JPH09134192A - Statistical language model forming device and speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the prediction accuracy and reliability of a transition probability by setting part-of-speech bigrams into an initial state and executing two kinds of state sepn. of word separation and integration of concatenated words from part-of-speech classes, thereby forming a statistical language model of variable length N-grams. SOLUTION: A word hypothesis converging section 6 converges the word hypothesis in such a manner that the word hypothesis of the same words of the equal end time and the different start time is represented by the one word hypothesis having the highest likelihood by referencing the statistical language model 22. The statistical language model 22 used here separates the words reliable with the words alone from the part-of-speech classes and handles these words as the independent classes on the basis of the bigram (N=2) between the part-of-speech classes. As to frequently appearing word strings, the model integrates these word strings and handles the same as one class, thereby making the expression of the long word concatenation possible. The statistical language model 22 having the characteristics of the part-of-speech bigrams and the variable length word N-grams in combination is thus formed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a statistical language model generating apparatus for generating a statistical language model based on learning text data, and a statistical language model using the statistical language model.
The present invention relates to a voice recognition device that recognizes a voice signal of an input uttered voice sentence.

[0002]

2. Description of the Related Art In recent years, a method of using a language model has been studied to improve the performance of a continuous speech recognition apparatus. This aims to improve the recognition rate and reduce the calculation time by predicting the next word and reducing the search space using a language model. N-Gram (N-Gram) is a language model that has been widely used recently.
m). It learns large-scale text data and statistically gives the transition probability from the previous N-1 words to the next word. Multiple L word strings w ₁ ^L = w
The generation probability P (w ₁ ^L ) of ₁ , w ₂ , ..., W _L is expressed by the following equation.

[0003]

(Equation 1)

[0004] Here, w _t represents a t-th one word of the word string w ₁ ^L, w _i ^j represents the j-th word string from the i-th. In the above _equation 1, the probability P (w _t |
wt _{+ 1-} ^Nt-1 ) is a word sequence wt _{+ 1-} ^Nt-1 composed of N words.
Is the probability that the word w _t will be uttered after is uttered, and similarly, the probability P (A | B) means the probability that the word A will be uttered after the word or word string B has been uttered. Also,
“Π” in Equation 1 represents the probability P (w _t from t = 1 to L
| W _{t + 1−N} ^t−1 ), and so on.

N-grams are extremely simple, yet easy to construct, compatible with statistical acoustic models,
It is very effective for continuous speech recognition because it has a large effect of improving the recognition rate and shortening the calculation time (for example, in the conventional document 1, “LR Bahl et al.,“ A Maximummum ”).
Likelihood Approach to C
ontinous Speech Recognit
Ion ”, IEEE Transaction on
Pattern Analysis and Mach
ine Intelligence, pp. 179-
190, 1983 ", the conventional document 2" PC Wood "
Land et al., "THE 1994 HTK Larg"
e Vocabulary Speech Recog
Nation System ”, Proceeding
s of ICASSP95 ', Vol. 1, pp. 7
3-76, 1995 ”, conventional document 3“ Murakami et al., “Sentence Speech Recognition Using Word Trigram and Extension to Free Speech Recognition”, IEICE Technical Research Report, SP9.
3-127, pp71-78, 1994 ". ).

Generally, in the N-gram language model, the accuracy of the next word is improved by handling a long word chain when N is increased, but the number of parameters is large and the frequency of appearance is low when the learning data amount is small. It is not possible to give a word a reliable transition probability. For example, the number of vocabulary is 5,
When there are 000 words, trigram (N
= 3), the transition set of all words is (5,000) ³ = 1,
Since it is 25 billion, a huge amount of text data of several hundred billion words or more is required to obtain a reliable transition probability. It is virtually impossible to collect such a huge amount of text data. Conversely, if N is made small, the reliability of the transition probability becomes high, but only a short word chain can be handled, and the prediction accuracy of the next word becomes low.

[0007]

In order to solve this problem, the following method has been proposed. (1) Method of Estimating Unlearned Transition Probability by Interpolation This method is based on, for example, Deleted Interpo
(deletion interpolation method) (for example, conventional document 4
"F. Jelinek et al.," Interpolate
d estimation of Markov So
urce Parameters from Spar
se Data ”, Proceedings of Wo
rkshop Pattern Recognitio
n inPractice, pp. 381-37, 19
80 years ". ), Back-off Smooth
ing method (conventional document 5 “SM Katz,“ Esti.
marriage of Probabilities f
rom Sparse Data for the L
Angular model component of
a Speech Recognizer ", IEEE
E Transaction on Acoustic
s, Speech, and Signal Pro
cessing, Vol. ASSP-35, No. 3,
pp. 400-401, March 1987 ". ) Etc., a small N N-gram (N-gra).
By interpolating the transition probability with the value of m), the transition probability can be given to a word transition that does not exist in the learning text data. However, there is a possibility that a reliable transition probability cannot be given to a word having a low appearance frequency.

(2) Method for reducing the number of parameters by class N-gram This method uses clustering based on mutual information (see, for example, the conventional document 6 “PF Brown et al.,“ Clas.
s-Based n-gram model of n
atatural language ”, Computat
ional Linguistics, Vol. 18, N
o. 4, pp 467-479, 1992 ". )
And part-of-speech (Conventional document 7 "Zhou et al.," Large vocabulary continuous speech recognition of Japanese by probabilistic model ", Information Processing Society of Japan, Proc. Of the 51st National Convention, pp. 119-120, 1995).
reference. ) Etc., the sentence generation probability P (w ₁ ^L ) of ^L words is generally expressed by the following equation.

[0009]

(Equation 2)

Here, c _t represents a class to which the word w _t belongs, and c _i ^j represents an i-th to j-th class string. In the above _equation 2, P ( _ct | _{ct-N +} ^{1t + 1} ) is the immediately preceding (N-1)
Represents the transition probability from the class to which this word belongs to the class to which the next word belongs. When the number of classes is 50, the set of transitions between all the classes of the trigram is 50 ³ = 125,00.
Since it is 0, it is considered that the transition probability can be obtained with text data of several hundred thousand words, which is considerably smaller than the word N-gram. However, since it is not possible to express a unique concatenation relationship between words, it is considered that the prediction accuracy of the next word becomes poor.

An object of the present invention is to solve the above-mentioned problems, and to generate a statistical language model capable of generating a statistical language model capable of improving the prediction accuracy and reliability of transition probability as compared with the conventional example. An object of the present invention is to provide a generation device and a speech recognition device that can perform speech recognition at a higher speech recognition rate than a conventional example using the statistical language model.

[0012]

According to a first aspect of the present invention, there is provided a statistical language model generating apparatus according to the first aspect of the present invention, in which all vocabularies are part-of-speech based on learning text data in which uttered voice sentences of a predetermined speaker are written down. Based on the generating means for classifying each part of speech into a part-of-speech class and generating a bigram between these part-of-speech classes as a statistical language model of the initial state, and the statistical language model of the initial state generated by the generating means. , A first separation class candidate that can be separated from a part-of-speech class of a word, a combination of a word and a word, a combination of a word string of a word and a plurality of words, a word of a plurality of words Separation from a word part-of-speech class by a concatenated word or a concatenation of concatenated word strings, including a concatenation of a sequence and a word, a concatenation of a plurality of word sequences and a concatenation of a plurality of word sequences. Search means for searching for possible second separation class candidates and predetermined entropy representing the difficulty level of prediction of the next word for the first and second separation class candidates searched by the search means. , Calculating means for calculating a reduction amount of the entropy by separating the classes, and selecting the largest class separation among the reduction amounts of entropy for the first and second separation class candidates calculated by the calculating means And performing the separation of the selected class to generate a statistical language model including the bigram of the part of speech and the N-gram of the word of variable length N, and the statistics generated by the separation means. Until the number of classes of the dynamic language model reaches a predetermined number of classes, the statistical language model generated by the separating means is used as the processing target model, And processing, the processing of the calculation means, by repeating the process of the separating means, characterized in that a control means for generating a statistical language model having a predetermined number of classes.

According to a second aspect of the present invention, there is provided a voice recognition device comprising voice recognition means for recognizing voice using a predetermined statistical language model based on a voice signal of an input utterance voice sentence. In the device, the voice recognition means,
It is characterized in that speech recognition is performed using a statistical language model including a part-of-speech bigram and a variable length N word N-gram.

According to a third aspect of the speech recognition apparatus, the statistical language model is generated by the statistical language model generation apparatus according to the first aspect.

A continuous speech recognition apparatus according to a fourth aspect of the present invention continuously detects the word hypothesis of the uttered voice sentence based on the input voice signal of the uttered voice sentence and calculates the likelihood, In a continuous voice recognition device equipped with voice recognition means for recognizing voices, the voice recognition means comprises:
With reference to the statistical language model generated by the described statistical language model generation device, with respect to the word hypothesis of the same word having the same end time but different start time, utterance start for each head phoneme environment of the word It is characterized in that the word hypotheses are narrowed down so as to be represented by one word hypothesis having the highest likelihood of the total likelihood calculated from the time to the end time of the word.

[0016]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows a block diagram of a continuous speech recognition apparatus according to an embodiment of the present invention. The continuous speech recognition apparatus of the present embodiment uses a known one-pass Viterbi decoding method to detect the word hypothesis of the uttered voice sentence based on the characteristic parameters of the voice signal of the input uttered voice sentence. In a continuous speech recognition apparatus having a word matching unit 4 for calculating and outputting a degree, a word hypothesis of the same word output from the word matching unit 4 via the buffer memory 5 and having the same end time but different start time is used. On the other hand, with reference to the statistical language model 22, one having the highest likelihood of the total likelihood calculated from the utterance start time to the end time of the word for each head phoneme environment of the word. It is characterized in that a word hypothesis narrowing unit 6 for narrowing down the word hypotheses so as to be represented by the word hypotheses is provided.

The statistical language model 22 used here is generated by the language model generator 20 based on the learning text data, and the statistical language model 22 is used.
Is based on a bigram (N = 2) between parts-of-speech classes, but separate and reliable words are separated from the parts-of-speech class and treated as a single class. For the sequence, those words are combined and treated as one class to enable the expression of a long word chain. In this way, the generated model is a statistical model that has characteristics of a part-of-speech bigram and a variable length word N-gram. It is a language model, and is characterized in that the accuracy and reliability of transition probabilities are balanced.

First, the variable length N used in this embodiment.
-The concept of grams is explained below. The N-gram is a (N-1) -weighted Markov model, which is as simple as (1) to remember past (N-1) state transitions.
Heavy) It is interpreted that each state of the Markov model is separated. As an example, FIG. 3 shows a state transition diagram in which a bigram is modeled as a Markov model, and FIG. 4 shows a state transition diagram in which a trigram is modeled as a Markov model.

[0019] In FIG. 3, but remains in state s ₁ when output symbol a in state s _1, the state s ₁
Makes a transition to the state s ₂ in which the symbol b is output. But it remains in the state s ₂ when outputting the symbol b in the state s _2,
When the symbol a is output in the state s ₂ , the state returns to the state s ₁ . In the trigram of FIG. 4, the state s ₁ of the bigram is changed to the state s ₁₁
And state s ₁₂ and state s ₂ is separated into state s ₂₁ and state s ₂₂ . Furthermore, by proceeding with the separation of all the states, a higher-order N-gram is obtained.

The variable length N-gram shown in FIG. 5 is a partial separation of the states of a simple Markov model. That is, in the bigram of FIG. 3, when the symbol a is output from the state s ₂ , when the symbol b is continuously output (this is referred to as ab and the symbol ab is output), then other than b is continuously output. The symbol is output (this is referred to as a (/ b) and the symbol a (/ b) is output. Here, / represents a negative meaning and is a bar (overline)). In the former case, the state s _{1 is changed} to the state s ₁₂ , while in the latter case, the state s _{2 is changed} to the state s ₁₁ . That is, in the former case,
The state s ₁ is separated into the state s ₁₂ , and the remaining transition (a (/ b)) that outputs the symbol a is left in the state s ₁₁ . In this model, when the symbol ab is output in the state s ₁₁ , the state transitions to the state s ₁₂ while the state s 11
_{When the} symbol a (/ b) is output at ₁₁ , the state remains s ₁₁ . Further, while in the state s ₁₂ and remain in the state s ₁₂ when outputting the symbol ab, while s ₁₂ symbols a (/
When b) is output, the state transits to the state s ₁₁ .

This model is characterized in that a long chain can be represented with the structure of the simple Markov model as it is by treating a plurality of continuous symbols as new symbols. By repeating the same state separation, it is possible to locally represent a longer chain. This is a variable length N-gram. That is, the variable-length word N-gram as a language model in which the symbols are regarded as words is represented as a bigram between word strings (including one word).

Next, the operation of the variable length N-gram will be described. Statistical language model 22 used in this embodiment
Is a variable length N-gram of a part-of-speech class and a word, and is expressed as a bigram between the following three types of classes. (1) Part of speech class (hereinafter referred to as the first class),
(2) A class of words separated from the part-of-speech class (hereinafter referred to as a second class), and (3) a class formed by combining concatenated words (hereinafter referred to as a third class).

The words belonging to the first class have a low frequency of occurrence, and the reliability of the transition probability is improved as compared with the case of handling the words alone. In addition, the words belonging to the second class are mainly high in frequency of occurrence, and are sufficiently reliable to be handled alone, and further, concatenated words are combined and classified into the third class. , Variable length N-gram, the prediction accuracy of the next word is improved. However, in the present embodiment, the concatenation of the part-of-speech class and the part-of-speech class and the combination of the part-of-speech class and the word that are connected are not considered. Generation probability P (w of a sentence composed of a plurality of L words
₁ ^L ) is given by the following equation.

[0024]

(Equation 3)

Here, ws _t means a t-th word string (including a single word) when sentences are classified into the above classes. _{Therefore, P (ws t | c t} ) is the probability that the word string ws _t appears when found t-th _{class, P (c t | c t} -1) is the previous (t- 1) Probability that the word of the t-th class appears from the 1st class. Further, K of the sentence represents the number of word strings, and K ≦ L. Therefore, Π in Equation 3 is a product of t = 1 to K.
Here, as an example, consider a sentence of an uttered voice sentence consisting of the following seven words.

[0026]

[Equation 4] "I-Murayama-and-Word-I-Mas"

The generation probability P (w ₁ ^L ) of this sentence is given by the following equation using the equation (3).

[0028]

[Formula 5] P (w ₁ ^L ) = P (Watashi | {Watashi}) ・ P ({Watashi}) ・ P (Murayama ｜ <proper noun>) ・ (<Proper noun> | {Wataku}}) P ( And | {and}) ・ P ({and} | <proper noun>) ・ P (say |｜ say]) ・ P ([say] | {and})

However, <>, {}, and [] represent that they belong to the first class, the second class, and the third class, respectively. However, each word and word string belong as follows. (1) "Murayama" belongs to the first class because it is a noun. (2) "Watakushi" and "to" are a combination of a noun and a particle, and belong to the second class. (3) “I say” is a combination of a verb, a verb suffix, an auxiliary verb, and an auxiliary verb suffix, and belongs to the third class. Here, in the second and third classes, the appearance frequency of the word and the class is the same, so P (Watashi | {Watashi})
= 1 and P (and | {and}) = 1, P (say | [say]) = 1, and therefore, the above equation 5 becomes the following equation.

[0030]

[Equation 6] P (w ₁ ^L ) = P (Watakushi) -P (Murayama ｜ <proper noun>)-P (<proper noun> ｜ Watakushi) -P (and | <proper noun>)-P (say | And)

Next, a language model generation process for generating the statistical language model 22 which is a variable length N-gram used in this embodiment will be described with reference to FIG. The statistical language model 22 used in the present embodiment is generated in the form of class separation based on a criterion of entropy minimization, with a part-of-speech class bigram as an initial state. The reduction of entropy is guaranteed to be positive, and the class separation reduces the entropy monotonically with respect to the training text data. The entropy used here is generally
It represents a measure of "ambiguity", and a small entropy in a language model means that the language is less ambiguous and the next word can be easily predicted. That is, entropy represents the difficulty of predicting the next word. Entropy H (X | Y) of conditional probability P (x | y) that is the probability of x under the condition of y
Is represented by the following equation.

[0032]

## EQU7 ## H (X | Y) =-. SIGMA.P (y) .SIGMA.P (x | y) l
og ₂ (x | y)

Therefore, the entropy used in this embodiment is calculated by the following equation based on the above equation (7).

[0034]

(Equation 8) Where w _k ∈ c _j

FIG. 6 is a flow chart showing the details of the language model generation processing executed by the language model generation unit 20, and the processing will be described below with reference to FIG. First, in step S1, all vocabulary included in the learning text data in which the uttered voice sentence of a predetermined speaker is written is made into a part-of-speech class (here, a part-of-speech class is a class clustered for each part-of-speech). We classify and make bigram between these classes a statistical language model in the initial state. Next, classes are separated in the next steps S2 to S4. That is, in step S2, a list is made by searching for a separated class candidate that can be separated into classes. Here, consider the following two types of class separation. (1) Separation of words from a part-of-speech class (hereinafter referred to as first class separation), (2) Class separation by combining concatenated words or concatenated word strings (hereinafter referred to as second class separation). Here, the concatenation of concatenated words or concatenated word strings means
Concatenation of one word and one word that are concatenated (input adjacently in time), one word and a word string of a plurality of words, a word string of a plurality of words and a word Binding,
It includes a word string of a plurality of words and a combination of the word strings of a plurality of words.

In separating the former word from the part-of-speech class, words originally belonging to the part-of-speech class are separated from the class, and the separated words form a single class with the word.

[0037]

## EQU9 ## c _ξ → {w _x } + c _ξ \ {w _x } where w _x ∈ c _ξ

Here, c _ξ \ {w _x } means a class excluding the class of the word w _x from the class c _ξ ,
The word w _x belongs to the class c _ξ . Therefore, what is meant by Equation 9 is, for example, that the noun class c _ξ is “desk”.
The class {w _x } of the word w _{x and} the word w of "desk"
_This means separating the class {w _x } of _x into a class excluding the class c _ξ .

In the latter class separation by concatenating concatenated words or concatenated word strings, concatenation between two concatenated classes is considered for a word class and a word string class already separated from the initial class. The combined word strings form a single class with the word strings.

[0040]

[Expression 10] {w _x } → {w _x , w _y } + {w _x , / w _y }

Here, {w _x , w _y } is a concatenated word string w _x ,
represent classes of w _y, representing the class of {w _x, / w _y} is not word w _y is subsequent to the next word w _x word w _x. That is, / w _y represents a word other than the word w _y . The meaning of the expression 10 is, for example, the class of the word "desk" {w
_x } is the class of the word sequence “desk” {w _x , w _y }
And a class of word strings such as “desk is” and “desk is” other than “of desk” {w _x , / w _y }. The above formula 10 is an expression relating to the word combination, but the word string-word combination and the word string-word string combination are also similarly expressed. Therefore, the second class separation includes these class separations.

Then, in step S3, the entropy reduction amount is calculated using the following equations 11 and 12 for the first and second separated class candidates listed in step S2. Here, Equation 11 is used for the first class separation, which is the initial class separation, while the second class separation is used.
Equation 12 is used for class separation by concatenating concatenated words or concatenated word strings, which is the class separation of.

[0043]

ΔH = H ({c _i }) − H ({c _i \ c _ξ } + {w _x } + {c _ξ \ w _x })

Equation 12] _{ΔH = H ({c i}} ) - H ({c i \w x} + {w x, w y} + {w x, / w y})

Here, in equations 11 and 12, H
({C _i }) is the entropy for all the original part-of-speech classes c _i , and H ({c _i \ c _ξ } in Eq.
_{+ {W x} + {c} ξ \w x}) is the entropy of the time separating the classes of word w _x from all the original word class class c _i, [Delta] H having 11 separate classes of word w _x It is the amount of decrease in entropy. Further, the number _{12 H ({c i \w x} } + {w x, w y} + {w x, / w y})
Is a word string {w _x , from all original part-of-speech classes c _i
is the entropy when the class of w _y } is separated,
ΔH in Expression 12 is the amount of decrease in entropy when the classes of the word string {w _x , w _y } are separated.

Next, in step S4, the entropy reduction amount ΔH calculated in step S3 among all the separated class candidates listed in step S2.
Actually separate only the class that maximizes. Then, in step S5, it is determined whether or not the number of separation classes has become equal to or larger than the desired number of separation classes of a predetermined threshold value (for example, 500, 1000, etc.). Repeat the process. On the other hand, if the number of classes is equal to or larger than the desired number of separated classes in step S5, the obtained statistical language model 22 is stored in the memory in step S6, and then the language model generation process is terminated. Since the algorithm of this language model generation processing does not combine parts of speech and between parts of speech and words, at the time of completion of generation, a statistical language model that combines the features of the bigram of parts of speech and the N-gram of variable-length words is used. 22.

In FIG. 1, connected to the word matching unit 4,
For example, a phoneme HMM1 stored in a hard disk memory
1 is represented including each state, and each state has the following information. (A) State number (b) Acceptable context class (c) List of preceding states and succeeding states (d) Parameter of output probability density distribution (e) Probability of self transition and transition probability to subsequent states The phoneme HMM 11 used in the embodiment is generated by converting a predetermined speaker mixed HMM because it is necessary to specify which speaker each distribution is derived from. Here, the output probability density function is a Gaussian mixture mixture having a 34-dimensional diagonal covariance matrix.

The word dictionary 12 connected to the word matching unit 4 and stored in, for example, a hard disk is a phoneme HMM.
For each of the eleven words, a symbol sequence indicating a reading represented by a symbol is stored.

In FIG. 1, the uttered voice of the speaker is input to the microphone 1 and converted into a voice signal, and then input to the feature extraction unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and obtains a 34-dimensional feature parameter including logarithmic power, 16th cepstrum coefficient, Δlog power, and 16th Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the word matching unit 4 via the buffer memory 3.

The word matching unit 4 uses the one-pass Viterbi decoding method, and uses the phoneme HMM 11 and the word dictionary 12 based on the feature parameter data input via the buffer memory 3 to perform a word hypothesis. Is detected and the likelihood is calculated and output. Here, the word matching unit 4 determines whether each HMM
The likelihood within a word and the likelihood from the start of utterance are calculated for each state. The likelihood is individually provided for each word identification number, word start time, and difference between preceding words. Further, in order to reduce the amount of calculation processing, the grid hypothesis of a low likelihood among the total likelihoods calculated based on the phoneme HMM 11 and the word dictionary 12 is reduced. The word collating unit 4 outputs the resulting word hypothesis and likelihood information to the word hypothesis narrowing unit 6 via the buffer memory 5 together with time information (specifically, a frame number, for example) from the utterance start time. .

The word hypothesis narrowing unit 6 refers to the statistical language model 22 based on the word hypothesis output from the word matching unit 4 via the buffer memory 5, and the end times are the same and the start times are different. The word hypothesis of the word is represented by one word hypothesis having the highest likelihood of the total likelihood calculated from the utterance start time to the end time of the word for each head phoneme environment of the word. After narrowing down the word hypotheses as described above, the word line of the hypothesis having the maximum total likelihood is output as the recognition result among the word strings of all the word hypotheses after the narrowing down. In the present embodiment, preferably, the first phoneme environment of the word to be processed is the three phonemes including the final phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. Say a line.

[0051] For example, as shown in FIG. 2, the (i-1) th word W _i-1 of the following, a phoneme string a _1, a _2, ..., comes i-th word W _i, which consists of a _n Sometimes, six hypotheses Wa, Wb, Wc, Wd, We, Wf are used as word hypotheses for the word W _i-1.
Exists. Here, the former three word hypotheses Wa, W
It is assumed that the final phonemes of b and Wc are / x /, and the final phonemes of the latter three word hypotheses Wd, We and Wf are / y /. Of the hypotheses in which the end time t _e is equal to the head phoneme environment (in FIG. 2, the top phoneme environment is “x / a ₁ / a ₂ ”, the three word hypotheses from the top), the hypothesis with the highest total likelihood (for example, FIG. Two
Are deleted except for the top hypothesis). Since the fourth hypothesis from the top has a different phoneme environment, that is, since the last phoneme of the preceding word hypothesis is y instead of x,
Do not delete the fourth hypothesis from the top. That is, only one hypothesis is left for each final phoneme of the preceding word hypothesis. In the example of FIG. 2, one hypothesis is left for the final phoneme / x /, and one hypothesis is left for the final phoneme / y /.

In the above embodiment, the leading phoneme environment of the word is defined as three phoneme sequences including the final phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. Although defined, the present invention is not limited to this, and the phoneme string of the preceding word hypothesis including the final phoneme of the preceding word hypothesis and at least one phoneme of the preceding word hypothesis continuous with the final phoneme, and the word May be a phoneme sequence including a phoneme sequence including the first phoneme of the word hypothesis.

In the above embodiment, the feature extraction unit 2
The word collation unit 4, the word hypothesis narrowing unit 6, and the language model generation unit 20 are composed of, for example, a digital computer, the buffer memories 3 and 5 are composed of, for example, a hard disk memory, and the phoneme HMM 11 and words are composed. The dictionary 12, the learning text data 21, and the statistical language model 22 are stored in a storage device such as a hard disk memory.

In the above embodiment, the speech recognition is performed using the word collating unit 4 and the word hypothesis narrowing unit 6.
The present invention is not limited to this, and includes, for example, a phoneme collation unit that refers to the phoneme HMM 11 and a speech recognition unit that performs speech recognition of a word by referring to the statistical language model 22 using, for example, the One Pass DP algorithm. May be.

[0055]

EXAMPLES In order to confirm the performance of the statistical language model 22 used in the present embodiment, the present inventor compared the perplexity and the number of parameters with the conventional word N-gram. The data used in the experiment is a natural speech travel conversation database (hereinafter referred to as “Mori”) owned by the applicant.
Moto et al., “A Speech and Lang”
uage Database for Speech
Translation Research ”, ICS
LP, pp 1791-1794, 1994 ". )
And is composed of 354 dialogues and 354,700 words. Of these, 828 dialogues, 347,299 words are used as learning text data (also referred to as a training set) for generating the statistical language model 22, and the remaining data is used as test text data (also referred to as a test set). .) In the statistical language model 22 according to this embodiment, the initial class has 80 parts-of-speech including the inflected forms, is separated up to 1000 pieces, and data is collected every 100 pieces. In addition to the statistical language model 22 according to the present embodiment and the word N-gram, as a measure against the unknown word transition, the transition probabilities of the class and the word are interpolated by the deletion interpolation method (see conventional document 4), and the test set is set. In, when an unknown word appeared, a predetermined fixed value (for example, 7.0 × 10 ⁻⁶ ) was given. Here, perplexity is used to evaluate the statistical language model 22 according to the present invention. For example, entropy H when there is a long word string w ₁ ⁿ = w ₁ w ₂ ... W _n consisting of a plurality of n words
(N) is expressed by the following equation.

[0056]

[Equation 13] H (n) =-(1 / n) · log ₂ P (w ₁ ⁿ )

Here, P (w ₁ ⁿ ) is the generation probability of the word string w ₁ ⁿ , and the perplexity PP (n) is expressed by the following equation.

[0058]

## EQU14 ## PP (n) = 2 ^{H (n)}

Here, the perplexity when the test text data is used as the word string is called the test set perplexity, and the perplexity when the learning text data is used as the word string is the training set perplexity. Tee.

FIG. 7 shows how the value of the test set perplexity changes with the experimental result. As is clear from FIG. 7, the test set perplexity decreases as the number of separation classes increases, and the number of separation classes increases by 2
It can be seen that the value of 00 is the same as that of the word trigram, and the number of separated classes is 600, which is about the same value as the word trigram. When the number of separation classes is 600 or more, the rate of decrease in perplexity becomes extremely small. Therefore, it is considered that the statistical language model 22 of the present embodiment works most effectively when the number of separation classes is about 600. Therefore, the statistical language model 22 of the present embodiment is considered to be a language model having a prediction accuracy equal to or higher than the word trigram and higher than the word trigram.

In Table 1, the number of separation classes is 0,500,
The value of perplexity at the time of 1000 and the number of parameters are shown.

[0062]

[Table 1] Performance comparison of each language model ─────────────────────────────────── Bigram Trigram Form (number of separated classes) 0 500 1000 ─────────────────────────────────── Test set 20.31 16.96 41.68 17.61 16.75 Perplexity ─────────────────────────────────── Training set 13.50 5.99 48.77 18.77 15.05 Perplexity ─ ────────────────────────────────── Number of parameters (1) 4.10 × 10 ⁷ 2.62 × 10 ¹¹ 1.28 × 10 ⁴ 3.43 × 10 ⁵ 1.17 × 10 ⁶ ─────────────────────────────────── Number of parameters (2) 52 , 244 165, 139
7,991 27,830 43,075 ────────────────────────────────────

Here, the number of parameters (1) means the number of combinations of transitions of all classes (words), and the number of parameters (2) means the number of combinations of class (words) transitions that actually exist in the training set. To do. From Table 1, in the statistical language model 22 of the present embodiment, the difference in perplexity between the test set and the training set is
It can be seen that it is very small compared to the word bigram and the word trigram. The number of parameters is 1000
Even when the classes are separated, less than the word bigram,
Much less than the word trigram. Therefore,
It can be said that the statistical language model 22 of the present embodiment is an excellent language model that can efficiently express language features with given parameters. Therefore, the statistical language model 22
Is considered to be a more reliable language model than the conventional word bigram and word trigram.

The statistical language model 22 of this embodiment is also used.
FIG. 8 shows the result of examining the change in the value of the test set perplexity by changing the number of learning words in order to confirm the reliability of. As is clear from FIG. 8, when all learning sets (about 350,000 words) are used, the word bigram and the statistical language model 22 (200 classes) of the present embodiment (the numbers in parentheses are the separation classes). The same applies to the statistical language model 2 even if the number of learning words is reduced.
It can be seen that the increase in perplexity of 2 is relatively small and lower than the word bigram. Similarly, even when the word trigram is compared with the statistical language model 22 (600 class), the statistical language model 22 exhibits lower perplexity when the number of learning words decreases.

Next, the present inventor applied the statistical language model 22 of this embodiment to the continuous speech recognition apparatus of FIG. 1 and confirmed the effect of the statistical language model 22. Table 2 shows the experimental conditions for phoneme recognition. Table 2 also shows the sound parameters.

[0066]

[Table 2] Experimental conditions ─────────────────────────────────── Analysis conditions Sampling frequency: 12 KHz, Hamming window : 20 ms, frame period: 10 ms ─────────────────────────────────── Used parameter 16th LPC cepstrum + 16th ΔCepstrum + log power + Δlog power ─────────────────────────────────── Acoustic model HM net by gender Speaker model 400 states, 5 mixed ───────────────────────────────────

In Table 2, for the male-and-female unspecified speaker models of the HM network, the conventional literature 9 "Kosaka et al.," Unspecified speaker speech recognition by mixed speaker SSS ", Proceedings of the Acoustical Society of Japan, 2- 5-9, pp135-136, 1992 "
Is disclosed. In this experiment, an acoustic model and a language model were applied to a continuous speech recognizer using a continuous speech recognition method using a word graph. The recognition target is 16 dialogues in the training set of the statistical language model 22, and the test set not used for learning is 18 dialogues. Table 3 shows the correct word content rate of the sentence recognition candidate having the first highest likelihood in each language model.

[0068]

[Table 3] Correct word content rate ─────────────────────────────────── Bigram This embodiment (separation class Number) 0 500 ─────────────────────────────────── Dictionary subset test set 71.4 67.3 72 .2 Training Set 69.4 63.4
69.7 ─────────────────────────────────── Dictionary full set Test set −− 57.1 58. 4 Training set --- 54.6 56.0 ────────────────────────────────────

In Table 3, the dictionary subset is a dictionary in which only the words included in the recognition target are registered (750).
, Dictionary full set is a dictionary (6,400 words) consisting of all words used for learning to generate a statistical language model.
Represents However, in the conventional word bigram, due to the memory capacity and the calculation time, recognition of the full dictionary is not performed in this experiment. In this case, in other words, a large capacity memory and a huge processing time are required.

Regarding the test set, the perplexity is ascending, that is, the statistical language model 2 of the present embodiment.
The correct word content rate is improved in the order of 2 (0 class) → word bigram → statistical language model 22 (500 class) of this embodiment, and the statistical language model 22 of this embodiment is shown.
In the (500 class), the correct word content rate is slightly higher than the word bigram. Regarding the training set, the statistical language model 22 of the present embodiment is used.
(500 class) has higher perplexity than the word bigram, but the correct word content rate is high.
Further, since the statistical language model 22 of this embodiment has a small number of parameters, it can be easily expanded to recognize a large vocabulary. Therefore, the statistical language model 22 of the present embodiment is considered to be a more effective language model than the word bigram even for continuous speech recognition.

As described above, the accuracy of N-gram
It is possible to realize a device for generating a statistical language model 22 of variable length N-gram and a continuous speech recognition device using the same for the purpose of improving reliability. The statistical language model 2
2 is generated by initializing a part-of-speech bigram, separating words from a part-of-speech class, and concatenating concatenated words into two kinds of states. Part-of-speech bigram and variable-length word N-gram are generated. It is a model that also has features. As a result of the evaluation experiment of the statistical language model 22, it was found that the statistical language model 22 can achieve perplexity equal to or larger than the word bigram and equal to the word trigram with far fewer parameters, and the target performance is It was confirmed that it was realized. Moreover, as a result of applying it to continuous speech recognition, we were able to obtain the correct word content rate similar to the word bigram. Since the statistical language model 22 can be realized with a small number of parameters, it can be easily extended to speech recognition of a large vocabulary.

Accordingly, it is possible to provide a statistical language model generation device capable of generating the statistical language model 22 capable of improving the prediction accuracy and reliability of the transition probability, and also the statistical language model 22. It is possible to provide a continuous speech recognition device capable of continuously recognizing speech with a higher speech recognition rate by using.

[0073]

As described above in detail, according to the statistical language model generating apparatus according to the first aspect of the present invention, the statistical language model generating apparatus according to the present invention is based on the learning text data in which the uttered voice sentence of the predetermined speaker is written. The vocabulary of is classified into a part-of-speech class clustered for each part-of-speech, and a generation means for generating a bigram between these part-of-speech classes as a statistical language model of the initial state, and a statistical language of the initial state generated by the generation means. A first separation class candidate capable of separating a word from a part-of-speech class based on a model, a combination of one word and one word, a combination of a word string of one word and a plurality of words, and a plurality of combinations. From the word part-of-speech class by a concatenated word or a concatenation of concatenated word strings including a concatenation of a word string of one word and a single word, a combination of a word string of a plurality of words and a combination of a word string of a plurality of words. Search means for searching for a possible second separation class candidate, and a predetermined entropy representing the difficulty level of prediction of the next word for the first and second separation class candidates searched by the search means. And a calculation means for calculating a reduction amount of the entropy by separating the classes, and a maximum class among the reduction amounts of entropy for the first and second separation class candidates calculated by the calculation means. Separation means for generating a statistical language model including a bigram of parts of speech and an N-gram of words of variable length N by selecting separation and performing separation of the selected class; The statistical language model generated by the separating means is set as a processing target model until the number of classes of the statistical language model is a predetermined number of classes, Comprising a processing of the serial retrieval means, the processing of the calculation means, by repeating the process of the separating means, and control means for generating a statistical language model having a predetermined number of classes. Therefore, it is possible to generate a statistical language model that can improve the prediction accuracy and reliability of the transition probability. Also, since the statistical language model can be realized with a small number of parameters,
It has the unique advantage that it can be easily extended to large vocabulary speech recognition.

In the speech recognition apparatus according to the second aspect of the present invention, the speech recognition means includes a speech recognition means for recognizing the speech signal of the uttered speech sentence by using a predetermined statistical language model. In the recognition device, the voice recognition means performs voice recognition using a statistical language model including a bigram of a part of speech and an N-gram of a word of variable length N. Therefore, since the speech recognition is performed using the statistical language model that can improve the prediction accuracy and reliability of the transition probability, it is possible to provide the speech recognition device that can perform the speech recognition with a higher speech recognition rate.

Further, in the speech recognition apparatus according to claim 3, the statistical language model is generated by the statistical language model generation apparatus according to claim 1. Therefore, since the speech recognition is performed using the statistical language model that can improve the prediction accuracy and reliability of the transition probability, it is possible to provide the speech recognition device that can perform the speech recognition with a higher speech recognition rate.

A continuous speech recognition apparatus according to a fourth aspect of the present invention continuously detects the word hypothesis of the uttered voice sentence based on the input voice signal of the uttered voice sentence and calculates the likelihood. In a continuous voice recognition device equipped with voice recognition means for recognizing voices, the voice recognition means comprises:
With reference to the statistical language model generated by the described statistical language model generation device, with respect to the word hypothesis of the same word having the same end time but different start time, utterance start for each head phoneme environment of the word The word hypotheses are narrowed down so that one word hypothesis having the highest likelihood of the calculated total likelihood from the time to the end time of the word is represented. That is, as compared with the conventional word pair approximation method in which each preceding word is represented by one word hypothesis, to handle collectively the preceding phonemes of the first phoneme of a word (that is, the final phoneme of the preceding word) , The number of word hypotheses can be reduced, and the approximation effect is large. In particular, when the number of words increases, the reduction effect is large. Therefore, even when the continuous speech recognition device is used for recognizing natural utterances in which interjections are inserted, stagnant, and rephrased frequently, the calculation cost required for merging or dividing word hypotheses is lower than in the conventional example. It will be smaller than that. That is, the amount of processing required for speech recognition is reduced, and therefore, the storage capacity required for the storage device for speech recognition is reduced, while the amount of processing is reduced, so that the processing time for speech recognition is reduced. can do. Further, since the speech recognition is performed by using the statistical language model capable of improving the prediction accuracy and reliability of the transition probability, the continuous speech recognition device capable of continuously recognizing the speech with a higher speech recognition rate is provided. be able to.

[Brief description of the drawings]

FIG. 1 is a block diagram of a continuous speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a timing chart showing a process of a word hypothesis narrowing section 6 in the continuous speech recognition device of FIG.

FIG. 3 is a state transition diagram showing a statistical language model of a bigram.

FIG. 4 is a state transition diagram showing a statistical language model of a trigram.

5 is a state transition diagram showing a model under a variable length N-gram used in the continuous speech recognition apparatus of FIG.

FIG. 6 is a flowchart showing a language model generation process executed by the language model generation unit 20 of FIG.

7 is a graph showing test set perplexity with respect to the number of separated classes in the statistical language model generated by the language model generation unit 20 of FIG.

8 is a graph showing test set perplexity with respect to the number of words of learning data in the statistical language model generated by the language model generation unit 20 of FIG.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3, 5 ... Buffer memory, 4 ... Word collation part, 6 ... Word hypothesis narrowing part, 11 ... Phoneme HMM, 12 ... Word dictionary, 20 ... Language model generation part, 21 ... Text data for learning, 22 ... Statistical language model.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Yoshinori Kosaka, No. 5 Mihiradani, Seiji-cho, Seika-cho, Soraku-gun, Kyoto Prefecture (72) In-T-V Speech Translation Laboratory (72) Shoichi Matsunaga, Kyoto Prefecture Soraku, Kyoto Prefecture Gunma Seika-cho, Osamu Osamu, Osamu Osamu, 5 Hiratani, A-T Co., Ltd.

Claims (4)

[Claims] 1. Based on learning text data in which a uttered voice sentence of a predetermined speaker is written, all vocabularies are classified into a part-of-speech class clustered for each part-of-speech, and a bigram between these part-of-speech classes is initialized. And a first separation class candidate that can be separated from the part-of-speech class of the word based on the initial-state statistical language model generated by the generation means. Combining one word with one word Combining one word with a word string of multiple words, combining word strings with multiple words with one word, word string with multiple words with multiple words Search means for searching a concatenated word including a combination with a word string or a second separated class candidate that can be separated from a part-of-speech class of a word by combining the concatenated word strings; and the searching means. And a calculation means for calculating a reduction amount of the entropy due to class separation, using a predetermined entropy representing the difficulty level of prediction of the next word for the first and second separation class candidates retrieved by By selecting the largest class separation among the reduction amounts of entropy for the first and second separation class candidates calculated by the calculating means, and executing the separation of the selected class, the part-of-speech bigram And a separating means for generating a statistical language model including N-grams of words of variable length N, and the separating means until the number of classes of the statistical language model generated by the separating means reaches a predetermined number of classes. Repeating the processing of the search means, the processing of the calculation means, and the processing of the separation means with the statistical language model generated by And a control means for generating a statistical language model having a predetermined number of classes. 2. A voice recognition device comprising a voice recognition means for recognizing a voice using a predetermined statistical language model based on a voice signal of an uttered voice sentence inputted, wherein the voice recognition means is a part-of-speech bigram. A speech recognition apparatus, which performs speech recognition using a statistical language model including a variable length N word N-gram. 3. The speech recognition device, wherein the statistical language model is generated by the statistical language model generation device according to claim 1. 4. Continuous speech recognition provided with a voice recognition means for continuously recognizing a voice by detecting a word hypothesis of the uttered voice sentence based on an input voice signal of the uttered voice sentence and calculating a likelihood. In the apparatus, the speech recognition means refers to the statistical language model generated by the statistical language model generation apparatus according to claim 1, and refers to the word hypotheses of the same word having the same end time but different start times. , For each head phoneme environment of the word, the word hypothesis is narrowed down so that it is represented by one word hypothesis having the highest likelihood of the total likelihood calculated from the utterance start time to the end time of the word. A continuous speech recognition device characterized by performing.