JP3415585B2 - Statistical language model generation device, speech recognition device, and information retrieval processing 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 generation device for generating a statistical language model based on learning data and learning text data, and an uttered voice sentence input using the statistical language model. Recognition device for recognizing a voice signal of a user, and an information retrieval processing device using the voice recognition device Regarding

[0002]

2. Description of the Related Art In recent years, with the progress of voice recognition technology, the application of voice recognition to large vocabulary tasks has been actively performed. However, the paradigm of large vocabulary speech recognition (meaning the method of grasping the dominant scientific object in a specific area or era) does not completely solve the problem of unregistered words.
In particular, with regard to proper names such as personal names, there is an essential problem that it is difficult to cover all of them. On the other hand, many proper nouns are important information for accomplishing the task, and the unregistered word processing technology of proper nouns is an important issue when considering the operation of voice recognition on a real task.

Conventionally, the following method has been proposed as a method of detecting unregistered words including phoneme sequences (reading) in a continuous speech recognition apparatus. (1) A method of using a subword decoder such as a phoneme typewriter together (hereinafter, referred to as a method of a first conventional example), and (2) A method of incorporating a subword into a language model as a pseudo word (hereinafter, referred to as a second method). It is called the conventional method.).

[0004]

However, the method of the first conventional example is not desirable from the viewpoint of processing amount because it is necessary to drive another decoder. In addition, since the maximum likelihood phoneme sequence score is used as the acoustic score of the estimated unknown word section, heuristics (heuristics) such as penalties and thresholds are involved in the integration with the in-vocabulary word sequence hypothesis.

On the other hand, the method of the second conventional example has an advantage that it can be realized without changing the decoder. However, in order to perform effective language processing on unregistered words obtained as a subword sequence, morphological analysis using a vocabulary larger than the recognized vocabulary is required as post-processing. Moreover, it is unlikely that a model that sufficiently reflects linguistic characteristics can be modeled by the word and the subword, or the N-gram probability between the subwords, and the effectiveness as a recognition constraint remains questionable.

Further, in a small-scale information search device such as a voice recognition and automatic dialing function in a telephone or a car navigation, the number of dictionary registrations is limited, so that the number of proper proper nouns is limited. In such a case, when the proper noun of the object is managed by a system different from the voice recognition device, it cannot be registered in the voice recognition device and the voice recognition rate cannot be improved.

The object of the present invention is to solve the above problems and to improve the accuracy of voice recognition for unregistered unregistered words in a word dictionary as compared with the conventional example. (EN) Provided are a statistical language model generation device capable of generating a statistical language model for identifying and a speech recognition device using the statistical language model generation device.

Another object of the present invention is to provide a voice recognition and automatic dialing function in a telephone and a small-scale information retrieval processing device such as a car navigation system for voice recognition of unregistered words in a word dictionary. An object of the present invention is to provide an information search processing device capable of executing information search by using a voice recognition device whose accuracy is higher than that of the conventional example.

[0009]

[0010]

[0011]

[0012]

[0013]

[0014]

Statistical language model generating apparatus according to claim 1, wherein according to the present invention SUMMARY OF THE INVENTION comprises a learning data storage means for storing the learning data including a word list of common nouns proper nouns or foreign language, Based on the learning data stored in the learning data storage means, the parameter of the gamma distribution of the mora length when the ratio of the number of words to the mora length in the learning data substantially assumes the gamma distribution depends on the class. It is not possible to estimate and calculate the first N-gram that has a class that is a subclass of the proper noun or the common noun of the foreign word, in units of subwords that are mora or mora chain. Based on a first generation unit that generates a subword unit N-gram model that models a registered word, and a predetermined text database The generated word class N-gram
The word class and the uniqueness based on the model, the sub-word unit N-gram model generated by the first generating unit, and the gamma distribution parameter of the mora length calculated by the first generating unit. Second, a statistical language model including unregistered words based on subword units is generated by calculating the probability of occurrence of a second N-gram depending on a class that is a subclass of a noun or a common noun of a foreign word. And a means for generating

Further, the statistical language model generating apparatus according to claim 2, in a statistical language model generating apparatus according to claim 1, based on the subword unit N-gram model generated by said first generating means A third extraction means for extracting the sub-word unit and assigning the extracted label to the sub-word unit to generate a plurality of labeled sub-word unit data per sub-word unit; and the word extracted from the text database. And fourth generating means for generating a word dictionary by adding a phoneme arrangement to the plurality of labeled subword unit data generated by the third generating means. To do.

Furthermore, a voice recognition device according to a third aspect of the present invention comprises a voice recognition means for recognizing a voice by using a predetermined statistical language model based on a voice signal of an input uttered voice sentence. In the voice recognition device described above, the voice recognition means is a statistical language model generated by the statistical language model generation device according to claim 1 or 2, and
It is characterized in that voice recognition is performed using the word dictionary generated by the fourth generating means described in 2 .

According to a fourth aspect of the present invention, there is provided an information retrieval processing device, which comprises a database storage means for storing a database including word data of common nouns corresponding to the word list and information corresponding to the data. Using the character string of the voice recognition result output from the voice recognition device described in No. 3 as a key, the database stored in the database storage means is searched, and the information corresponding to the matching word data is retrieved from the database storage means. And a retrieval means for reading and outputting.

Further, the information search processing device according to a fifth aspect is the information search processing device according to the fourth aspect , further comprising processing execution means for executing a predetermined process based on the information output from the search means. It is characterized by having.

[0019]

[0020]

DETAILED DESCRIPTION OF THE INVENTION Embodiments of the present invention will be described below with reference to the drawings.

<First Embodiment> FIG. 1 is a block diagram of a continuous speech recognition system according to a first embodiment of the present invention. The continuous speech recognition system according to the first embodiment of the present invention includes an unregistered word model generation unit 20, a subword unit data generation unit 21, a word dictionary generation unit 22, and a word class N-gram model generation unit 23. And a statistical language model generation device including the language model generation unit 24.

The present embodiment discloses a statistical language model generation device for generating a new statistical language model, which enables highly accurate recognition of speech including unregistered words. The statistical language model of this embodiment is (1) learning data memory 30.
Subword unit N-gram model and mora length gamma distribution data, which are a plurality of unregistered word models for unregistered word recognition, which are generated by the unregistered word model generation unit 20 based on the learning data in ) The word class Ng based on the text data in the text data memory 31.
The language model generation unit 2 based on the word class N-gram model generated by the ram model generation unit 23.
4 produces a statistical language model.

These unregistered word models are constructed depending on each vocabulary class. Here, the subword means a unit smaller than a word, and specifically means a mora or a mora chain. A mora is a relative length of a sound that is a unit of stress or intonation in prosody theory, and one mora corresponds to one syllable including a short vowel. In Japanese, this corresponds to almost one "kana" character (two characters in Japanese syllabary). In the following, the description is limited to unregistered words of Japanese surnames and surnames, which are subclasses of proper nouns.

The present inventor has (1) about 3 as shown in Table 1.
The Japanese surname file 30a containing the data of the mora arrangement (reading) of the 0,000 Japanese surnames and (2) the data of the mora arrangement (reading) of the approximately 300,000 Japanese first names as shown in Table 2. The learning data including the Japanese name file 30b including the learning data is stored in the learning data memory 30.

[0025]

[Table 1] ―――――――――――― Suzuki Ta, Ka, Ha, Shi Sa, Lee, To, Oh Ta, na, mosquito Hi, La, Tsu, Ji A, sa, gi, no ...... ――――――――――――

[0026]

[Table 2] ――――――――――――――――――――――――――――――――――― Yo, Oh, Ko Tomo, D To, mo, ko Ke, Lee, Ko   ...... ―――――――――――――――――――――――――――――――――――

The analysis and the analysis result of the Japanese surname and first name data by the present inventor based on the above learning data will be described. When viewing Japanese surnames and given names as a series of subwords, it is easily expected to have the following characteristics. (1) Features concerning length: Suzuki, Sato, Takahashi, etc. are the surnames, and Hiroshi, Akira, Ichiro, etc. are the names with 3 to 4 moras. (2) Features related to phoneme alignment: Japanese surnames and given names are basically composed of kanji, and there are high-frequency units such as Yama, Mura, and Naka for surnames, and Rou, Ichi, and Hiro for names. To do.

From this viewpoint, the present inventor analyzed the statistical characteristics of reading Japanese surnames and given names. As the personal name data, a publicly known personal name list in which names of approximately 300,000 famous persons are collected was used. From this learning data, we extract the surname and given name consisting of only kanji and hiragana as Japanese names,
The obtained 303,552 first names and 295,148 first names were analyzed, and the results are shown in Table 3. For comparison, we also analyze the characteristics of words other than Japanese surnames and given names. As learning data for comparison, a total of 1,155,183 words excluding Japanese surnames and given names from the natural speech travel conversation database owned by the applicant of the present patent application were used.

[0029]

[Table 3] Model learning data ―――――――――――――――――――――――――――――――――――― Japanese First Name First Name Last Name Travel Conversation― ―――――――――――――――――――――――――――――――――― Total number of words 303,552 295,148 1,161,576 Different vocabulary 19 , 018 20,413 13,453 ――――――――――――――――――――――――――――――――――― (Note) Different Japanese names The vocabulary was evaluated by words with different phoneme sequences or different readings, and differences in kanji notation were ignored.

FIG. 6 shows statistics about the length of words analyzed by the present inventor. The Mora number was used as the unit of length. From this result, it can be confirmed that the Japanese surnames and given names have a distribution that is extremely biased around 3, 4 mora.
If 3 and 4 moras are combined, almost 90% of the names correspond to both surname and given name. Next, Table 4 shows statistics on the arrangement of mora. As an index of the bias of mora alignment,
The coverage of the mora array by the two most-frequent mora two chains was investigated. Here, the coverage means a ratio of two chains in all mora.

[0031]

[Table 4] Bias of mora arrangement ―――――――――――――――――――――――――――――――――――― Two-chain mora types Mora arrangement Coverage rate (%) (top N types by frequency) Japanese first name First name First name Travel conversation ―――――――――――――――――――――――――――――――― ――― 1 3.8 4.9 0.1 10 23.3 28.3 5.1 100 100 59.8 66.6 19.4 1000 84.3 82.4 35.6 ――――――― ―――――――――――――――――――――――――――― (Note) Based on the most frequent N types of two-chain mora. Coverage (%) of mora. Since there are words of odd length, the coverage will not be 100%.

For example, for Japanese surnames and given names, 80% or more of the mora arrangements in the surnames and given names are covered only by the high-frequency 1000 kinds of mora two chains.

Next, a method of generating a statistical language model based on unregistered word models of Japanese surnames and given names will be described in detail. Based on the knowledge obtained above, a statistical language model based on the unregistered word model of Japanese surname and first name class is constructed. Further, from the viewpoint of decoding, it is desirable that the statistical language model can be handled in the N-gram format which has been widely used in recent years, and in this embodiment, the unregistered word model is implemented in the word N-gram format.

First, modeling of a word series including unregistered words will be described. The basis of the statistical language model according to this embodiment is the class N-gram model of words. The word class N-gram model is generated by the word class N-gram model generation unit 23 based on text data called a corpus composed of a large number of Japanese written sentences stored in the text database memory 31, for example, a known maximum likelihood. Word clustering is performed using the estimation method to generate a word class N-gram model, and this model is stored in the word class N-gram model memory 43.

In the word class N-gram model, the linguistic likelihood p of the word sequence W = w ₁ , w ₂ , w ₃ , w ₄ , ..., W _T.
h (W) is generally given by the following equation. However, w _t is the t-th word of the word sequence W, and c ^wt represents the vocabulary class of the word w _t .

[0036]

[Equation 1]

By the way, the word w _t (hereinafter, one word is represented by w) includes an unregistered word which is not in the recognition vocabulary. The occurrence probabilities of these unregistered words are arranged in phonemes (or read).
When estimated on the basis of the statistical characteristics of, the in-class word 1-gram probability p (w | c ^w ) in the above equation 1 is given by the following equation. However, M ^w represents the mora sequence of the word w.

[0038]

(2) (1) If w ∈ word dictionary, p (w | c ^w ) ≡ (1-p (OOV | c ^w )) · p (w |
c ^w , inVoc), (2) If other than the above, p (w | c ^w ) ≡p (OOV | c ^w ) · p (M ^w | c ^w , O
OV)

Here, the word dictionary is the word dictionary memory 12
Is a vocabulary dictionary stored in, OOV is an unregistered word, and inVoc represents a condition in the word dictionary. In the above equation 2, the probability p (OOV | c ^w ) is the probability that an unregistered word will occur from the class, and for example, a known method (for example,
Prior art document "Masaki Taki et al.," Morphological analysis using complex N-gram of part-of-speech and variable length morpheme sequence ", Journal of the Language Processing Society" Natural Language Processing ", Vol. 6, No. 2, pp.
41-57, 1999 ". ) Can be estimated.
This estimation method will be described. When a known Turing estimation method is used, a morpheme that appears r times in the data is estimated to be r ^* times in the following equation.

[0040]

## EQU3 ## r ^* = {(r + 1) n _{r + 1} } / n _r

Here, n _r represents the number of types of morphemes that appear r times in the data. Therefore, the appearance probability P (w | c _ξ ) from the part of speech of the morpheme w that appears r times is represented by the following equation.

[0042]

## EQU4 ## P (w | c _ξ ) = r ^* / N (c _ξ )

The appearance probability P (w | c _ξ ) is given as a class c
All the morphemes belonging to _ξ are calculated, and the remainder subtracted from 1 is the probability P (c _{h ξ} ) of the unknown word appearing from the part of speech c _ξ as shown in the following expression.

[0044]

[Equation 5]

In the present embodiment, focusing on effective evaluation of unregistered word models on a limited evaluation set,
Modeling is performed using the following formula. That is, the occurrence of unregistered words should be allowed only to some predefined classes (herein, a set of classes is called C ^OOV ), and all the word occurrences from these classes should be explained by the unregistered word model. And do not make a registered word.

[0046]

[Equation 6]

[Equation 7]

Next, the unregistered word model of Japanese surnames and given names will be described. As described above, the reading of Japanese surnames and surnames showed characteristic tendencies for the mora length and mora arrangement. Therefore, the unregistered word model p (M ^w | c ^w ) of the above equation 7 can be modeled with high accuracy by expanding as in the following equation. However, le
n (M ^w ) represents the mora length of the word w.

[0048]

## EQU8 ## p (M ^w | c ^w ) = p (len (M ^w ) | c ^w ).
p (M ^w | c ^w , len (M ^w ))

The probability p (len (M) | c) of the above equation 8
Is the length len (M) in the Japanese surname or surname class
Is the probability that the word will occur. In this embodiment, it is assumed that the probability distribution substantially follows the gamma distribution given by the following equation. That is, the right side of the above equation 8 is a product of the probability of the gamma distribution of the mora length of the first term and the probability of the subword unit bigram of the second term. However,
α and λ are parameters depending on the class c, and are determined by the average and variance of the mora length.

[0050]

[Equation 9] here,

[Equation 10]

On the other hand, the probability p (M ^w | c ^w , le of the above equation 8 is
n (M ^w )) has length len (M ^w ) in class c ^w
Is a probability that the mora arrangement of M ^w = m ₁ ^w , m ₂ ^w , ... Is modeled by the subword unit N-gram of the following equation. However, U = u ₁ , u ₂ , ... Is a subword unit (mora or mora chain) automatically acquired by the method described later in detail.
It is a series of. In addition, subword unit N-gra in the formula
The transition to the terminal symbol is not included in m.

[0052]

[Equation 11]

In the statistical language model according to this embodiment described above, “... Asagino Yoko and
． ． ． 7 shows an example in which "" is output. In the example, the mora series “Asagino (day surname)” and “Yoko (day name)” with class labels are output for the words “Asagino” and “Yoko” in the Japanese surname and given name class. In this model, linguistic restrictions are imposed on the occurrence of Japanese surnames and given names from the following three levels.

<Three-Level Linguistic Constraints> (1) Interword Constraints: Using a class N-gram of words,
Evaluate the likelihood of Japanese surnames and given names (classes) occurring in a word context. The modeling of surnames and given names by subwords is hidden in the lower hierarchy, and therefore does not adversely affect the modeling of registered word sequences. (2) Continuation length constraint of family name and first name section: The family name and first name likeness of the section are evaluated using the gamma distribution regarding the mora length of each family name and first name. With this restriction, it is possible to prevent the emergence of an unduly short or long mora sequence. (3) Subword arrangement constraint: A subword unit N-gram of a family name and a given name in units of a mora and a mora chain is used. By using the mora chain as a unit, higher accuracy of N-gram can be expected. Here, the mora chain as a modeling unit is automatically acquired in the iterative learning described later.

Next, an unregistered word model learning generation process executed by the unregistered word model generation unit 20 will be described. The unregistered word model generation unit 20, based on the Japanese surname file 30a and the Japanese name file 30b stored in the learning data memory 30, the Japanese surname class (abbreviated as Japanese surname in the label) and Japanese. An unregistered word model of the first name class (abbreviated as day name in the label) is constructed. In this learning generation process, specifically, subword unit N-gram model and mora length gamma distribution data are generated. In the following embodiments, it is assumed that individual names appear with equal probability, and the same surname or the number of same names in the personal name list is used as the observation frequency of each surname or each name. Only a single mora is given to the sub-word unit N-gram as an initial unit set, and new mora chains are sequentially added to the unit set in iterative learning described later. The mora chain, which is a unit candidate, is pre-selected according to frequency to improve learning efficiency.

FIG. 3 shows the unregistered word model generator 20 of FIG.
FIG. 4 is a flowchart showing an unregistered word model generation process executed by FIG. 4, and FIG. 4 is a subword 2-gram unit determination process (step S in FIG. 3) which is a subroutine of FIG.
It is a flow chart which shows 4).

As the learning data of the unregistered word model, a list of surnames and first names expressed in the Mora series is used, and the list data are stored in the Japanese surname file 30a and the Japanese name file 30b in the learning data memory 30, respectively. ing. This learning data is as shown in Tables 1 and 2 above, and each mora is described by being separated by a comma ",". Based on this learning data, an unregistered word model, that is, gamma distribution of mora length and subword unit N-gr
Generate an am model. In the following, a procedure for generating an unregistered word model will be described for the case where the degree N of N-gram is 2.

In step S1 of FIG. 3, the learning data is read from the learning data memory 30, and in step S2, the family name or first name is set to 1 based on the read learning data.
After calculating the average μ and the variance V of the number of mora per word,
The parameters of the gamma distribution of the mora length are estimated using the following formula.

[0059]

(12) λ = V / μ

[Formula 13] α = μ ² / V

Further, in step S3, a high-frequency mora chain, which is a unit candidate of the subword unit 2-gram, is extracted based on the learning data, and the extracted unit candidate is called a "unit candidate set". Here, the frequency of all mora chains appearing on the learning data, for example, all the mora chains having a length of 2 or more is examined, and the frequency is a predetermined value (= 10).
0) The above mora chains are extracted as unit candidates. Next, in step S4, the unit determination process of the sub word 2-gram, which is the subroutine shown in FIG. 4, is executed,
Finally, in step S5, the provisional unregistered word model after the learning is stored in the subword unit model N-gram model memory 40 as a subword unit N-gram model, and the mora length gamma distribution data is stored. The data is stored in the data memory 41.

In the unit determination process for the subword 2-gram in FIG. 4, first, at step S11, all single moras are inserted into the fixed unit set, and then step S11.
At 12, one mora chain is selected from the unit candidate set. Next, it is determined whether or not the mora chain selected in step S13 is included in the fixed unit set. If YES, the process proceeds to step S15, while NO
If so, the process proceeds to step S14. Add the mora chain selected in step S14 to the fixed unit set,
The provisional subword 2-gram using the known maximum likelihood estimation method
A model is generated and the process proceeds to step S15. Where 2-
The gram model is one between the training data and the additional set.
-Gram and 0-gram of only the additional set are used to perform a known deletion interpolation method (for example, the prior art document "Seiichi Nakagawa," Speech recognition by stochastic model ", edited by the Institute of Electronics, Information and Communication Engineers, pp. 63). -64, issued July 1, 1988, etc.). This provisional subword 2-gram and gamma distribution data of mora length are collectively referred to as a "provisional unregistered word model". Step S
At 15, it is determined whether or not the steps S13 and S14 have been performed for all the mora chains, and NO.
If so, the process returns to step S12 to repeat the above process, but if YES, the process proceeds to step S16. In step S16, the average likelihood is calculated for each of the temporary unregistered word models using Equation 8, and the unit set of the temporary unregistered word model that maximizes the average likelihood is set as a new definite unit set. Then, in step S17, it is determined whether or not the number of mora chains included in the definite unit set ≧ a predetermined threshold value N _th (for example, 150). If NO, the process returns to step S12 and the above-mentioned is executed. While repeating the processing, when YES is returned to the original main routine.

FIG. 9 shows the unregistered word model generator 20 of FIG.
In the unregistered word model generation process executed by
It is a graph which shows improvement of average likelihood by unitization of a mora chain, and is a graph which shows average likelihood with respect to the number of kinds of mora chain. That is, FIG. 9 shows changes in the average likelihood (Equation 8) in iterative learning. The mora chain as a unit candidate has a frequency of 100 or more. From the learning data shown in Tables 1 to 3, 1,829 kinds of surname models and 1,660 kinds of surname models are mora chains as unit candidates. The subword unit N-gram is N = 2, and 1
Interpolation was performed by a known deletion interpolation method using -gram and 0-gram. As shown in FIG. 9, by adding the mora chain in subword units, the average likelihood for the learning data monotonically increases. The average likelihood of the model in which 150 mora chains were added was 3.9 times in the surname model and 3.2 times in the surname model as compared to the model without the mora chain. If we consider the unregistered word model as the word 1-gram, the learning set perplexity of the word is 74% for the surname model,
It will be improved by 69% with the famous model.

Next, the unregistered word model is set to the word N-gra.
A method of implementing the m-type word dictionary will be described. The unregistered word model described above can be handled in the class N-gram format without approximation by the method described below. Therefore, as a statistical language model, class N-
Any decoder that can handle gram can recognize unregistered words by this method without changing the decoder. However, extremely long unregistered words (in this embodiment,
The condition is that family names and surnames of 10 mora or more) are not recognized. The mora and mora chain used as a unit in the sub-word unit N-gram are treated as a pseudo word and are incorporated in the recognition dictionary and the class N-gram.
At this time, each subword unit is expanded by the following labeling to generate a plurality of identical subword units having different labels.

That is, the labeled sub-word unit data generating section 20 uses the sub-word unit N-gram memory 4
Subword unit N- of Japanese surname and given name stored in 2
The following processing is executed for each gram, and the resulting data of the labeled subword unit group is stored in the labeled subword unit data memory 40. That is, all subword units (single mora, mora chain) used as units in the Japanese surname (or Japanese name) subword unit N-gram are extracted. Then
A plurality of labels are generated according to each extracted subword unit. By assigning each generated label to the corresponding subword unit, a plurality of labeled subword units are generated per subword unit. Here, the label means (a) class (in the present embodiment, “day family name” or “day name”), and (b) start position (1, 2, ...) In the word.
LenMax + 1; number of mora in subwords: Here, LenMax is a preset value, and means the number of mora of the longest family name or given name to be recognized. ), And (c) whether or not it is the end of a word (“end” or “−” indicating the end).

The expansion by the starting mora position in the above (b) is performed so that the ending position is up to 9 mora in accordance with the longest surname and given name appearing in the learning data. In the subword unit to which the word end label is added in (c) above, a pause is allowed at the end of the phoneme sequence (reading). Here, the labeled subword is subject to the following restrictions on its transition. (I) The transition from the registered word class to the labeled subword is allowed only when the starting mora position of the labeled subword is 1. On the contrary, (ii) the transition from the labeled subword to the class of the registered word is allowed only when the word end label is given to the labeled subword. (Iii) Transitions between labeled subwords are allowed only when the mora positions within a word are concatenated and belong to the same class.

Further, the word dictionary generator 22 generates a word dictionary as follows based on the data in the labeled subword unit data memory 42 and the text data in the text data memory 31, and the word dictionary memory 12 To store. First, the word dictionary generation unit 22 extracts all the words that appear in the text data in the text database memory 31 and stores them in the word dictionary memory 12. Then, all the labeled subword units in the labeled subword unit data memory 42 are stored in the word dictionary memory 12
To store. For all the entries stored in the word dictionary memory 12, the reading (phoneme sequence) is performed manually or by using a known phoneme assignment processing program that assigns phonemes using a correspondence table between the reading and the phoneme sequence. ) Is added to generate a word dictionary. As a result, the word dictionary is generated and stored in the word dictionary memory 12.

Next, the statistical language model generation process of the language model generation unit 24 will be described. When dealing with a label with the sub-word units u _d as a word, class N-gram
The probability at (N = 2) is given as follows (N>
2 also applies). However, u represents the subword unit before labeling, # represents the mora start position symbol in the word,
c represents a vocabulary class, len (u) is the mora length of the subword unit u, p _sw (u _j | u _i , c) is the subword unit 2-gram depending on the class, and p _LM (c ′ |
c) represents the transition probability between classes of class 2-gram.
In addition, classof (u _d), startof
_{(U d), isend (u} d) , respectively, is a function to refer vocabulary class is the label of the labeled sub-word units u _d, mora start position in the word, the word end whether the, endof (u _d ) is startof (u _d) + le
It is given by n (u). G _c (*) is a probability function based on the class mora length gamma distribution g _c (x) (Equation 9), and is given by the following definite integral.

[0068]

[Equation 14]

[Equation 15]

[0069] First of all, is the probability of occurrence of the first sub-word last name and the name of the class within a 1-gram probability p (u _d |
c) it is permitted only if the start mora position label labeled subword unit u _d is 1. Therefore, it can be expressed by the following equation.

[0070]

Equation 16] (1) If (classof (u _d) = c)
_{∧ (startof (u d) =} 1) ∧ (isend (u d) = false) p when _{(u d | c) = p} sw (u | #, c) · G c (x> len
(U)) (2) if _{(classof (u d) = c} ) ∧ (sta
_{rtof (u d) = 1)} ∧ (isend (u d) = true) when _{p (u d | c) =} p sw (u | #, c) · G c (x = len
(U)) (3) If other than the above, p (u _d | c) = 0

Next, the transition between labeled subword units, which is the first 2-gram probability between classes, is allowed only when both classes are the same and the mora positions in both words are concatenated. Therefore, 2-g between the first class
The ram probability p (u _dj | u _di ) can be expressed by the following equation.

[0072]

(17) (1) If (classof (u _dj ) = cl
assof (u _di ) = c) ∧ (startof (u _dj ) = endof (u _di ) +
1) When ∧ (isend (u _dj ) = false) p (u _dj | u _di ) = p _sw (u _j | u _i , c) · G _c (x>
(Endof (u _di ) + len (u _dj )) / G _c (x
> Endof (u _di )) (2) If (classof (u _dj ) = classof
(U _di ) = c) ∧ (startof (u _dj ) = endof (u _di ) +
1) When ∧ (isend (u _dj ) = true) p (u _dj | u _di ) = p _sw (u _j | u _i , c) · G _c (x =
(Endof (u _di ) + len (u _dj )) / G _c (x
> Endof (u _di )) (3) If other than the above, p (u _dj | u _di ) = 0

Furthermore, the transition from the labeled subword unit to the class of the next word, which is the second inter-class 2-gram probability, is allowed only when the word end label is assigned to the labeled subword unit. Thus, the second class between 2-gram probability p (c | u _d) is expressed by the following equation.

[0074]

Equation 18] (1) If iSEND (u _d) = true when _{p (c | u d) =} p LM (c | classof (u d)) (2) If the time other than the p (c | u _d ) = 0

FIG. 5 is a flow chart showing the language model generation processing executed by the language model generation unit 24 of FIG. In FIG. 5, first, in S21, each stored data is read from each of the memories 40, 41, and 42, and in Step S22, the probability function of the probability function is calculated using Equations 14 and 15 based on the data of the gamma distribution of the mora length. Calculate the value. Next, in step S23, the in-class-gram probability that is the occurrence probability of the leading subword is calculated using Equation 16, and in step S24, the first inter-class 2-gram that is the transition probability between subwords.
The m probability is calculated using Expression 17, and the second inter-class 2-gram probability, which is the probability of transition from the terminal subword to the class of the next word, is calculated using Expression 17 in step S25. Then, in step S26, the probabilities calculated above are put together and stored in the statistical language model memory 44 as a statistical language model based on the unregistered word model.

FIG. 8 shows another example of the statistical language model generated by the above language model generation processing. The transition probabilities in the example of FIG. 8 are as defined above.

Next, the structure and operation of the continuous speech recognition system shown in FIG. 1 will be described. In FIG. 1, a phoneme hidden Markov model (hereinafter,
Hidden Markov model is called HMM. The phoneme HMM in the memory 11 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) parameters of output probability density distribution, and (e) self-transition probability and transition to succeeding states probability. The phoneme HMM used in the present embodiment is generated by converting a predetermined speaker mixed HMM because it is necessary to specify which speaker each distribution originates from. Here, the output probability density function is a Gaussian mixture mixture having a 34-dimensional diagonal covariance matrix. Further, the word dictionary in the word dictionary memory 12 connected to the word matching unit 4 stores a symbol string indicating the reading expressed in symbols for each word of the phoneme HMM in the phoneme HMM memory 11.

In FIG. 1, the vocalized 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. The feature extraction unit 2 performs, for example, LPC analysis after A / D conversion of the input voice signal, and a 34-dimensional feature parameter including logarithmic power, 16th-order cepstrum coefficient, Δ logarithmic power, and 16th-order Δ cepstrum coefficient. To extract. The time series of the extracted characteristic 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 based on the data of the characteristic parameter input via the buffer memory 3, the phoneme HMM memory 1
Using the phoneme HMM in 1 and the word dictionary in the word dictionary memory 12, the word hypothesis is detected, the likelihood is calculated, and the likelihood is output.
Here, the word matching unit 4 calculates the likelihood within a word and the likelihood from the start of utterance for each state of each HMM at each time. The likelihood is individually held for each word identification number, word start time, and preceding word difference. Also, in order to reduce the amount of calculation processing,
Reduce the grid hypothesis of low likelihood out of the total likelihood calculated based on the phoneme HMM and the word dictionary. The word matching unit 4 uses the resulting word hypothesis and likelihood information as time information from the utterance start time (specifically, for example, a frame number).
Together with this, it outputs to the word hypothesis narrowing unit 6 via the buffer memory 5.

The word hypothesis narrowing unit 6 refers to the statistical language model in the statistical language model memory 44 on the basis of the word hypothesis output from the word matching unit 4 via the buffer memory 5, and determines the end time. For the same word with the same start time but different start times, the highest likelihood of the calculated total likelihood from the utterance start time to the end time of the word is calculated for each head phoneme environment of the word. After narrowing down the word hypotheses so that they are represented by one word hypothesis that they have,
Among the word strings of all the word hypotheses after narrowing down, the word string of the hypothesis having the maximum total likelihood is output as the recognition result. The task-adapted statistical language model has one statistical language model for each task, and the word hypothesis narrowing unit 6 selectively selects the statistical language model corresponding to the task to be speech-recognized. Refer to. 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.

[0081] For example, as shown in FIG. 2, the (i-1) th word W _i-1 of the following phoneme string a _1, a _2, ..., come i th word W _i consisting 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
The final phoneme of b and Wc is / x /, and the final phoneme of the latter three word hypotheses Wd, We, and Wf is / 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
Delete all but the first hypothesis). Since the fourth phoneme from the top has a different first phoneme environment, that is, 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 remains 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 embodiments, Japanese surnames and given names, which are subclasses of proper nouns, are used as vocabulary classes of unregistered words, but the present invention is not limited to this, and the following proper nouns are used. It can be applied to common nouns of foreign words. (1) First and last name of foreigner, (2) Place name, (3) Company name,
(4) Various facility names, (5) Various product names, etc. Therefore, in the present invention, each subclass of proper nouns and common nouns of foreign words can be used as the vocabulary class of unregistered words.

The class-dependent unregistered word model used in this embodiment can have a different parameter structure for each class. Therefore, modeling that strongly reflects the statistical characteristics of reading in each class is possible. In the embodiment, the unregistered word model construction of the Japanese surname and given name includes (1)
The gamma distribution of word length and (2) subword unit N-gram which does not include transition to terminal symbol were used. However, when modeling other classes, such as accommodation names (“Kyoto Daiichi Kanko Hotel”, “Akasaka Prince Hotel”, “Ito Ryokan”, ...) (1) There are cases where the restriction on the length is not valid. In that case, the constraint of the above (1) is omitted, and instead, the subword unit N-gram based on the constraint of the above (2) is modeled so as to include a transition to the terminal symbol (for example, “hotel” or “hotel”). It is possible to build a highly accurate unregistered word model for such a class by making a transition with a high probability from "inn" to a terminal symbol.

In the above embodiment, the leading phoneme environment of the word is a three phoneme sequence 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
, Word collation unit 4, word hypothesis narrowing unit 6, unregistered word model generation unit 20, and subword unit data generation unit 21.
The word dictionary generation unit 22, the word class N-gram model generation unit 23, and the language model generation unit 24 are configured by, for example, a computer such as a digital electronic computer, and include buffer memories 3 and 5 and a phoneme HMM memory. 11
A word dictionary memory 12, a learning data memory 30,
Text database memory 31 and subword unit N
-Gram model memory 40, mora length gamma distribution data memory 41, labeled subword unit data memory 42, word class N-gram model memory 43
The statistical language model memory 44 is composed of 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 for example, the phoneme HMM memory 11
Phoneme collating unit that refers to a phoneme HMM in the
It may be configured with a voice recognition unit that performs voice recognition of a word by referring to a statistical language model using the Pass DP algorithm.

[0087]

EXAMPLE The present inventor conducted a speech recognition experiment in order to confirm the effectiveness of the statistical language model according to this embodiment.
Below, a comparative evaluation is performed using two types of statistical language models. Both language models, as a common base model,
Class N-gr generated only from travel conversation data in Table 3
am is used. For this base model, in each method, Japanese surname class and first class word 1
-Replace gram.

The statistical language model used for evaluation is as follows. (1) Statistical language model according to the present embodiment: The unregistered word model of each family name and given name is used as the word 1-gram of the Japanese family name and given name class. Subword unit N-gr
As for the mora chain used as a unit in am, 150 cases are evaluated unless otherwise specified. The recognition vocabulary consists of 12,755 words plus subwords other than Japanese surnames and given names, and Japanese registered surnames and given names are not created. (2) Registered word method (hereinafter, referred to as a comparative example): The word 1-gram based on the personal name data in Table 3 is used as the word 1-gram of Japanese surname and given name class. The recognition vocabulary is 12,755 words other than Japanese surnames and given names + 39,431 Japanese surnames and given names. This method has a vocabulary that covers almost all the names in the evaluation set, and because the method of the present embodiment directly uses the person name data used for the maximum likelihood estimation of the unregistered word model as the word 1-gram, It is considered that the upper limit of the recognition accuracy according to the method of the present embodiment is approximately given.

The speech recognition rates of these two methods are evaluated according to the following criteria. (1) Word recognition rate: The recognition rate of all words appearing in the evaluation data is evaluated. For Japanese surnames and given names, the correct answer is only when the class (“day surname” or “day name”), reading (arranged by mora), and position (correspondence by DP) are all correct. However, with regard to reading, apparently equivalent long sounds (Yoko and Yoko) were manually corrected and evaluated. (2) Recall and precision of surnames and given words: The recall and precision of only Japanese surnames and given names are evaluated using dynamic programming matching (DP matching) when evaluating word recognition rates.

Here, 42 unilateral conversations 4,990 words in the travel conversation domain were used in the evaluation set. The Japanese name appearing in the evaluation set is 70 words in total including the family name and given name (different word number is 52). Of these, the surnames that did not appear in the personal name list in Table 3 were 3 words (1 word for Asagino, 2 words for Chinzai).

Next, Table 3 shows the voice recognition rates of the method of this embodiment and the method of the comparative example.

[0092]

[Table 5] Speech recognition rate ――――――――――――――――――――――――――――――――――― Recognition rate (%) This embodiment Comparative example ――――――――――――――――――――――――――――――――――― Word recognition rate 87.51 87.30 Surname and surname word recall 70 73 Compliance rate 67 75 ――――――――――――――――――――――――――――――――――― (Note) The surname and given name will be evaluated as correct only if the reading, class and section are all correct. It was It is considered that the recognition rate of the method of the comparative example roughly corresponds to the upper limit of the method of the present embodiment. To be

According to the method of this embodiment, the unregistered words, namely the family name and the first name, can be recognized with substantially the same accuracy as when they are recognized as the registered words. Contrary to expectations, one of the reasons why the word recognition rate of the method of the present embodiment exceeds the method of the comparative example is as follows. For some surnames and given names with low acoustic likelihood,
It is considered that the method of the present embodiment correctly detects a section although there is a reading error, and as a result, it rarely induces a recognition error in the preceding and following words. This is supported by the fact that the method of the present embodiment is excellent in the recall and precision of the family name and first name sections in which the reading error shown in Table 6 below is ignored.

[0094]

[Table 6] Section detection rate for surnames and given words ――――――――――――――――――――――――――――――――――― Recognition rate (%) This embodiment Comparative example ――――――――――――――――――――――――――――――――――― Surname and surname recall rate 87 80 Compliance rate 84 82 ――――――――――――――――――――――――――――――――――― (Note) If the class and section of family name and given name are correct, it is evaluated as the correct answer (acoustic likelihood shadow). Ignore misleading reading errors. )did.

FIG. 10 is a graph showing the experimental results of the experiment by the present inventor showing the unitization effect of the mora chain on the recall rate of Japanese surnames and given names, and word reproduction with respect to the number of types of the mora chain. It is a graph which shows a rate. As is clear from FIG. 10, it is considered that the likelihood of surname and given name by the model is increased and the recall is improved by increasing the mora chain to be unitized. This is consistent with the trend of improving the average likelihood for the learning set described above.

Next, the voice recognition rate for rare family names and first names will be described. The advantage of the unregistered word model proposed in the present embodiment is that rare words that cannot be predicted in advance can be correctly recognized. Here, the method of the present embodiment is evaluated by simulating such rare family names and first names. 52 types of Japanese family names and surnames appear in the evaluation set. Therefore, after deleting all surnames and given names having the same readings as these words from the learning data in Table 1, a comparison experiment of speech recognition rates by the method of the present embodiment and the method of the comparative example was performed as in the previous section. . Table 7 shows the results.

[0097]

[Table 7] Rare surname and first name voice recognition rate ―――――――――――――――――――――――― Recognition rate (%) Comparison of this embodiment ――――――――――――――――――――――― Word recognition rate 86.66 86.08 Surname and first name word recall rate 31 6 Relevance rate 36 8 ―――――― ―――――――――――――――――― (Note) From the surname and given name data used for learning, delete all entries that have the same reading as the surname and given name that appear in the evaluation set. . The surname and given name were evaluated as correct only when the reading, class and section were all correct. Registered word recall rate
The precision does not reach 0% because some surnames became "common nouns" due to morpheme imperfections.

As is clear from Table 5, in the method of this embodiment, even if a surname and given name that do not exist in the learning data are given,
With a recall of 31%, the reading, class, and section were correctly recognized. As a result, even with the word recognition rate, the registered word method is 0.58.
I exceeded the points.

As described above, according to the embodiment of the present invention, the following unique effects can be obtained by making the unregistered word model class-dependent. (1) By limiting the modeling target, it is possible to clarify the statistical characteristics of reading and to introduce the parameter constraint peculiar to the class. Therefore, the accuracy of the unregistered word model can be improved. (2) Language processing of the detection section is possible. Unregistered words are
In addition to reading, the class is also identified at the same time. Yomi and class are considered to be necessary and sufficient information in the language processing of proper nouns in many cases. (3) By using the generated statistical language model for voice recognition, it is possible to perform voice recognition with a higher recognition rate as compared with the related art.

<Second Embodiment> FIG. 11 is a block diagram showing the configuration of a continuous speech recognition system according to a second embodiment of the present invention. FIG. 12 shows the continuous speech recognition system of FIG. It is a block diagram which shows the structure of the telephone with an automatic dialing function used.

The continuous speech recognition system of FIG. 11 differs from the continuous speech recognition system of FIG. 1 in the following points. (1) The learning data memory 30 is a Japanese surname file 30.
In addition to "a" and the Japanese name file 30b, files such as the place name file 30c are included. The place name file is, for example, a file containing words of place names in Japan and foreign countries. (2) Word class N-gram model generation unit 23 in FIG.
Instead, the finite state automaton model generation unit 23a is provided. (3) Word class N-gram model memory 43 of FIG.
Instead, it has a finite state automaton model memory 43a.

Here, the finite state automaton model generator 23a generates a finite state automaton model based on text data called a corpus composed of a large number of Japanese written sentences stored in the text database memory 31, and this model is generated. It is stored in the finite state automaton memory 43a.

In the continuous speech recognition system of FIG. 11,
The part surrounded by the broken line is called the voice recognition device 100. That is, the speech recognition device 100 includes a circuit and a processing unit from the microphone 1 to the word hypothesis narrowing unit 6, a phoneme HMM memory 11 and a word dictionary memory 1 connected to the word matching unit 4.
2, and a statistical language model memory 44 connected to the word hypothesis narrowing unit 6. In the second embodiment, the voice recognition device 100 is used to configure the telephone with the automatic dialing function of FIG. The data in the word dictionary memory 12 and the statistical language model memory 44 are generated and stored in advance by the system shown in FIG. Therefore, the voice recognition device 100 is the microphone 1
In response to a speaker's voice of a word such as a person's name, the voice recognition processing is executed and the character string of the voice recognition result is output.

FIG. 12 shows the configuration of a telephone with an automatic dialing function according to the present embodiment, and the main control unit 50
Is composed of a CPU and controls the overall operation of this telephone by executing a predetermined operation program stored in the ROM 51. The RAM 52 stores data required when the operation program is executed by the main control unit 50, and is used as a temporary working memory for the main control unit 50. The display unit 53 is, for example, a display device such as a liquid crystal display device (LCD), and displays the operating state of the telephone, the name of the transmission destination, and the telephone number. The operation unit 54 also includes character keys, a numeric keypad for dialing, speed dial keys, various function keys, and the like necessary for operating the telephone. Further, the network control unit (NCU) 55 is an MTD that performs an operation of closing and opening a DC loop of an analog public telephone line L and has an automatic dial function.
A hardware circuit including an F dialer, which is connected to the handset 59 as necessary, or the voice synthesis output unit 5
The output from 6 is connected to the public telephone line L. Furthermore, the voice synthesis output unit 56 includes, for example, a pulse generator, a noise generator, a variable gain amplifier, and a filter, and is controlled by the main control unit 50 using a well-known voice synthesis method. After converting the text data of the character string to a predetermined parameter time series, the pulse generator is controlled based on the pitch, and the pulse generator and the noise generator are selectively selected based on voiced / unvoiced switching. By switching and using, the variable gain amplifier is controlled based on the amplitude value, and the filter is controlled based on the filter coefficient value, the voice of the character string is voice-synthesized and output through the speaker 57, Alternatively, the voice signal of the voice synthesis is NC.
It is transmitted to the other party of communication via the U55 and the public telephone line L. The above circuits 51 to 56 and the telephone number search unit 60 are connected to the main control unit 50 via the bus 58.

The telephone number table memory 61 stores in advance a person's name and a telephone number corresponding thereto in the form of a table. Then, the telephone number search unit 60 uses the voice recognition device 10
Based on the word of the person's name in the character string following the word "call" in the voice recognition result from 0, the information of the telephone number corresponding to the person's name is read from the telephone number table memory 61, and the information of the telephone number is stored in the bus. Main control unit 5 via 58
Output to 0. In response to this, the main control unit 50 outputs the telephone number information to the MTDF dialer in the NCU 55. At this time, after the NCU 55 goes off-hook to make a call, the MTDF dialer changes the input telephone number information. A corresponding dial signal is generated and sent to the public telephone line L. This allows the user to make a call to the telephone having the telephone number corresponding to the person's name uttered by the microphone 1.

According to the second embodiment, in a telephone having a voice recognition and automatic dialing function, the accuracy of voice recognition regarding an unregistered word that is not registered in the word dictionary is higher than that of the conventional example. It is possible to perform information retrieval using a voice recognition device capable of
Since a large number of proper names such as personal names can be recognized by voice in a limited memory, it is possible to retrieve information with higher accuracy in a telephone equipped with a database as compared with the prior art. Also, automatic dialing can be performed with a high voice recognition rate.

Although the finite state automaton model generation unit 23a and the finite state automaton memory 43a are provided in the above embodiment, the present invention is not limited to this.
The word class N-gram model generation unit 2 in FIG.
3, the word class N-gram model memory 43 of FIG. 1 may be provided.

<Third Embodiment> FIG. 13 is a block diagram showing the configuration of a private branch exchange (PBX) according to a third embodiment of the present invention. This embodiment is characterized in that the voice recognition device 100 of FIG. 11 and the telephone number search unit 60 and the telephone number table memory 61 of FIG. 12 are applied to extension transfer or outside line transfer of a private branch exchange.

In FIG. 13, the main control unit 150 displays the CP
The entire operation of this private branch exchange is controlled by executing a predetermined operation program which is composed of U and is stored in the ROM 151. The RAM 152 stores data required when the operation program is executed by the main control unit 150, and is used as a temporary working memory for the main control unit 150. The display unit 153 is, for example, a display device such as a liquid crystal display device (LCD), and displays an operating state of the private branch exchange, a destination name, and a telephone number. In addition, the operation unit 154 includes character keys necessary for operating the private branch exchange, numeric keys for dialing, various function keys, and the like. Further, the network control unit (NCU) 155 includes a telephone exchange switch circuit that alternately connects a plurality of public telephone lines L1 to LN that are external lines and extension lines that are connected to the extension telephones T1 to TM, and also has an analog switching circuit. MTDF for closing and opening DC loops of each public telephone line L1 to LN and having an automatic dialing function
A hardware circuit including a dialer, which outputs the output from the voice synthesis output unit 156 to the public telephone line L1 as necessary.
To LN. Furthermore, the voice synthesis output unit 15
6 is configured similarly to the voice synthesis output unit 56 of FIG. 12,
Under the control of the main control unit 50, the text data of the character string to be voice-synthesized is voice-synthesized, the voice of the character string is voice-synthesized, and the voice signal is communicated via the NCU 155 and the public telephone lines L1 to LN. Send to the other party. The above circuits 151 to 156 and the telephone number search unit 6
0 is connected to the main control unit 150 via the bus 158.

The telephone number table memory 61b stores in advance a person's name and corresponding extension telephone numbers and outside telephone numbers in the form of a table. Then, the main controller 150
The NCU 155 automatically responds to an incoming call from one of the public telephone lines L1 to LN.
It is a BC company. Which extension should I connect to? "
The voice synthesis signal is generated by the voice synthesis output unit 156 and output to the other party. On the other hand, the voice signal of the person's name to be transferred by extension, which is uttered by the other party, is output from the NCU 155 to the feature extractor 2 of the voice recognition device 100. At this time, the voice recognition device 100 executes the voice recognition process and outputs the character string of the voice recognition result to the telephone number search unit 60. In response to this, the telephone number search unit 60 causes the voice recognition device 1 to
The information of the extension telephone number corresponding to the person's name is read from the telephone number table memory 61b based on the word of the person's name of the character string of the voice recognition result from 00, and the information of the extension telephone number is mainly transmitted via the bus 158. Output to the control unit 150. In response to this, the main control unit 150 controls the NCU 155 based on the information of the extension number to connect the incoming public telephone line to the extension telephone of the corresponding extension number, thereby completing the extension transfer. To do.

In the above embodiment, an example of extension transfer has been described, but the same applies to the voice recognition device 1 when an extension call is made from the extension telephones T1 to TM to the public telephone line.
00, the telephone number search unit 60, and the telephone number table memory 61b can be configured so that the user can make an outside line by simply saying the "personal name" that the user wants to make an outside line.

According to the third embodiment, in the private branch exchange equipped with the voice recognition and automatic transfer functions, the accuracy of voice recognition regarding unregistered words that are not registered in the word dictionary is higher than that of the conventional example. It is possible to perform information retrieval by using a voice recognition device capable of performing voice recognition, which enables voice recognition of a large number of proper names such as personal names in a limited memory. Information can be searched with higher accuracy than Also, automatic transfer is possible with a high voice recognition rate.

<Fourth Embodiment> FIG. 14 is a block diagram showing the configuration of a car navigation system according to a fourth embodiment of the present invention. This embodiment is shown in FIG.
The voice recognition device 100 is applied to a car navigation system.

In FIG. 14, the main controller 250 uses the CP
A predetermined basic program which is composed of U and is stored in the ROM 251 and C in the CD-ROM drive device 259.
The entire operation of this car navigation system is controlled by executing the application program loaded from the D-ROM to the flash memory 258.
The RAM 252 stores data necessary when the main control unit 250 executes the basic program or the application program, and is used as a temporary working memory for the main control unit 250. Display unit 253
Is a display device such as a liquid crystal display device (LCD), and displays the operation state of the car navigation,
Display the map and navigation information around the designated place name. In addition, the operation unit 254 includes character keys necessary for operating the navigation system, numeric keys for dialing, various function keys, and the like. further,
The voice synthesis output unit 256 is the voice synthesis output unit 56 of FIG.
Is configured in the same manner as described above, and by the control from the main control unit 250,
The text data of the character string to be voice-synthesized is voice-synthesized, the voice of the character string is voice-synthesized, and the voice signal is output to the speaker 257. CD-ROM drive device 2
A CD-ROM storing car navigation information such as an application program for car navigation and map information is inserted in 59, and these pieces of information are transferred from the CD-ROM to a CD-ROM drive device 259.
And is loaded into the flash memory 258 via the bus 258 and used. The above circuits 251-256, 2
59 and the place name search unit 60a are connected to the main control unit 250 via a bus 258.

The place name table memory 61a stores place names and position information (latitude and longitude information) corresponding thereto in advance in the form of a table. Then, the user recognizes the voice recognition device 1
When the place name is uttered to the microphone 1 of 00, the voice recognition device 100 executes the voice recognition process and outputs the character string of the voice recognition result to the place name search unit 60a. In response to this, the place name search unit 60a reads the position information corresponding to the place name from the place name table memory 61a, based on the word of the place name in the character string of the voice recognition result from the voice recognition device 100, and the position concerned. The information is output to the main control unit 250 via the bus 258. In response to this, the main control unit 250
On the basis of the position information.
The car navigation information such as map information in the inside is searched, the corresponding searched information is displayed on the display unit 253, and the searched voice information is output to the voice synthesis output unit 256.
To output the synthesized voice from the speaker 257.

According to the fourth embodiment, in the car navigation system having the voice recognition and car navigation functions, the accuracy of voice recognition regarding unregistered unregistered words can be increased as compared with the conventional example. It is possible to perform information retrieval using the device, which makes it possible to recognize a large number of proper names of place names with voice in a limited memory, which is higher in a car navigation system with a database than in the prior art. Information can be retrieved with accuracy. Further, the place name can be recognized by voice with a high voice recognition rate, and the car navigation process can be appropriately executed.

In the above second, third and fourth embodiments, examples of a telephone set, a private branch exchange, and a car navigation system are explained, but the present invention is not limited to this, and an ordinary word list corresponding to Stores a database memory containing word data of nouns and corresponding information,
Using the character string of the voice recognition result output from the voice recognition device 100 as a key, the database stored in the storage device of the database is searched, and the information corresponding to the matching word data is read from the database memory. It is possible to output and further execute a predetermined process based on the retrieved information.

<Fifth Embodiment> FIG. 15 is a block diagram showing the arrangement of a Kana-Kanji conversion apparatus according to the fifth embodiment of the present invention. The same components as those in FIG. 1 are designated by the same reference numerals. is doing. The kana-kanji conversion device according to this embodiment includes a keyboard 71, a keyboard interface 72, a phoneme HMM memory 11 and a word dictionary memory 12.
A word matching unit 4a connected to, a buffer memory 5,
And a word hypothesis narrowing unit 6 to which a statistical language model memory 44 is connected.

Here, the word dictionary memory 12 stores the word dictionary generated by the word dictionary generator 22 of FIG.
Here, the word dictionary is a file stored in the learning data memory 30 (Japanese surname file 30a, Japanese name file 30b, place name file 30 shown in FIGS. 1 and 11).
Not limited to c, as described above, the foreigner's surname and first name, company name,
Words such as various facility names and various product names may be included. )
And a plurality of word data in Kanji notation corresponding to the words of the text data in the text data memory 31. Further, the statistical language model memory 44 stores the statistical language model generated by the language model generation unit 24 of FIG. 1, and this statistical language model is the learning data memory 3 described above.
File and text data memory 31 stored in 0
Generated based on the words in the text data in.

In FIG. 15, a keyboard 71 is an input means for inputting a kana character string, and a keyboard interface 72 temporarily stores the data of the kana character string input using the keyboard 71 and then a predetermined signal. After performing processing such as conversion, it outputs to the word collating unit 4a. The word matching unit 4a uses the one-pass Viterbi decoding method, based on the data of the kana character string input via the keyboard interface 72, and the phoneme HMM in the phoneme HMM memory 11 and the word dictionary memory 12. The word hypothesis is used to detect the word hypothesis, and the likelihood is calculated and output. Here, specifically, the word matching unit 4a refers to the word dictionary, performs word matching and likelihood calculation between the input kana character string and the word in the word dictionary, and matches. At times, it is converted into a word in Kanji notation and output as a character string of a word hypothesis with likelihood, while when it does not match, it is output as a kana character string as a character string of word hypothesis with likelihood. The output data from the word matching unit 4a is input to the word hypothesis narrowing unit 6 via the buffer memory 5. The word hypothesis narrowing unit 6 refers to the statistical language model in the statistical language model memory 44 on the basis of the word hypothesis output from the word matching unit 4a via the buffer memory 5 and starts the end times at the same time. For the word hypothesis of the same word at different times, one of the highest likelihoods of the calculated total likelihoods from the utterance start time to the end time of the word is calculated for each head phoneme environment of the word. After narrowing down the word hypotheses so that they are represented by the word hypotheses, out of the word strings of all the word hypotheses after narrowing down, the word string of the hypothesis that has the maximum total likelihood and contains the Kanji notation is output as the recognition result. To do.

According to the fifth embodiment, the kana-kanji conversion device can provide the kana-kanji conversion device which can improve the kana-kanji conversion rate as compared with the prior art by using the statistical language model. . Therefore, for example, unregistered proper nouns can be converted.

[0122]

As described above in detail, according to the present invention, since the statistical language model including the unregistered word model constructed depending on the class is generated, the following unique effect can be obtained. . (1) By limiting the modeling target, it is possible to clarify the statistical characteristics of reading and to introduce the parameter constraint peculiar to the class. Therefore, the accuracy of the unregistered word model can be improved. (2) Language processing of the detection section is possible. Unregistered words are
In addition to reading, the class is also identified at the same time. Yomi and class are considered to be necessary and sufficient information in the language processing of proper nouns in many cases. (3) By using the generated statistical language model for voice recognition, it is possible to perform voice recognition with a higher recognition rate as compared with the related art.

Further, according to the information retrieval processing device of the present invention, in a small-scale information retrieval processing device such as voice recognition and automatic dialing function in a telephone or car navigation, unregistered words not registered in the word dictionary. It is possible to perform information retrieval using a voice recognition device in which the accuracy of voice recognition is higher than that in the conventional example. Therefore,
Since a large number of personal names and place names such as place names can be recognized by voice in a limited memory, information retrieval processing devices such as small portable devices equipped with a database can retrieve information with higher accuracy than conventional technologies. Becomes

[0124]

[Brief description of drawings]

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

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

FIG. 3 is a flowchart showing an unregistered word model generation process executed by an unregistered word model generation unit 20 in FIG.

FIG. 4 is a sub word 2-g which is a subroutine of FIG.
It is a flowchart which shows the unit determination process (step S4) of ram.

5 is a flowchart showing a language model generation process executed by the language model generation unit 24 of FIG.

FIG. 6 is a graph showing the distribution of word names in Japanese family names and first names and travel conversations, which is analyzed by the present inventors, and is a graph showing the ratio of the number of words to the mora length.

FIG. 7 is a state transition diagram showing an example of a statistical language model based on the class-dependent unregistered word model according to the first embodiment.

FIG. 8 is a state transition diagram showing an example of the statistical language model according to the first embodiment.

9 is a graph showing the improvement of the average likelihood due to the unitization of the mora chain in the unregistered word model generation processing executed by the unregistered word model generation unit 20 of FIG.
It is a graph which shows the average likelihood with respect to the number of types of mora chain.

FIG. 10 is an experimental result of an experiment relating to the continuous speech recognition system of the first embodiment by the present inventor, and is a graph showing the unitization effect of the mora chain in the recall rate of Japanese surnames and given names; It is a graph which shows the word recall with respect to the number of types of mora chain.

FIG. 11 is a block diagram showing a configuration of a continuous speech recognition system according to a second embodiment of the present invention.

FIG. 12 is a diagram showing a case of using the continuous speech recognition system of FIG.
It is a block diagram showing a configuration of a telephone with an automatic dialing function.

FIG. 13 is a block diagram showing a configuration of a private branch exchange (PBX) according to a third embodiment of the present invention.

FIG. 14 is a block diagram showing a configuration of a car navigation system that is a fourth embodiment according to the present invention.

FIG. 15 is a block diagram showing a configuration of a kana-kanji conversion device according to a fifth embodiment of the present invention.

[Explanation of symbols]

1 ... Microphone, 2 ... Feature extraction section, 3, 5 ... Buffer memory, 4, 4a ... Word matching section, 6 ... Word hypothesis narrowing section, 11 ... Phoneme HMM memory, 12 ... Word dictionary memory, 20 ... Unregistered model Generation unit, 21 ... Subword unit data generation unit, 22 ... Word dictionary generation unit, 23 ... Word class N-gram model generation unit, 23a ... Finite state automaton model generation unit, 24 ... Language model generation unit, 30 ... Learning data memory , 30a ... Japanese surname file, 30b ... Japanese name file, 30c ... Place name file, 31 ... Text database memory, 40 ... Subword unit N-gram model memory, 41 ... Mora length gamma distribution data memory, 42 ... Labeled subword unit Data memory, 43 ... Word class N-gram model memory, 43a ... Finite state State automaton model memory, 44 ... Statistical language model memory, 50 ... Main control unit, 51 ... ROM, 52 ... RAM, 53 ... Display unit, 54 ... Operation unit, 55 ... Network control unit (NCU), 56 ... Speech synthesis Output unit, 57 ... Speaker, 58 ... Bus, 59 ... Handset, 60 ... Telephone number search unit, 60a ... Place name search unit, 61, 61b ... Telephone number table memory, 61a ... Place name table memory, 71 ... Keyboard, 72 ... Keyboard interface, 100 ... Voice recognition device, 150 ... Main control unit, 151 ... ROM, 152 ... RAM, 153 ... Display unit, 154 ... Operation unit, 155 ... Network control unit (NC)
U), 156 ... Voice synthesis output section, 158 ... Bus, 250 ... Main control section, 251 ... ROM, 252 ... RAM, 253 ... Display section, 254 ... Operation section, 256 ... Voice synthesis output section, 257 ... Speaker, 258 ... bus, 259 ... CD-ROM drive device, L, L1 to LN ... public telephone line, T1 to TM ... extension telephone.

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yoshinori Kosaka 5 Seiji-cho, Seika-cho, Soraku-gun, Kyoto Prefecture Mihiratani No. 5, A-T-R, Inc. Voice Translation Research Institute (56) References (JP, A) JP-A-6-308994 (JP, A) JP-A-2000-99082 (JP, A) Masaaki NAGATA, A Part of Speech Estimation Method for Golf Japan, Waisting in the United States of America. Annual Meeting of the Association for Computational Linguistics Pro eedings o f the Conference, the United States, Association for Computational Ling uistics, 6 May 20, 1999, 277-284 (58) investigated the field (Int.Cl. ^7, DB name) G10L 15/18 JICST file (JOIS)

Claims (5)

(57) [Claims] 1. A learning data storage means for storing learning data including a word list of proper nouns or common nouns of foreign words, and a mora length in the learning data based on the learning data stored in the learning data storage means. The parameter of the gamma distribution of the mora length is assumed to be calculated depending on the class, assuming that the ratio of the number of words to is substantially following the gamma distribution. Or a first N-gram having a class that is a subclass of a common noun of a foreign word
First generation means for generating a subword unit N-gram model in which an unregistered word is modeled by calculating an appearance probability of, a word class N-gram model generated based on a predetermined text database, A sub-word unit N-gram model generated by the first generation means,
The word class based on the parameters of the gamma distribution of the mora length calculated by the first generating means;
A statistical language model including unregistered words based on subword units is generated by calculating the probability of occurrence of the second N-gram depending on a class that is a subclass of the proper noun or the common noun of a foreign word. A statistical language model generation apparatus comprising: a second generation unit. 2. A plurality of subword units are extracted for each subword unit by extracting the subword unit based on the subword unit N-gram model generated by the first generating unit and applying the extracted label to the subword unit. Third to generate data in units of labeled subwords
Generating a word dictionary by adding a phoneme arrangement to the word extracted from the text database and the data in units of a plurality of labeled subwords generated by the third generating means. statistical language model generating apparatus according to claim 1, wherein, further comprising fourth and generating means. 3. Based on the audio signal of the utterance sentence input, the speech recognition apparatus having speech recognizing speech recognition means by using a predetermined statistical language model, the speech recognition means, according to claim 1 or speech recognition, characterized in that the speech recognition using a statistical language model generated by 2 statistical language model generating apparatus according, a word dictionary generated by the fourth generating means according to claim 2, wherein apparatus. 4. A database storage means for storing a database containing word data of common nouns corresponding to the word list and information corresponding thereto, and a character of a voice recognition result output from the voice recognition device according to claim 3. Search means for searching the database stored in the database storage means, using the column as a key, and reading and outputting information corresponding to the matching word data from the database storage means. Information retrieval processing device. 5. The information search processing device according to claim 4, further comprising a processing execution means for executing a predetermined processing based on the information output from the search means. apparatus.