating the standard patterns in word units. In this conventional example, the initial standard pattern is created from syllables that are uttered independently,
Further, averaging processing is performed with respect to the syllables cut out from the learning voice data uttered in word units. By this processing, the standard pattern in syllable units is corrected for word utterances, but no countermeasures against articulatory coupling are taken into consideration. Also, there is no consideration for deformation by the speaker.

[0012]

SUMMARY OF THE INVENTION The object of the present invention is to create and utilize a standard pattern capable of coping with articulatory coupling and deformation of a voice pattern by a speaker, which were not sufficiently taken into consideration in the above prior art. To provide the means.

[0013]

The object of the present invention is to prepare a highly reliable standard pattern of a voice unit smaller than a phoneme, and in the process of creating a standard pattern, cluster a group of similar speakers by clustering to create a standard pattern. Is created,
Furthermore, this can be achieved by creating a standard pattern of a large unit while utilizing the correspondence between the phoneme and this standard pattern.

[0014]

According to the present invention, by setting a standard pattern smaller (shorter in time length) than a phoneme, the degree of deformation to be covered in each standard pattern can be suppressed to a small level, and each standard pattern can be suppressed. The pattern is created with high accuracy.
Also, in the process of creating the standard pattern, the standard pattern is created while clustering similar speakers by clustering, so that the standard pattern that can accurately cover the variety of speakers is created. Further, the acoustic transformation depending on the phoneme environment is described in the phoneme-recognition basic unit standard pattern correspondence information. The above three effects are combined to enable highly accurate recognition.

[0015]

Embodiments of the present invention will be described below with reference to the drawings.
FIG. 1 is a block diagram showing an embodiment of a voice recognition device of the present invention. The input voice is converted into an electric signal by the voice input means 1. The voice converted into the electric signal is further analyzed by the voice analysis means 2, and the time series of the feature vector is output. On the other hand, in the standard pattern connection unit 33, the standard pattern of the recognition basic unit stored in advance in the standard pattern storage unit 4 is converted into the phoneme-recognition basic unit correspondence storage information stored in the phoneme-recognition basic unit correspondence storage unit 34. They are connected according to the word dictionary 35 to form a word standard pattern. The collating unit 31 collates the standard pattern of each word created by the standard pattern connecting unit 33 with the feature vector time series of the input voice, and obtains the similarity between the input voice and each word to be recognized. The determination unit 32 outputs the recognition result based on the similarity. For example, the word with the highest similarity is used as the recognition result.

Next, the standard pattern of the recognition basic unit used in the present invention will be described. FIG. 2 shows a probability model (Hidede) corresponding to the standard pattern used in the present invention.
nMarkov Model, hereinafter abbreviated as HMM). Each circle in the figure represents a state, and arrows represent transitions between states. The symbol a _ij attached to the arrow represents the probability that the transition from the state i to the state j occurs, and the symbol b _ij (k)
Represents the probability that a feature vector belonging to the kth classification will be observed when a transition from state i to state j occurs. This parameter can be used to calculate the posterior probability that the time series of feature vectors of voice data is output from this probabilistic model. A _ij , b _ij (k) using the speech data for learning
Estimating the value of is equivalent to standard pattern creation. A specific method of calculating the posterior probability and a specific method of parameter estimation are, for example, Kluwer Acad.
emic Publishers, Norwel, M
A, 1989 "Automatic Speech
Known methods such as those described in "Recognition", pp. 20-26 can be used.

Next, the phoneme-recognition basic unit correspondence relationship information stored in the phoneme-recognition basic unit correspondence relationship storage means 34 used in the present invention will be described. There are several possible methods for expressing the correspondence between phonemes and basic recognition units. A method of expressing phonemes by a sequence of recognition basic units, and a method of expressing phonemes by a network of recognition basic units. For example, there is a method of expressing a phoneme by an HMM using a basic unit of recognition as a symbol. FIG. 3 shows a method of expressing phonemes by a series of basic recognition units. In the example of FIG. 3, a series of syllables and recognition basic units is associated and expressed. For example, syllable / a /
Is a sequence of recognition basic units N1 and N2, or
It is said to be a series of recognition basic units N9 and N10.
Similarly, it is stated that the syllable / i / is a recognition basic unit N3, a series of recognition basic units N8 and N9, or a series of recognition basic units N32 and N90. The other syllables have the same description, and finally, the syllable / N / consists of the recognition basic unit N130.

Next, the standard pattern connecting means used in the present invention will be described with reference to FIG. FIG. 4 is an example in which HMMs of basic recognition units are connected to create an HMM of a word. Take the word "Hitachi (/ hitachi /)" as an example. First, from the word dictionary 35, "Hitachi" is syllable string / hi
It can be seen that it is composed of /, / ta /, / chi /. Further, the phonological-recognition basic unit correspondence relationship information stored in the phonological-recognition basic unit correspondence storage unit 34 makes it possible to know what kind of recognition basic unit each syllable is composed of. In the example of FIG. 4, the syllable / hi / is the recognition basic unit N1.
And N2, or recognition basic units N1 and N
It is a series of 4. The syllable / ta / is a sequence of recognition basic units N5 and N6. The syllable / chi / is a sequence of recognition basic units N12 and N9. The concatenation of HMMs can be realized by putting the last state transition of the preceding HMM into the first state of the succeeding HMM. Then the syllable sequence / hi /, / ta /, / c corresponding to the word
If the HMMs corresponding to each syllable of hi / are sequentially connected, the HMM of the word "Hitachi" is created. In the example of FIG. 4, syllables / h
Since there are two series of recognition basic units for i /, as a result, two "Hitachi" HMMs can be created. The first is a series of recognition basic units N1, N2, N5, N6, N12, N9. The second is the basic recognition unit N1, N4, N5, N6, N
12 and N9 series.

Next, the learning method of the recognition basic unit used in the present invention will be described with reference to FIG. Figure 5
It is a flowchart explaining the learning flow of the recognition basic unit used in this invention. First, the learning voice is segmented according to a predetermined criterion and divided into small (short time length) voice intervals. The partitioning based on a predetermined standard may be a method such as partitioning at regular time intervals. Next, one HMM is created for each segmented speech section, and this is used as an initial model. Specifically, the HMM is created by, for example, Kluwer Academic Pub.
lishers, Norwel, MA, 1989.
"Automatic Speech Recogni
section ", pp. 20-26. Then, parameter estimation is performed again on the initial model and each HMM is updated. Next, each HMM is updated.
Cluster. Clustering is similar HMM
Is a process for collecting clusters to form a cluster. There are various clustering methods, but here, for example, consider a method in which two HMMs having the highest degree of similarity are combined into one cluster. According to this method, one cluster reduces the number of clusters by one. HMM
As a measure of the degree of similarity of Kullback, for example, Kullback described in Technical Report of IEICE, SP90-64, (1990-12) "Prediction and evaluation of acoustic variation in unknown phoneme context by tree structure phoneme model". It is sufficient to use the information amount of. The process of parameter re-estimation and clustering is repeated, the process is terminated under a predetermined convergence condition, for example, the condition that the number of HMMs becomes smaller than a predetermined number, and the resulting HMM is used as a standard pattern of the recognition basic unit. Store.

Next, another learning method of the recognition basic unit used in the present invention will be described with reference to FIG. FIG. 6 is a flowchart explaining another learning flow of the basic recognition unit used in the present invention. While the learning flow described in FIG. 5 is a method of starting from a large number of initial models and sequentially reducing the number by clustering, the method described in FIG. 6 creates only one initial model and sequentially This is a method of dividing and increasing the number of models. First, the learning voice is segmented according to a predetermined criterion and divided into small (short time length) voice intervals. The partitioning based on a predetermined standard may be a method such as partitioning at regular time intervals. Next, one common HMM is created for each segmented voice section, and this is used as an initial model. Next, parameter estimation is performed again on the initial model to update each HMM. Next, each HMM is divided into two. This division may be performed by finely modifying the parameters of each HMM. For example, H shown in FIG.
Assuming that MM is an HMM before division, two models in which the value of b _ij (k) is finely corrected may be created. The two resulting models will be slightly different from the original model. Next, the speech for learning is recognized using the HMM that has been doubled. As a result, each HMM is associated with each segmented voice section, and parameter re-estimation is performed using this new correspondence. The process of parameter re-estimation, division into two, and recognition of learning speech is repeated, and a predetermined convergence condition, for example, the number of HMMs is a predetermined number (for example, 256).
The processing is terminated under the condition that the number of individual objects has been reached, and the resulting HMM is stored as a standard pattern of the recognition basic unit.

Next, the phoneme-recognition basic unit correspondence relationship learning means used in the present invention will be described with reference to FIG. FIG. 7 is a flowchart for explaining the flow of learning the phoneme-recognition basic unit correspondence. The phonological-recognition basic unit correspondence relationship learning method described here is an extremely simple method. That is, a voice whose utterance content is known is recognized using the standard pattern of the recognition basic unit, and the recognition result is obtained as a series of recognition basic units. Next, the phoneme sequence of the speech whose utterance content is known is associated with the sequence of the recognition basic unit of the recognition result. This association may be linear association, for example. Since the number of phonemes in the voice and the number of recognition basic units of the recognition result are not necessarily divisible by an integer,
Plural relations of phonological pair recognition basic units are provided, such as 1: 1, 1: 2, 1: 3. After the association processing, the frequencies are accumulated while weighting each of the plurality of association relationships according to the linear relationship. By the above, the phoneme-recognition basic unit correspondence is obtained, for example, in the form as shown in FIG.

[0022]

As described above, according to the present invention, even if a small voice unit is used as a standard pattern for a voice recognition apparatus, a standard voice pattern such as articulatory coupling is sufficiently reflected when recognition is performed. Because you can use patterns,
Highly accurate recognition is possible.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a voice recognition device of the present invention.

FIG. 2 is a diagram illustrating a hidden Markov model of a basic recognition unit used in the speech recognition apparatus of the present invention.

FIG. 3 is a diagram illustrating correspondence information between phonemes and standard patterns of basic recognition units used in the speech recognition apparatus of the present invention.

FIG. 4 is a diagram illustrating a method of connecting standard patterns of basic recognition units used in the speech recognition apparatus of the present invention.

FIG. 5 is a flowchart illustrating a learning flow of a standard pattern of recognition basic units used in the voice recognition device of the present invention.

FIG. 6 is a flowchart illustrating another learning flow of the standard pattern of the recognition basic unit used in the voice recognition device of the present invention.

FIG. 7 is a flowchart illustrating a learning means for learning a correspondence relationship between phonemes and standard patterns of recognition basic units used in the speech recognition apparatus of the present invention.

[Explanation of symbols]

1 ... voice input means, 2 ... voice analysis means, 3 ...
-Voice recognition means 31 ... collation means, 32 ... determination means, 33 ...
Standard pattern connecting means 34 ... Word dictionary, 4 ... Standard pattern storing means.

Claims (14)

[Claims] 1. A speech input means for inputting speech, a speech analysis means for analyzing the inputted speech and outputting a time series of feature vectors, and a speech section having a time length shorter than a phoneme as a basic recognition unit, A standard pattern storage means for storing a standard pattern for a recognition basic unit, a phonological-recognition basic unit correspondence relationship storage means for storing a correspondence relationship between a phoneme and a recognition basic unit, and a word dictionary describing recognition target words as a sequence of phonemes. And a standard pattern connecting means for connecting a standard pattern for a recognition basic unit to form a standard pattern for a recognition target word, and a time series of the feature vector of the input speech and the standard pattern formed by the connection. In the voice recognition device, which comprises a collating means, and performs recognition based on the collation result output from the collating means, the standard pattern connecting means includes the sound - speech recognition apparatus being characterized in that so as to constitute a standard pattern for recognition terms using the correspondence between the stored recognition basic unit corresponding relation storage unit phoneme recognition basic unit. 2. The correspondence between the phoneme and the basic recognition unit is
The speech recognition apparatus according to claim 1, wherein the phoneme is given as a series of at least one recognition basic unit. 3. The correspondence relation between the phoneme and the basic recognition unit is:
The speech recognition apparatus according to claim 1, wherein the phoneme is provided as a network of at least one recognition basic unit. 4. The correspondence between the phoneme and the basic recognition unit is:
The speech recognition apparatus according to claim 1, wherein the recognition basic unit is given as a hidden Markov model having a label. 5. A voice for learning is divided according to a predetermined reference, an acoustic model of a recognition basic unit is created for each divided voice section, and parameter estimation and estimation are performed for a group of the acoustic models. A method for learning a standard pattern of a basic recognition unit, which is characterized by repeatedly performing clustering. 6. Each of the divided sections of the learning voice is
The method for learning a standard pattern of a recognition basic unit according to claim 5, wherein a fixed section is used in the re-estimation of the parameters of the repeated acoustic model. 7. The divided sections of the learning voice are:
6. The learning method for a standard pattern of a recognition basic unit according to claim 5, wherein the section obtained as a result of recognition of the learned speech using the acoustic model repeatedly performed is corrected and used as needed. 8. A voice for learning is divided according to a predetermined reference, one acoustic model of a recognition basic unit common to each divided section is set, and thereafter, the acoustic model of the acoustic model using the voice for learning is set. Parameter estimation, a process of dividing each model into two models by modifying the parameters of each acoustic model in which the parameters are estimated, and using the acoustic model of which the total number is doubled for the learning, Voice recognition process,
A method for learning a standard pattern of a recognition basic unit, characterized in that the above three processes are repeatedly performed until a predetermined criterion is satisfied. 9. The divided sections of the learning voice are:
The method for learning a standard pattern of a recognition basic unit according to claim 8, wherein a fixed section is used in the re-estimation of the parameters of the acoustic model that is repeatedly performed. 10. The divided section of the learning voice is used by being corrected as needed to a section obtained as a result of recognition of the learning voice using the acoustic model repeatedly performed. A method for learning the standard pattern of the recognition basic units described. 11. A phonological unit and a recognition basic unit are obtained by recognizing a voice whose utterance content is known by using a standard pattern of a recognition basic unit and associating a sequence of recognition basic units obtained as a recognition result with a phoneme sequence of the utterance content. A learning method characterized by learning the correspondence between a phoneme for obtaining a correspondence with a unit and a recognition basic unit. 12. The learning method according to claim 11, wherein the correspondence relationship between the phoneme and the recognition basic unit is given as a sequence of at least one recognition basic unit. 13. The learning method according to claim 11, wherein the correspondence relationship between the phoneme and the recognition basic unit is given as a sequence of at least one recognition basic unit. 14. The learning method according to claim 11, wherein the correspondence relationship between the phoneme and the recognition basic unit is given as a hidden Markov model having the recognition basic unit as a label.