JPH0827638B2 - Phoneme-based speech recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an improvement of a speech recognition device capable of recognizing a large vocabulary continuous speech by using a phoneme, which is a basic unit constituting speech, as a recognition unit. Is.

(Prior Art) Conventionally, there has been a method of recognizing a voice by using a phoneme, which is a basic unit forming a voice, as a unit. In this method, first, a phoneme standard pattern which is the basis of recognition is obtained by learning. A word voice is recognized by using a standard pattern of words synthesized from this phoneme standard pattern according to a word dictionary in which phonemes are written. In this case, the term "phoneme" means a unit of recognition, and is used in a broader sense including not only phonemes in phonetics but also syllables and plural phoneme chains. Also, in addition to words as recognition targets,
Although there are clauses, sentences, etc., the case of recognizing a word will be described below.

As an example of a phoneme-based recognition method, for example, the Acoustical Society of Japan, Speech Research Material S85-62 (December 20, 1985)
, Pp. 477 to 484, in a paper titled "Syllabic-based Japanese Speech Recognition" (hereinafter referred to as Reference 1), consonant + vowel (CV; consonant is C, vowel is V And a vowel + consonant + vowel (VCV) as a recognition unit (phoneme). In this method, the learning voice continuously uttered word by word is converted into CV and VCV.
The phoneme standard pattern is created from the segmented speech and the phoneme standard pattern is created from the segmented speech (the creation and update of such a standard pattern is called learning). At the time of recognition, words (phrases) are recognized by the DP matching method based on the combination of the phoneme standard patterns obtained.

In addition, IEEE, International Conference on Acoustic
s, Speech, and Signal Processing, 1986, 30.9, page 1593, "The Role of Word-Dependent Coarticulatory Effe.
cts in a Phoneme-Based Speech Recognition System "
In a paper published as (1986) (hereinafter referred to as reference document 2), the Hidden Markov model (hereinafter referred to as "HMM") method is used as a unit of recognition of phonemes. The method of recognition is shown. This method will be described below.

First, we describe the method of finding the phoneme-based HMM (called a phoneme HMM, which is equivalent to the phoneme standard pattern) by learning. The learning voice is converted into the observation label string O (t); 1≤t≤T by the vector quantization method. Second
The figure shows an example of a phoneme HMM. The HMM is represented by a state transition network as shown in the figure. In the HMM, the output establishment bi (O (t)) of the observation label sequence O (t) in the state i, the state i
Each parameter of the state transition establishment a (i, j) from the state to the state j is defined.

When learning a phoneme HMM, first, a phoneme HMM is created in advance based on segmented speech data of a typical speaker. This is the initial model that is the initial value for phoneme HMM learning. For a new speaker, the learning model uttered by the speaker is used to update the initial model by the learning process, and the phoneme HMM of the speaker is created. This learning process can be executed using an FB (Forward-Backward) algorithm. For this FB algorithm, see, for example, "Structural Methods in Automatic Speech Recognit" on page 1652 of Proceedings of The IEEE, Vol.73, No.11.
This is described in detail in the paper entitled "ion" (November 1985) (hereinafter referred to as Reference 3). A method for learning a phoneme HMM using a learning voice uttered in word units. As a method, a method of repeating the operation of obtaining the parameters of a new phoneme HMM by collecting the intermediate parameters in the FB algorithm updated word by word in the phoneme HMM unit is used.

When recognizing the input speech, use the word HMM that combines the phoneme HMMs according to the word dictionary in which phonemes are written, and use the above F
The word HMM as the forward probability in the B algorithm
The appearance probability of the input voice for is calculated. The word with the highest appearance probability becomes the recognition result.

(Problems to be Solved by the Invention) In speech recognition using phonemes as a unit, the recognition unit is generally described in the cited document 1 rather than a single phoneme such as C or V (called a single phoneme). It is preferable to use a phoneme chain (called a compound phoneme) having a certain length such as CV and VCV. This is because the voice pattern greatly changes depending on what kind of phoneme exists before and after, and the change part from one phoneme to the next phoneme has a great feature.

However, in general, the types of compound phonemes are extremely large compared to the types of single phonemes. For example, the number of single phonemes such as C and V in Japanese is about 20, but there are more than 1000 kinds of compound phoneme VCV. Therefore, in order to learn all VCVs, a huge amount of learning speech including those compound phonemes is required, and there is a drawback that the amount of processing required for learning becomes extremely large.

Further, when a voice is uttered, voicing transformation such as devoicing and lengthening of vowels may occur. In order to deal with such voicing transformation, there is a method of preparing phoneme HMMs of unvoiced phonemes and long vowel phonemes in addition to normal phoneme HMMs. However, the presence / absence of voicing transformation occurs probabilistically, and if there is no particular voicing transformation in the training voice, or if all phonemes that may have transformations such as unvoiced voicing transformation occur. There is a case. In such a case, there is a drawback that a phoneme in which voicing deformation occurs or a standard pattern of phonemes without voicing deformation is not learned.

Furthermore, although the initial model is generally created based on the utterance of a typical speaker, when learning a phoneme HMM, the learning voice of the new speaker is the voice of the typical speaker. In the case of a large difference from the pattern, there is a drawback that the learning may not be performed correctly because the correspondence between the phoneme segment of the learning speech and the section of the phoneme HMM is largely deviated.

The present invention realizes a high-performance speech recognition apparatus by eliminating various drawbacks as described above and enabling various voicing transformations and standard patterns of compound phonemes of many types to be learned with a small amount of learning speech. Especially.

(Means for Solving Problems) A phoneme-based speech recognition device according to the first invention of the present application includes a phoneme learning unit that obtains a phoneme standard pattern from a learning voice, and the obtained phoneme standard. A phoneme combination unit that creates a compound phoneme standard pattern by combining one or more patterns, and a compound phoneme learning unit that learns using learning speech based on the created compound phoneme standard pattern,
And a voice recognition unit that recognizes an input voice using the composite phoneme standard pattern.

In addition to the first invention of the present application, a phoneme-based speech recognition device according to the second invention of the present application is based on the voicing modification detection unit for obtaining voicing modification information of a learning voice, and the voicing modification information. It has a single phoneme learning unit for learning and a compound phoneme learning unit for learning on the basis of the utterance transformation information.

In addition to the first and second inventions of the present application, the phoneme-based speech recognition device according to the third invention of the present application uses, as a standard, an initial standard pattern obtained in advance and an initial standard pattern obtained from learning speech. It has a phoneme selection unit that selects each pattern, and a single phoneme learning unit that obtains a single phoneme standard pattern from the learning voice using the selected initial standard pattern as an initial value.

(Operation) The operation of the phoneme-based speech recognition device according to the present invention will be described. In the following description, C is a single phoneme,
Phonemes such as V, CV and VCV at the beginning of compound phonemes
We will use phoneme chains such as. As a recognition method, phoneme-based HM as described in the cited document 2 is used.
I will use M. The same applies to other methods.

In the present invention, a compound phoneme HMM is used as a recognition unit. When learning this compound phoneme HMM by learning, if there are few learning voices or if voicing transformations are included, it is possible that there is no speech pattern corresponding to all defined compound phonemes in the learning voice. is there. As a result, there is a possibility that an unlearned compound phoneme HMM can be created.

In order to cope with this, in the present invention, when obtaining a compound phoneme HMM, first, a monophoneme HMM is obtained by learning. Monophone HMM
Is equivalent to a composite phoneme HMM divided. Compound phoneme V1C
1V2 (eg [asi]) is a single phoneme V1 ([a]), C1
([S]) and V2 ([i]). Monophone HMM
Since there are a limited number of types (about 20 types), it is easy to prepare a learning voice including all phonemes. It is possible to deal with the voicing deformation by replacing the phoneme in which the voicing deformation has occurred with a similar phoneme by using the monophone HMM. For example, a monophone HMM with unvoiced vowels has fricatives [s]
It can be dealt with by replacing with a single phoneme HMM. It is assumed that the content of the utterance transformation of the learning voice (for example, the presence or absence of unvoiced voice) is known in advance as the utterance transformation information.

As a method for learning the monophone HMM, for example, the method described in the cited document 2 can be used. FIG. 3 shows an example of a single phoneme HMM of a single phoneme [a]. Here, an HMM having two states as shown in the figure is used as a monophone HMM.
When learning a monophone HMM, a monophone HMM obtained from the voice of a typical speaker is used as an initial model. These monophoneme HMMs are combined according to a monophoneme notation word dictionary to create a word HMM. Figure 5 shows the monophone HMM.
[A], [a], [s], [a], [a], [h],
An example of the word HMM [asahi] obtained by combining [i] and [i] is shown. The single phonemes [a] and [i] of the vowel overlap because the division into the compound phonemes [asa] and [ahi] is taken into consideration.

Based on such a word HMM, a phoneme HMM is learned by the method described in the cited document 2 by using a learning voice in which a word is uttered. In the case of using the learning voice including the utterance transformation, the phoneme HMMs are combined according to the utterance transformation information to create the word HMM used for learning.

A compound phoneme HMM can be created by combining the single phoneme HMMs thus obtained. FIG. 4 is an example of a compound phoneme HMM [asa] obtained by combining [a], [s], and [a] of a single phoneme HMM. Due to the influence of articulatory coupling, the voice pattern may change even for the same phoneme depending on the type of phonemes before and after. Thus, although simple combination of single phoneme HMMs is not enough, it can be used as a compound phoneme HMM approximately.

Furthermore, in the present invention, the compound phonemes present in the learning speech are
Regarding the HMM, the compound phoneme HMM is trained by using the compound phoneme HMM created by synthesizing the single phoneme HMM as an initial model.
This learning can be performed as in the case of the monophone HMM.

As a result, it is possible to create a model that includes the influence of articulatory coupling on the compound phonemes existing in the learning speech.

As described above, according to the present invention, not only the compound phoneme existing in the learning speech but also the compound phoneme HMM can be created approximately even when it does not exist. Can be used for learning compound phoneme HMM.

In the above description, it is assumed that the voicing transformation in the learning voice is known in advance. However, in order to obtain voicing transformation information, it is necessary to use a method such as designating the presence or absence of voicing transformation when uttering a learning voice in advance.
For example, one vowel is made unvoiced and the other vowel is uttered so as not to be unvoiced. However, such a method imposes a burden on the user. On the other hand, in the present invention, it is possible to use a method of automatically detecting the vocalization deformation of the learning voice. The automatic detection method of this vocal deformation is described below.

First, prepare a word dictionary that covers all vocal variants,
A word HMM is created by combining the monophone HMMs obtained by a representative speaker according to those dictionaries. These words HMM
Then, the appearance probability of the learning voice is calculated, and the utterance transformation in the word dictionary with the highest probability is used as the utterance transformation of the learning voice.
This appearance probability can be obtained using the forward probability described in the cited document 3, as in the case of recognizing a word.

For example, [u], which is the fourth and sixth phonemes of "clapping" (pronunciation: [hakusyu]), may be devoiced. So all possible combinations [hakusy
u], [haku-syu], [hakusyu-], [haku-syu-]
The dictionaries (incapacitated [u] are represented as [u-]) are prepared, and voicing variants are obtained using the word HMMs created according to these dictionaries. For example, if the appearance probability of [haku-syu] is the highest, it is assumed that the first [u] is devoiced.

By using the vocalization deformation information obtained in this way, it becomes possible to learn by voice including the vocalization deformation.

Also, in the above description, the initial model is based on the utterance of a typical speaker. Usually, this allows learning of a single phoneme HMM, but if the new speaker's training speech is significantly different from the typical speaker's speech pattern, it may already be learned that learning may not be successful. Stated. Therefore, in the present invention, a method of directly creating an initial model of a single phoneme, such as a vowel, whose pattern variation may be large depending on the speaker, is used.

Here, a case where an initial model of a vowel is created from learning voice is described, but the same applies to an initial model other than a vowel. First, a speech that is easy to segment is prepared for each phoneme as a learning speech. For example, if a vowel uttered alone is used as a learning voice, a voice section of the voice data (for example, a portion having an amplitude larger than a certain level) can be a segment of a vowel monophoneme.
Further, for example, various voices can be used for initial model creation by cutting out a vowel section using the segmentation method described from page 73 of Kyoritsu Publishing “Speech Recognition”.

This learning speech can be segmented into single phonemes by the above method, and the single phoneme HMM can be learned using the FB algorithm for each single phoneme, using the speech in the single phoneme segment as the training speech. In this case, there is no risk that the learning voice segment and the monophone HMM are associated with each other, so a random number value or a monophone HMM made by a typical speaker can be used as an initial model for learning. In addition, only the observation label output probability bi (x) is used based on the appearance frequencies of all the observation labels in the corresponding segment, and the state transition probability a (i, j) is that of a typical speaker. Can be used as is.

In this way, stable learning can be performed by obtaining the initial model of a single phoneme from the voice of the speaker who is learning.

Although the method using the HMM has been described above, the same applies to the case of using the DP matching method as described in the cited document 1. In this case, as the phoneme standard pattern,
A single phoneme standard pattern and a compound phoneme standard pattern are prepared. As a standard pattern iterative learning method using DP matching, the following method can be used. Using the word standard pattern that is obtained by connecting the initial phoneme standard patterns obtained in advance according to the word dictionary, the learning speech is segmented into phoneme units by DP matching. The obtained phoneme unit segments are averaged among the same phonemes to create a new phoneme standard pattern. This operation is repeated to update the phoneme standard pattern.

By using such an iterative learning method, the learning method of the standard pattern used for DP matching can be treated in the same manner as the method using the FB algorithm of HMM, and thus the learning method according to the present invention can be applied.

(Example) An example of a phoneme-based speech recognition apparatus according to the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the present invention. First, a learning method for obtaining a compound phoneme HMM used for recognition will be described.

The initial model memory 1 holds parameters (initial model) of the monophone HMM that are initial values for learning the monophone HMM. This initial model is obtained in advance by segmenting the voice uttered by a typical speaker into individual phonemes. This can be realized by applying the FB algorithm as described in Reference 3 for each phoneme. As the monophone HMM, a model consisting of two states as shown in FIG. 3 is used here.

The initial model learning voice memory 2 holds the voice data of the initial model learning voice converted into the observation label sequence by the vector quantization method. This voice data is input to the segmentation unit 3 and segmented into single phonemes. Here, a vowel that is uttered alone is used as the initial model learning voice, and a portion of the voice data whose amplitude is larger than a predetermined value is used as vowel monophoneme data. The parameter creating unit 4 obtains the parameters of the monophone HMM by the FB algorithm based on the input monophone data.

The initial model selection unit 5 uses the single phoneme HMM in the initial model memory 1 and the single phoneme HMM obtained by the parameter creation unit 4 according to a predetermined rule.
Is selected and output as the initial model. For example, the rule is applied that the initial model of the vowel (single phoneme HMM) is used from the parameter creating unit 4, and the others are selected from the initial model memory 1. If the initial model learning speech is not used, the single phoneme in the initial model memory 1
Output the HMM as the initial model.

The learning voice memory 6 holds the learning voice data represented by the observation label string. Speech deformation detection unit 7
Then, a dictionary that covers all utterance variants of words corresponding to the learning voice data is read from the word dictionary memory 8. The initial model is combined according to the notation of this word dictionary, and the word HMM for each voicing transformation is created. Subsequently, with respect to these word HMMs, the appearance probability of the learning voice data is obtained as a forward probability, and the voicing transformation of the word HMM having the maximum appearance probability is calculated.
The utterance modification information memory 9 holds the utterance modification information of the learning voice data. Further, as the vocalization transformation information, in addition to the information obtained in this way, information obtained by investigating the learning voice in advance can be used.

The phoneme learning unit 10 learns a phoneme HMM using the voice data in the learning voice memory 6. This is to first select a word dictionary containing utterance transformations in the word dictionary memory 8 according to the utterance transformation information in the utterance transformation information memory 9 and combine the initial models to create a word HMM corresponding to the learning voice. Similar to the method described in the reference document 2, the parameters of this word HMM are updated, and the updated parameters are collected in units of single phonemes to advance the learning processing of the single phoneme HMM. This learning process is repeated until the parameters converge.

The learned single phoneme HMMs are combined according to a predetermined rule in the single phoneme combination unit 11 to form a compound phoneme HM.
M is made. As a rule, for example, compound phoneme [asa]
There is a method in which monophonemes [a], [s], and [a] are combined and created as shown in FIG.

In the compound phoneme learning unit 12, the compound phoneme HMM created by the single phoneme combination unit 11 is used as an initial model, and the learning voice memory 6
The compound phoneme HMM is trained by using the voice data in it. The learning method is the same as the method used in the monophone learning unit 10.

The obtained compound phoneme HMM is held in the compound phoneme HMM memory 13.

Next, a method of recognizing the input voice will be described.
The recognition process is performed in the recognition unit 14. This recognition method is the same as the method described in Reference 2. That is, the input speech is converted into the observation label string by the vector quantization method. Next, according to the word dictionary in the word dictionary 8, the compound phoneme HMMs in the compound phoneme HMM memory 13 are combined to sequentially create the word HMMs. The appearance probability of the input speech with respect to this word HMM is obtained as a forward probability, and the word with respect to the word HMM having the maximum appearance probability becomes the recognition result.

(Effect of the Invention) According to the present invention, since many voicing variations and standard patterns of compound phonemes having many types can be learned with a small amount of learning speech, a high-performance speech recognition device can be realized. .

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG.
FIG. 4 is a diagram showing an example of a phoneme HMM, FIG. 3 is a diagram showing an example of a single phoneme HMM, FIG. 4 is a diagram showing an example of a compound phoneme HMM, and FIG.
The figure is a diagram showing an example of the word HMM. In the figure, 1 ... Initial model memory, 2 ... Initial model learning voice memory, 3 ... Segmentation unit, 4 ... Parameter generating unit, 5 ... Initial model selecting unit, 6 ... Learning voice memory, 7
... voicing transformation detector, 8 ... word dictionary memory, 9 ... voicing transformation information memory, 10 ... single phoneme learning unit, 11 ... single phoneme combining unit, 12
… Composite phoneme combination unit, 13… Composite phoneme HMM memory, 14… Recognition unit.

Claims (3)

[Claims] 1. A monophoneme learning unit that obtains a monophoneme standard pattern from a learning voice, and a phoneme combination unit that creates a compound phoneme standard pattern by combining one or more of the obtained monophoneme standard patterns, Phoneme-based speech recognition including a compound phoneme learning unit that learns using a learning voice based on the created compound phoneme standard pattern, and a recognition unit that recognizes an input voice using the compound phoneme standard pattern. apparatus. 2. A voicing deformation detection unit for obtaining voicing deformation information of learning voice, a single phoneme learning unit for learning based on the voicing deformation information, and a compound phoneme for learning based on the voicing deformation information. The phoneme-based speech recognition device according to claim 1, further comprising a learning unit. 3. An initial standard pattern obtained in advance, a phoneme selection unit for selecting an initial standard pattern obtained from a learning voice for each standard pattern, and a learning voice with the selected initial standard pattern as an initial value. 3. A phoneme-based speech recognition device according to claim 1, further comprising a phoneme learning unit that obtains a phoneme standard pattern.