JP2886121B2 - Statistical language model generation device and speech recognition device - Google Patents

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. Recently, N-gram (N-gram) has been widely used as a language model.
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.

Although the N-gram is extremely simple, it is easy to construct, has good compatibility with the statistical acoustic model,
It is very effective for continuous speech recognition because it has a significant effect of improving the recognition rate and shortening the calculation time (for example, see LR Bahl et al., "A Maximum" in Reference 1).
Likelihood Approach to C
ontinous Speech Recognit
ion ", IEEE Transaction on
Pattern Analysis and Mach
intelIntelligence, pp. 179-
190, 1983 "and Conventional Document 2" PC Wood
Land et al., "THE 1994 HTK Large.
e Vocabulary Speech Recog
Nation System ”, Proceeding
s of CASSP95 ', Vol. 1, pp. 7
3-76, 1995 ", and conventional literature 3," Murakami et al., "Extension to Sentence Speech Recognition and Free Utterance Recognition Using Word Trigram", IEICE Technical Report, SP9.
3-127, pp71-78, 1994 ". ).

In general, the N-gram language model increases the accuracy of the next word by handling a long word chain when N is increased, but the frequency of appearance is low when the number of parameters is large and the amount of learning data is small. Words cannot be given reliable transition probabilities. For example, if the vocabulary number is 5,
For 000 words, trigram (N
= 3) is (5,000) ³ = 1
Since it is 25 billion, a large amount of text data consisting of hundreds of billions of 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, when N is reduced, the reliability of the transition probability increases, but only short word chains can be handled, and the prediction accuracy of the next word decreases.

[0007]

In order to solve this problem, the following method has been proposed. (1) Method of Estimating Unlearned Transition Probability by Interpolation This method uses, for example, Deleted Interpo.
ration (deletion interpolation method) (for example,
"F. Jelinek et al.," Interpolate
de estimation of Markov So
source Parameters from Spar
se Data ", Proceedingsof Wo
rkshop Pattern Recognition
n inPractice, pp. 381-37, 19
80 years ". ) And Back-off Smooth
ing method (conventional document 5, "SM Katz," Esti
nation of Probabilities f
rom Sparse Data for the L
angle model Componentof
a Speech Recognizer ", IEEE
E Transaction on Acoustic
s, Speech, and Signal Pro
sessing, Vol. ASSP-35, No. 3,
pp. 400-401, March 1987 ". )), A small N-gram (N-gram)
By interpolating the transition probabilities with the value of m), transition probabilities can be given to word transitions that do not exist in the learning text data. However, there is a risk that reliable transition probabilities may not be given for words that appear infrequently.

(2) Method for Reducing the Number of Parameters by Class N-gram This method is based on clustering based on mutual information (for example, see Reference 6 “PF Brown et al.,“ Class ”
s-Based n-gram modelofn
atarulanguage ", Computat
ionicLinguistics, 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 Using Probability Model ”, Information Processing Society of Japan, Proc. Of the 51st Annual Convention, pp119-120, 1995.)
reference. )), And the sentence generation probability P (w ₁ ^L ) of ^L words is generally expressed by the following equation.

[0009]

(Equation 2)

[0010] Here, c _t represents a class that belongs word w _t, c _i ^j represents the j-th class sequence from the i-th. In the above _equation 2, P ( _ct | _{ct-N +} ^{1t + 1} ) is ^{equal to} the immediately preceding (N-1).
It represents the transition probability from the class to which the word belongs to the class to which the next word belongs. When the number of classes is 50, the set of transitions between all classes of the trigram is 50 ³ = 125,000.
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 specific connection between words, the prediction accuracy of the next word is considered to be 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 which converts all vocabulary parts of speech based on learning text data in which uttered voice sentences of a predetermined speaker are written. Generating means for classifying each part-of-speech class into clustered part-of-speech classes, and generating a bigram between those part-of-speech classes as an initial state statistical language model; and an initial state statistical language model 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 one word and one word, a combination of one word and a word string of a plurality of words, a word of a plurality of words Separation from a word class by combining concatenated words or concatenated word strings, including concatenation of strings with one word, concatenation of word strings of multiple words, and word strings of multiple words Search means for searching for a possible second separation class candidate, and for the first and second separation class candidates searched for by the search means, using a predetermined entropy representing the difficulty of predicting the next word. Calculating means for calculating the amount of reduction in entropy by separating classes, and selecting the largest class separation from the amount of reduction in entropy for the first and second separation class candidates calculated by the calculation means And separating the selected class to generate a statistical language model including the part-of-speech bigram and the N-gram of the variable-length N word, and a statistic generated by the separating unit. Until the number of classes of the statistical language model reaches a predetermined number of classes, the statistical language model generated by the separating means is used as a processing target model, and the search method is performed. 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 speech recognition apparatus comprising: a speech recognition unit for recognizing a speech based on a speech signal of an input uttered speech sentence using a predetermined statistical language model. In the apparatus, the voice recognition means includes:
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 present invention, the statistical language model is generated by the statistical language model generating apparatus according to the first aspect.

According to a fourth aspect of the present invention, a continuous speech recognition apparatus detects a word hypothesis of the uttered speech sentence based on an input speech signal of the uttered speech sentence and calculates a likelihood to obtain a continuous speech. In a continuous speech recognition apparatus provided with a speech recognition means for recognizing a speech, the speech recognition means comprises:
With reference to the statistical language model generated by the described statistical language model generation device, for each word hypothesis of the same word having the same end time and different start time, the utterance starts for each head phoneme environment of the word. It is characterized in that word hypotheses are narrowed down so as to be represented by one word hypothesis having the highest likelihood among the total likelihoods 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 according to the present embodiment detects the word hypothesis of the uttered speech sentence based on the characteristic parameter of the speech signal of the input uttered speech sentence using a known one-pass Viterbi decoding method. In a continuous speech recognition device provided with a word matching unit 4 for calculating and outputting a degree, a word hypothesis of the same word having the same end time and different start time, which is output from the word matching unit 4 via the buffer memory 5, is obtained. On the other hand, with reference to the statistical language model 22, one of the total likelihoods calculated from the utterance start time to the end time of the word for each head phoneme environment of the word has one of the highest likelihoods. A word hypothesis narrowing unit 6 for narrowing down word hypotheses so as to be represented by 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.
Is based on bigrams between part-of-speech classes (N = 2), but words that can be independently trusted are separated from part-of-speech classes, treated as a single class, and frequent words are used to improve prediction accuracy. For sequences, those words are combined and treated as a class, allowing the representation of long word chains, and thus the model generated is a statistical model that combines the features of part-of-speech bigrams and variable-length word N-grams. It is a language model, characterized by a balance between accuracy and reliability of transition probability.

First, the variable length N used in this embodiment
-The concept of gram is explained below. The N-gram is a (N-1) -fold Markov model, which is simple (1-1) to store the past (N-1) state transitions.
Heavy) Each state of the Markov model is interpreted as being separated. As an example, FIG. 3 shows a state transition diagram in which a bigram is represented as a Markov model, and FIG. 4 shows a state transition diagram in which a trigram is represented 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
In a transition to a state s ₂ which has output the symbol b. But it remains in the state s ₂ when outputting the symbol b in the state s _2,
Back in state s ₂ to the state s ₁ when outputting the symbol a. Figure 4 trigram, state status s ₁ bigrams s ₁₁
Vital separated into state s ₁₂ and it is believed that separation of the state s ₂ in the state s ₂₁ and the state s _22. Further, by promoting the separation of all states, a higher-order N-gram is obtained.

The variable length N-gram shown in FIG. 5 is obtained by partially separating the states of the simple Markov model. That is, in the bigram of FIG. 3, when the symbol a is output from the state s ₂ and the symbol b is continuously output (this is referred to as ab, and the symbol ab is output), the state other than b is continued. (This is expressed as a (/ b) and the symbol a (/ b) is output. Here, / represents a negative meaning and is a bar (overline)). in the former case, while transitioning from state s ₁ to the state s _12, in the latter case, the transition from state s ₂ to the state s _11. That is, in the former case,
Separated from the state s ₁ to state s _12, in which left remaining transition for outputting the symbols a to (a (/ b)) in the state s _11. Incidentally, in this model, whereas a transition to the state s ₁₂ when outputting the symbol ab in the state s _11, the state s
_{When the} symbol a (/ b) is output in step ₁₁ , the state remains s11. Further, while in the state s ₁₂ and remain in the state s ₁₂ when outputting the symbol ab, while s ₁₂ symbols a (/
b) a transition to the state s ₁₁ when the output.

This model has a feature that long chains can be represented with a simple Markov model structure by regarding a plurality of consecutive symbols as new symbols. By repeating the same state separation, a longer chain can be represented locally. This is the variable length N-gram. That is, a variable-length word N-gram as a language model in which a symbol is regarded as a word 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 first class),
(2) A class of a word separated from the part of speech class (hereinafter, referred to as a second class), and (3) a class formed by combining connected words (hereinafter, referred to as a third class).

The words belonging to the first class have a low appearance frequency, and the reliability of the transition probability is improved as compared with the case where the words are handled alone. In addition, the words belonging to the second class are mainly those having a high frequency of appearance, and have sufficient reliability even if handled alone. Furthermore, the words connected to each other are combined and classified into the third class. , Operates as a variable-length N-gram, and the prediction accuracy of the next word is improved. However, in the present embodiment, it is not considered that the part-of-speech class and the part-of-speech class that are connected to each other and the combination of the part-of-speech class and the word are 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)

[0025] In this case, ws _t means the sentence at the time of the classification of the above class, t-th word string (a stand-alone word, is also included). _{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 a word of the t-th class appears from the class. Further, K in the text represents the number of word strings, and K ≦ L. Therefore, Π in Equation 3 is a product from t = 1 to K.
Here, as an example, consider a sentence of an uttered voice sentence composed of the following seven words.

[0026]

[Equation 4] "I, Murayama, and-Words-I-Mas-"

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

[0028]

P (w ₁ ^L ) = P (Watakushi | ｛Watakushi) · P (｛Watakushi) · P (Murayama | <proper noun>) · P (<proper noun> | ｛Watakushi) ・ P (And | ｛and｝) · P (｛and｝ | <proper noun>) · P (say | [say]) · P ([say] | ｛and｝)

However, <>, ｛｝, and [] indicate 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 proper noun. (2) “I” and “to” are words separated from nouns and words separated from particles, respectively, 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 frequency of appearance of the word and the class is equal, so that P (I |
(I) = 1, P (and | ｛and｝) = 1, and P (say | [say]) = 1, 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 a statistical language model 22 which is a variable length N-gram used in the present 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 the criterion for minimizing entropy with the bigram of the part of speech class as an initial state. The decrease in entropy is guaranteed to be positive, and the class separation causes the entropy to decrease monotonically with respect to the training text data. The entropy used here is generally
It represents a measure of "ambiguity". In a language model, a small entropy means that the language has small ambiguity and that the next word is easy to predict. That is, entropy indicates the difficulty of predicting the next word. Entropy H (X | Y) of conditional probability P (x | y) which is the probability of x under the condition of y
Is represented by the following equation.

[0032]

H (X | Y) = − {P (y)} P (x | y) log ₂ P (x | y)

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

[0034]

(Equation 8) Where w _k ∈c _j

FIG. 6 is a flowchart showing details of the language model generation processing executed by the language model generation unit 20. The processing will be described below with reference to FIG. First, in step S1, all the vocabulary included in the learning text data in which the uttered voice sentence of a predetermined speaker has been written is converted into a part-of-speech class (here, the part-of-speech class means a class that is clustered for each part of speech). Classify, and the bigram between those classes is used as an initial statistical language model. Next, classes are separated in the next steps S2 to S4. That is, in step S2, a list is made by searching for separated class candidates that can be separated into classes. Here, the following two types of class separation are considered. (1) Separation of a word from a part-of-speech class (hereinafter, referred to as a first class separation), (2) Class separation by combining a connected word or a connected word string (hereinafter, referred to as a second class separation). Here, the concatenation word or the concatenation of the concatenation word string is
Concatenation (joined in time) of one word and one word, concatenation of one word and plural word strings, plural word word string and one word Join,
It includes a word string of a plurality of words and a combination of the word strings of a plurality of words.

In the former separation of the word from the part-of-speech class, a word originally belonging to the part-of-speech class is separated from the class, and the separated word forms a single class by the word.

[0037]

[Mathematical formula 9] c _ξ → ｛w _x ｝ + c _ξ ＼ ｛w _x ｝ where w _x ∈c _ξ

Here, c _ξ {w _x } means a class obtained by removing the class of the word w _x from the class c _、,
Word w _x belongs to the class c _ξ. Therefore, the meaning of Expression 9 is, for example, that the class of the noun _cｃ is “desk”
Class {w _x } of the word w _{x and} the word w of “desk”
means separated into the classes excluding the _x classes {w _x} from class c _xi].

In the latter class separation by connecting the connected words or the connected word strings, a connection between two connected classes is considered for the word class and the word string class that have already been separated from the initial class. The combined word string forms a single class with the word string.

[0040]

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

Here, {w _x , w _y } is a concatenated word sequence 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. In other words, / w _y represents a word other than the word w _y. The meaning of Equation 10 is, for example, the class “w” of the word “desk”
_x ｝ is the class {w _x , w _yの
Means that the words are separated into classes {w _x , / w _y } of word strings other than “desk”, such as “desk” and “desk”. Equation (10) is an expression relating to the combination of words, but the combination of a word string and a word and the combination of a word string and a word string are similarly expressed. Therefore, the second class separation includes these class separations.

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

[0043]

Equation 11] _{ΔH = H ({c i}} ) - H ({c i \c ξ} + {w x} + {c ξ \w x})

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 original part-of-speech classes c _i , and H ({c _i ＼c _ξ }
+ ｛W _x ｝ + ｛c _ξ ＼w _x ｝) is the entropy when the class of the word w _x is separated from all the original part-of-speech classes c _i , and ΔH in Equation 11 separates the class of the word w _x It is the amount of decrease in entropy when doing. Further, the number _{12 H ({c i \w x} } + {w x, w y} + {w x, / w y})
The word string from the original of all the part-of-speech class c _i {w _x,
entropy when the class of w _yク_ラス is separated,
ΔH in Expression 12 is the amount of decrease in entropy when the class of the word sequence {w _x , w _y } is separated.

Next, in step S4, of all the separation class candidates listed in step S2, the entropy reduction amount ΔH calculated in step S3 is used.
Only the class that maximizes is actually separated. Then, in step S5, it is determined whether or not the number of separation classes is equal to or more than a desired number of separation classes of a predetermined threshold (for example, 500, 1000, etc.). Repeat the process. On the other hand, if the number is equal to or larger than the desired number of separation classes in step S5, the obtained statistical language model 22 is stored in the memory in step S6, and then the language model generation processing ends. Since the algorithm of this language model generation processing does not perform a part-of-speech or a connection between a part-of-speech and a word, a statistical language model combining features of a part-of-speech bigram and a variable-length word N-gram at the time of completion of generation. 22.

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

The word dictionary 12 connected to the word collating 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, a uttered voice of a speaker is input to a microphone 1 and converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order 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 collating unit 4 uses the one-pass Viterbi decoding method and the word hypothesis using the phoneme HMM 11 and the word dictionary 12 based on the characteristic parameter data input via the buffer memory 3. Is detected, 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 matching 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, for example, a frame number) from the utterance start time. .

The word hypothesis narrowing section 6 refers to the statistical language model 22 based on the word hypothesis output from the word matching section 4 via the buffer memory 5 and has the same end time and the same start time. Is represented by one word hypothesis having the highest likelihood among the total likelihoods 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 so as to cause them, the word string of the hypothesis having the maximum total likelihood is output as the recognition result among the word strings of all the narrowed word hypotheses. In the present embodiment, preferably, the first phoneme environment of the word to be processed is three phonemes including the last phoneme of the word hypothesis preceding the word and the first two phonemes of the word hypothesis of the word. I mean 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, and Wf are assumed as the word hypotheses of the word Wi _-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 /. The hypothesis with the highest total likelihood among the hypotheses in which the end time t _e is equal to the first phoneme environment (the top three word hypotheses in which the _first phoneme environment is “x / a ₁ / a ₂ ” in FIG. 2) (for example, FIG. 2
Are deleted except for the top hypothesis). Note that 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 head phoneme environment of the word is defined as a sequence of three phonemes including the last 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. The phoneme sequence of the preceding word hypothesis including the final phoneme of the preceding word hypothesis, and at least one phoneme of the preceding word hypothesis that is continuous with the final phoneme, And a phoneme sequence that includes a phoneme sequence that includes the first phoneme of the word hypothesis.

In the above embodiment, the feature extraction unit 2
The word matching unit 4, the word hypothesis narrowing unit 6, and the language model generation unit 20 are constituted by, for example, a digital computer. The buffer memories 3 and 5 are constituted by, for example, hard disk memories. 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, 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 matching 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. You may.

[0055]

EXAMPLE The present inventor compared the perplexity and the number of parameters with a conventional word N-gram in order to confirm the performance of the statistical language model 22 used in the present embodiment. The data used in the experiment is a naturally spoken travel conversation database owned by the present applicant (refer to the conventional document 8 “Mori”).
Moto et al., “A Speech and Lang
age Database for Speech
Translation Research ”, ICS
LP, pp 1791-1794, 1994. " )
846 dialogues, consisting of 354,700 words. Among them, 828 dialogues, 347,299 words are used as learning text data (also called a training set) for generating the statistical language model 22, and the remaining data is used as test text data (also called a test set). ). In the statistical language model 22 according to the present embodiment, the initial class is set to 80 parts of speech including the inflected form, separated up to 1000 parts, and data is collected every 100 parts. In addition to the statistical language model 22 according to the present embodiment and the word N-gram, as a measure against unknown word transition, the class and the transition probability of the word are interpolated by a deletion interpolation method (see conventional literature 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, the entropy H when there is a long word string _{^{w 1 n = w 1 w 2}} ... w n comprising a plurality of n words
(N) is represented by the following equation.

[0056]

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

FIG. 7 shows how the value of the test set perplexity changes in the experimental results. As is clear from FIG. 7, as the number of separation classes increases, the test set perplexity decreases, and the number of separation classes decreases by two.
It can be seen that the value of 00 is a word bigram and the number of separation classes is 600, which is almost the same value as the word trigram. When the number of separation classes exceeds 600, 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 with the separation classes of 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 greater than a word bigram and a word trigram.

Table 1 also shows that the number of separated classes is 0,500,
The value of the 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 ─数 Number of parameters (1) 4.10 × 10 ⁷ 2.62 × 10 ¹¹ 1.28 × 10 ⁴ 3.43 × 10 ⁵ 1.17 × 10 ⁶数 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 (word) transitions actually existing in the training set. I do. From Table 1, the statistical language model 22 of the present embodiment has a difference in perplexity between the test set and the training set.
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 classes are separated, less than a word bigram,
Much less than a 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 conventional word bigrams and word trigrams.

The statistical language model 22 of the present embodiment
FIG. 8 shows the result of examining the change in the value of the test set perplexity by changing the number of learning words to confirm the reliability of the test set. As is clear from FIG. 8, when all the 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 indicate the separation classes) Are the same perplexity values, but even if the number of learning words is reduced, the statistical language model 2
It can be seen that the increase in perplexity of 2 is relatively small and has a lower value than the word bigram. Similarly, even when the word trigram is compared with the statistical language model 22 (600 classes), when the number of learning words decreases, the statistical language model 22 exhibits lower perplexity.

Next, the inventor applied the statistical language model 22 of the present 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 使用 Parameters used 16th order LPC cepstrum + 16th order ΔCepstrum + log power + Δlog power ─────────────────────────────────── Acoustic model Unspecified gender of HM network Speaker model 400 states, 5 mixed ───────────────────────────────────

In Table 2, with regard to the gender-specific unspecified speaker model of the HM network, see Reference 9 "Kosaka et al.," Unspecified speaker speech recognition using mixed speaker SSS ", Proceedings of the Acoustical Society of Japan, 2- 5-9, pp 135-136, 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 conversations in the training set of the statistical language model 22, and the test set not used for learning is 18 conversations. Table 3 shows the correct word content rates of the sentence recognition candidates having the 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 obtained by registering only words included in the recognition target in the dictionary (750).
Word), 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, the recognition of the dictionary of the full dictionary was not performed in this experiment because of the memory capacity and the calculation time. In this case, in other words, a large-capacity memory and a huge processing time are required.

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

As described above, the accuracy of the N-gram
A variable-length N-gram statistical language model 22 generation apparatus for improving reliability and a continuous speech recognition apparatus using the same can be realized. Statistical language model 2
2 is generated by performing two types of state separation, that is, separating the part of speech bigram from the part of speech class and combining the concatenated words with the part of speech bigram as an initial state. It is a model that combines features. As a result of the evaluation experiment of the statistical language model 22, it was found that the statistical language model 22 can realize a perplexity equal to or larger than a word bigram and a word trigram with far fewer parameters. It was confirmed that it was realized. In addition, as a result of applying the method to continuous speech recognition, it was possible to obtain a correct word content rate equivalent to that of a 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.

Therefore, 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 to provide the statistical language model 22 , A continuous speech recognition device capable of continuously performing speech recognition at a higher speech recognition rate can be provided.

[0073]

As described above in detail, according to the statistical language model generating apparatus according to the first aspect of the present invention, based on the learning text data in which the uttered voice sentence of a predetermined speaker is written, Means for classifying the vocabulary of words into part-of-speech classes clustered for each part-of-speech, and generating a bigram between the part-of-speech classes as an initial state statistical language model; and an initial state statistical language generated by the generator. A first separation class candidate that can be separated from a part of speech class of a word based on a model, a combination of one word and one word, a combination of one word and a word string of a plurality of words, From the part of speech class of a word by combining a word string of a word with one word, a word string of a plurality of words and a word string of a plurality of words. Search means for searching for a second separation class candidate that can be performed, and a predetermined entropy indicating the difficulty of predicting the next word for the first and second separation class candidates searched for by the search means. Calculating means for calculating the amount of decrease in the entropy by separating the classes, and the largest class of the amount of decrease in entropy for the first and second separation class candidates calculated by the calculating means. Separating means for selecting a separation and performing separation of the selected class to generate a statistical language model including a part-of-speech bigram and an N-gram of a variable-length N word; Until the number of classes of the statistical language model thus obtained reaches a predetermined number of classes, the statistical language model generated by the separating means as a processing target model, 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, a statistical language model that can improve the prediction accuracy and reliability of the transition probability can be generated. Also, since the statistical language model can be realized with few parameters,
It has the unique advantage that it can be easily extended to large vocabulary speech recognition.

According to a second aspect of the present invention, there is provided a speech recognition apparatus comprising a speech recognition means for recognizing a speech by using a predetermined statistical language model based on a speech signal of an input uttered speech sentence. 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 having a variable length. 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 a speech recognition device that can perform speech recognition at a higher speech recognition rate.

In the speech recognition apparatus according to the third aspect, the statistical language model is generated by the statistical language model generation apparatus according to the first aspect. 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 a speech recognition device that can perform speech recognition at a higher speech recognition rate.

A continuous speech recognition apparatus according to a fourth aspect of the present invention detects a word hypothesis of the uttered speech sentence based on an input speech signal of the uttered speech sentence and calculates a likelihood to thereby obtain a continuous speech. In a continuous speech recognition apparatus provided with a speech recognition means for recognizing a speech, the speech recognition means comprises:
With reference to the statistical language model generated by the described statistical language model generation device, for each word hypothesis of the same word having the same end time and different start time, the utterance starts for each head phoneme environment of the word. The word hypotheses are narrowed down so as to be represented by one word hypothesis having the highest likelihood among the total likelihoods calculated from the time to the end time of the word. That is, as compared with the prior art word pair approximation method in which each preceding word is represented by one word hypothesis, those in which the leading phoneme of the word (that is, the last phoneme of the preceding word) is equal are collectively handled. Therefore, 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 compared to 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. Furthermore, since speech recognition is performed using a statistical language model that can improve the prediction accuracy and reliability of transition probability, a continuous speech recognition device that can continuously perform speech recognition at 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 bigram statistical language model.

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

FIG. 5 is a state transition diagram showing a model under a variable length N-gram used in the continuous speech recognition device of FIG. 1;

FIG. 6 is a flowchart illustrating a language model generation process executed by the language model generation unit 20 of FIG. 1;

FIG. 7 is a graph showing test set perplexity with respect to the number of separated classes in a statistical language model generated by the language model generation unit 20 of FIG. 1;

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 ... Learning text data, 22 ... Statistical language model.

────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Shoichi Matsunaga 5th place Sanraya, Daiyaku small character, Seika-cho, Soraku-gun, Kyoto A.T. 108704 (JP, A) JP-A-5-250405 (JP, A) Proceedings of the Acoustical Society of Japan (March 1996) 1-P-17, p. 195-196 (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00 535 G10L 3/00 561 JICST file (JOIS)

Claims (4)

(57) [Claims] 1. Classifying all vocabulary into part-of-speech classes clustered for each part-of-speech based on learning text data in which uttered voice sentences of a predetermined speaker have been written, and setting a bigram between these part-of-speech classes in an initial state A first separation class candidate capable of separating words from a part of speech class based on the initial state statistical language model generated by the generation unit; Combination of one word and one word, combination of one word and multiple word strings, combination of multiple word word strings and one word, multiple word word strings and multiple word A search unit for searching for a connected word including a combination with a word string or a second separation class candidate that can be separated from a word class based on the combination of the connected word strings; Calculating means for calculating, using the predetermined entropy representing the difficulty level of the prediction of the next word, for the first and second separation class candidates retrieved by the search, the amount of reduction of the entropy by separating the classes Selecting the largest class separation among the reduced amounts of entropy for the first and second separation class candidates calculated by the calculation means, and executing the separation of the selected class, whereby the bigram of the part of speech is obtained. 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. Using the statistical language model generated by the above as a processing target model, and repeating the processing of the search means, the processing of the calculation means, and the processing of the separation means And a control unit for generating a statistical language model having a predetermined number of classes. 2. A speech recognition apparatus comprising speech recognition means for recognizing speech based on a speech signal of an input uttered speech sentence using a predetermined statistical language model, wherein the speech recognition means comprises a part-of-speech bigram. A speech recognition apparatus characterized in that speech recognition is performed using a statistical language model including N-grams of words of variable length N. 3. A speech recognition device, wherein the statistical language model is generated by the statistical language model generation device according to claim 1. 4. A continuous speech recognition device comprising a speech recognition means for continuously recognizing speech by detecting a word hypothesis of the speech speech sentence based on an input speech signal of the speech speech and calculating a likelihood. In the apparatus, the speech recognition unit refers to the statistical language model generated by the statistical language model generating apparatus according to claim 1 and, for a word hypothesis of the same word having equal end times and different start times. For each leading phoneme environment of the word, the word hypothesis is narrowed down so as to be represented by one word hypothesis having the highest likelihood among the total likelihoods calculated from the utterance start time to the end time of the word. A continuous speech recognition apparatus characterized by performing.