JP6276516B2 - Dictionary creation apparatus and dictionary creation program - Google Patents

Description

  The present invention relates to a dictionary creation device and a dictionary creation program for creating a dictionary (vocabulary) used in speech recognition.

  Conventionally, in speech recognition, correct candidate word strings (correct candidate word strings) are generated from speech input using a statistical language model typified by n-gram. In the statistical language model, the linguistic likelihood of the correct candidate word string is calculated. In the n-gram model, which is one of the statistical language models, linguistic likelihood is calculated by predicting subsequent words based on a word string (context) that lies ahead.

For example, given a word string w = w ₁ , w ₂ , w ₃ , w ₄ , the likelihood of the word string when using a bigram that is one of the n-gram models is as follows (Equation 1 ). Here, P (w ₂ | w ₁ ), P (w ₃ | w ₂ ), and P (w ₄ | w ₃ ) are bigrams, respectively. The bigram value is a probability and is estimated from a large amount of text data by the maximum likelihood method.

Here, in speech recognition, consecutive word pairs in a word string are connected as one compound word and connected in order to improve the predictability of words in a statistical language model typified by n-gram. A compound word is registered in a dictionary (vocabulary) for speech recognition. In the above example, w ₂ and w ₃ are not treated as one independent word, but correspond to treating them as one word together like w _c = w ₂ and w ₃ . The likelihood of the word string w ′ = w ₁ , w _c , w ₄ using the compound word w _c can be calculated as (Equation 2) below.

  If P ′ (w ′) − P (w) is a positive value (> 0), the given word string is more likely to be linguistically, and it can be regarded that the word prediction accuracy by bigram has improved. However, P ′ (w ′) − P (w) is not always a positive value, and a word that becomes a compound word so as to improve linguistic likelihood from among a plurality of compound word candidates. It is necessary to select a pair.

Conventionally, compound words registered in a dictionary are created from text data using linguistic criteria. There are the following two linguistic criteria for creating a compound word.
[1] Criteria based on frequency (occurrence frequency) of word pairs constituting a compound word [2] Criteria based on difference in entropy before and after composing a compound word Here, entropy is calculated from a given language model An index indicating the complexity of the word string or an index indicating the difficulty of speech recognition. The higher the entropy, the more complex the word string and the more difficult the speech recognition by the language model.

  Since these are all used as a standard for composing compound words from the past, they will be briefly described here. According to the first criterion (a criterion based on the frequency (occurrence frequency) of word pairs constituting a compound word), a word string that appears in large numbers in text data is preferentially extracted as a candidate compound word to be additionally registered. . Further, according to the second criterion (a criterion based on a difference in entropy before and after composing a compound word), a low complexity word string (a word string that can be easily recognized by speech) is a candidate compound word to be additionally registered. As preferentially extracted.

  In addition, a technique has been proposed in which the two criteria are combined to sequentially obtain a set of words with the highest accuracy and the highest n-gram prediction performance (see, for example, Patent Document 1 and Non-Patent Document 1). . The procedure is as follows.

[1] The frequency of a set of two adjacent words (a set of two words) in the text data is calculated.
[2] A pair of the most frequently used M words is selected as a compound word candidate.
[3] Calculate entropy when a pair of words (compound word candidate) is regarded as a compound word and entropy when it is not regarded as a compound word.
In the procedure [3], for example, _a set of four words (w ₁ , w _a , w) existing in the learning data in consideration of the word set w _a , w _b and the preceding and following words w ₁ , w _{2. b} , w ₂ ). The entropy H ₁ for the word quadruple is obtained by the following (Equation 3).

Here, the simultaneous distribution P (w ₁ , w _a , w _b , w ₂ ) in which all the words w ₁ , w _a , w _b , w ₂ occur simultaneously can be approximated by the following (formula 4). it can.

On the other hand, entropy H ₂ when the pair of words w _a and w _b is regarded as a compound word w _c is obtained by the following (Equation 5).

Further, the simultaneous distribution P (w ₁ , w _c , w ₂ ) can be approximated by the following (Expression 6), similarly to (Expression 4).

Entropy _{H 1} and word 2 for words quadruplet set _w a, the difference _{ΔH (w c) ≡H 2 -H} 1 and entropy of _{H 2} when regarded _{w b} and compound words _{w c,} the following ( It can be calculated as Equation 7).

  A word satisfying the following (Equation 8) from a set of M words is defined as a compound word. Here, “argmax f (t)” in (Expression 8) indicates that a variable t that maximizes f (t) is obtained for the function f (t) related to the variable t. Thereafter, this procedure is repeated until a desired number of compound words are obtained.

JP 2007-171724 A (paragraphs 0032 to 0054, FIG. 2)

Akio Kobayashi et al., “Language model using compound words and classes for conversational speech”, Proceedings of the Acoustical Society of Japan, March 2006, 2-1-1, p.71-p.72

  However, two linguistic criteria that have been used in the past and a combination of these two criteria described in Patent Literature 1 and Non-Patent Literature 1 are texts for estimating a language model such as n-gram. Based on statistics obtained from the data. Therefore, the prior art is not always suitable for an application that estimates a word string from input speech, such as speech recognition.

  That is, in speech recognition, when input speech is given, a speech recognition result is output in consideration of not only linguistic likelihood but also acoustic likelihood. Therefore, the conventional technology that considers only linguistic likelihood does not always create a compound word that can improve speech recognition performance. This means that the prior art does not consider the error probability of the predicted word when speech recognition is performed.

  The present invention has been made in view of the above points, and is a dictionary creation device in which the prediction accuracy of words in a statistical language model of speech recognition is improved, and recognition errors are reduced as compared with a conventional statistical language model, and It is an object to provide a dictionary creation program.

  In order to solve the above problems, a dictionary creating apparatus according to an aspect of the present invention is a dictionary creating apparatus that registers a compound word used for speech recognition in a dictionary as a new vocabulary, and includes a spoken language resource storage unit, a storage unit, The speech recognition unit, the alignment unit, the compound word frequency counting unit, the entropy calculation unit, the compound word selection unit, and the dictionary registration unit are used.

  In this dictionary creation device, a speech and a correct word string obtained by characterizing the speech without error are stored in the spoken language resource storage unit. In the dictionary creation device, the storage unit stores the dictionary, a statistical language model indicating the degree of connection between words in the speech, and a statistical acoustic model indicating the relationship between the words in the speech and waveform patterns. The

In the dictionary creation device, the speech recognition unit recognizes the speech using the statistical language model and the statistical acoustic model, and generates a correct candidate word sequence indicating a correct speech candidate word sequence. Next, in the dictionary creation device, the alignment unit aligns the correct word string and the correct candidate word string in accordance with the utterance time. In the dictionary creation device, the compound word frequency counting unit counts the frequency of each pair of words included in the correct word sequence and the correct candidate word sequence aligned in the alignment unit, and the high frequency Accordingly, the two-word pair is extracted as a plurality of compound word candidates of compound words to be registered in the dictionary.

Further, the dictionary creation device uses each of the plurality of compound word candidates extracted by the compound word frequency counting unit by the entropy calculation unit, the correct word string and the correct candidate word string arranged in the alignment unit, and This is an index indicating the difficulty of speech recognition of the speech when a compound word candidate is the compound word, and an entropy indicating that speech recognition is difficult as the value is higher is calculated. Further, the dictionary creation device selects a compound word candidate that minimizes the entropy when the compound word selection unit sets each compound word candidate as the compound word. Then, the dictionary creation device registers the compound word candidate selected by the dictionary registration unit by the compound word selection unit as a new vocabulary of the dictionary.

  The dictionary creation device having such a configuration extracts compound word candidates from the correct word sequence of speech and the correct candidate word sequence that is the result of speech recognition by the speech recognition unit. Further, the dictionary creation device calculates entropy when the compound word candidate is a compound word using the extracted compound word candidate, and selects a compound word to be registered in the dictionary using the calculated entropy.

  Further, the dictionary creation device according to one aspect of the present invention is configured such that the entropy calculation unit calculates the entropy using a language score obtained from the statistical language model and an acoustic score obtained from the statistical acoustic model. did.

  The dictionary creation device having such a configuration calculates entropy, which is a reference for selecting a compound word to be registered in the dictionary, using a language score obtained by a statistical language model and an acoustic score obtained by a statistical acoustic model. .

  Further, the dictionary creation device according to one aspect of the present invention is configured such that the spoken language resource storage unit sequentially stores the broadcasted speech and the correct word string of the speech. For example, a subtitle creation device (speech and correct word string is output) provided with a manual speech recognition result error correction function (correct word word creation function) is connected to the preceding stage of the spoken language resource storage unit. This is realized.

  The dictionary creation device having such a configuration reflects the broadcasted voice and the correct word string in real time (real time) in entropy calculation, which is a reference for registering compound words in the dictionary.

  The dictionary creation program according to an aspect of the present invention is a dictionary creation program that causes a computer to function as a dictionary creation device that registers a compound word used for speech recognition in the dictionary as a new vocabulary, wherein the computer stores a spoken language resource storage. And a storage unit, and causes the computer to function as speech recognition means, alignment means, compound word frequency counting means, entropy calculation means, compound word selection means, and dictionary registration means.

  In the computer that executes the dictionary creating program, the speech and the correct word string obtained by characterizing the speech without error are stored in the speech language resource storage unit. The computer stores the dictionary, a statistical language model indicating the degree of connection between words in the speech, and a statistical acoustic model indicating a relationship between the words in the speech and waveform patterns.

Further, the computer that executes the dictionary creating program, the speech recognition means recognizes the speech using the statistical language model and the statistical acoustic model, and correct candidate word strings indicating the correct speech candidate word strings Is generated. Next, the computer arranges the correct word string and the correct candidate word string in accordance with the utterance time. Then, the computer counts the frequency of each pair of words included in the correct word sequence and the correct candidate word sequence in which the compound word frequency counting means are arranged, and the word is determined according to the high frequency. Two pairs are extracted as a plurality of compound word candidates of compound words to be registered in the dictionary.

Further, the computer that executes the dictionary creation program uses the plurality of compound word candidates extracted by the entropy calculation means and the aligned correct word strings and the correct candidate word strings to convert each compound word candidate into the compound word. In this case, the entropy indicating that the speech recognition is difficult as the value is higher is calculated. Further, the computer selects a compound word candidate that minimizes the entropy when the compound word selection unit sets each compound word candidate as the compound word. Then, the computer registers the selected compound word candidate as a new vocabulary of the dictionary by the dictionary registration means.

  A dictionary creation program that realizes such a function extracts compound word candidates from a correct word sequence of speech and a correct candidate word sequence that is a result of speech recognition by speech recognition means. The dictionary creation program calculates entropy when the compound word candidate is a compound word using the extracted compound word candidate, and selects a compound word to be registered in the dictionary using the calculated entropy.

  According to one aspect of the present invention, entropy that is a reference for registering a compound word in a dictionary is calculated from a correct word sequence of speech and a correct candidate word sequence that is a result of speech recognition. Compound words to be registered in the dictionary can be determined in consideration of elements. Therefore, the accuracy of word prediction in the statistical language model for speech recognition is improved, and recognition errors are reduced as compared with the conventional statistical language model.

  Further, according to one aspect of the present invention, a statistical language model and a statistical acoustic model used in speech recognition can be used for entropy calculation for registering a compound word in a dictionary.

  Further, according to one aspect of the present invention, when a speech and a correct word string can be acquired sequentially, it is possible to configure a dictionary that follows the speech content by gradually adding compound words to the recognition dictionary. is there.

It is a functional block diagram which shows the structure of the dictionary creation apparatus which concerns on embodiment. It is a flowchart which shows the process sequence of the entropy calculation part which concerns on embodiment. It is a flowchart which shows the process sequence of the language score calculation in the entropy calculation part which concerns on embodiment. It is a flowchart which shows the process sequence of the language model update part which concerns on embodiment.

[Outline of the present invention]
First, the outline of the present invention will be described. In the conventional method, text data is used to determine a change in the text corpus before and after creating a compound word, and a compound word to be registered in the dictionary is determined based on this change (based on this change).

  In the present invention, instead of using text data, speech data and a correct word sequence (speech of speech) are used, and before and after the creation of a compound word for a plurality of correct candidate word strings on which correct answers as speech recognition results are superimposed. The entropy change is determined, and based on this change (based on this change), a compound word to be registered in the dictionary is determined.

  In the speech recognition result, the acoustic likelihood of the word string is given as a probability (logarithmic acoustic score) by a statistical acoustic model. For this reason, the conventional method uses only text data information to determine compound words to be registered in the dictionary without considering acoustic elements, whereas the present invention considers acoustic elements in the dictionary. Determine the compound word to register.

<< Criteria for registering compound words in dictionary according to the present invention >>
First, the description policy of the terms used below will be described. In general, matrices and sets are shown in bold, and in formulas, matrices and sets are shown in bold. However, a portion that cannot be represented in bold in the specification is replaced with a normal character that is not bold. For specific symbols that cannot be represented in the specification, the names of the symbols are written in parentheses.

Next, the criteria for registering a compound word in the dictionary according to the present invention will be described.
According to Bayes' theorem, when a speech input x is given, an acoustically and linguistically likely word string w (hat) can be obtained by the following (formula 9).

  Here, P (x | w) is an acoustic likelihood for a word string (sentence hypothesis), and a logarithmic acoustic score (hereinafter sometimes referred to as “acoustic score”) is an HMM (Hidden Markov). Model) and a statistical acoustic model represented by a mixed Gaussian distribution (GMM: Gaussian Mixture Model) (hereinafter, sometimes referred to as “acoustic model”).

  On the other hand, P (w) is a linguistic likelihood for the word string w, and a logarithmic language score (hereinafter sometimes referred to as “language score”) is a statistical value such as a word n-gram model. It is calculated by a language model (hereinafter, sometimes abbreviated as “language model”). In the following description, it is assumed that n-gram is used as the language model.

  In speech recognition, the following (Equation 10) is defined as an evaluation function of a sentence hypothesis based on the Bayes' theorem described above, and the best word sequence w ( Hat).

Here, f _am (x | w) is the acoustic score of the hypothesis w by the acoustic model, f _lm (w) is the language score of the hypothesis by the language model, and λ _lm is the weight of the language score with respect to the acoustic score, f (w | x) is a score for hypothesis w.
The posterior probability of hypothesis w in Bayes' theorem is calculated by the following (formula 12).

Assume that N voices are given. n-th sound _{x n (n = 1, ···} , N) with respect to, correct word sequence _{w n, 0} ∈W _n and a plurality of correct candidate word sequence is a speech recognition result _{w n, k} ∈W _n It is assumed that (k = 1,...) Is obtained. The conditional entropy H for the hypothesis set when x _n is given is given by (Equation 13) below. In addition, the entropy of all N voices is represented by the following (formula 14). Here, the conditional probability P (w _{n, k} | x _n ) of the word strings in (Equation 13) and (Equation 14) is calculated from the acoustic score and the language score by (Equation 12).

  If the prediction accuracy of the correct word string by the acoustic model and the language model is high, the score of the correct word string increases and the score of the hypothesis including other errors decreases. This means that the uncertainty associated with hypothesis identification is small, resulting in a small conditional entropy value. Therefore, if a compound word is created so that conditional entropy is as small as possible, word prediction accuracy when an acoustic model and a language model are combined is improved.

Now, the entropy of the entire N speeches when H is a compound word w _c when a specific pair of words in the word string included in the word string set W _n (n = 1,..., N) is H ( If w _c ) and the entropy before creating a compound word is H (w _c (bar)), the entropy difference is obtained from (Equation 15) below.

When a plurality of compound word candidates are given, if a compound word w _c (hat) as shown in (Equation 16) below is selected, w _c (hat) is the word prediction accuracy among the compound word candidates. Is the highest pair. That is, the compound word corresponding to (Expression 16) is registered in the dictionary using the difference in entropy of (Expression 15) as a compound word selection criterion.

The conventional method uses a change in entropy obtained from a text corpus, which is equivalent to using only the language score (entropy derived from) of the language model in speech recognition. On the other hand, in the present invention, since entropy considering logarithmic acoustic scores is used, it becomes a selection criterion for compound words suitable for speech recognition.
This is the end of the description of the outline of the present invention.

[Embodiment]
Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings as appropriate.
Each figure is only schematically shown so that the invention can be fully understood. Therefore, the present invention is not limited to the illustrated example. In addition, in each figure, about the same component or the same component, the same code | symbol is attached | subjected and those overlapping description is abbreviate | omitted.

<< Configuration of Dictionary Creation Device According to Embodiment >>
FIG. 1 is a functional block diagram showing a configuration of a dictionary creation device 100 according to an embodiment of the present invention. The dictionary creation device 100 includes a language model learning data storage unit 11, a vocabulary / language model storage unit 12, an acoustic model storage unit 13, a spoken language resource storage unit 14, a speech recognition unit 21, and a correct word string alignment unit. 22, an alignment unit 23, a compound word frequency counting unit 24, an entropy calculation unit 25, a compound word selection unit 26, and a language model update unit 27.

  Here, the language model learning data storage unit 11, the vocabulary / language model storage unit 12, the acoustic model storage unit 13, and the spoken language resource storage unit 14 are general recording media such as an HDD (Hard Disk Drive).

  The speech recognition unit 21, the correct word string alignment unit 22, the alignment unit 23, the compound word frequency counting unit 24, the entropy calculation unit 25, the compound word selection unit 26, and the language model update unit 27 are a CPU (Central Processing Unit). Is realized by a program execution process realized by developing a program stored in the HDD or the like in a RAM (Random Access Memory), a dedicated circuit, or the like. When these functions are realized by a program execution process, the “XX section” may be called “XX means”.

(Language model learning data storage)
The language model learning data storage unit 11 stores learning data for creating a language model used for speech recognition. The learning data is the text data such as news manuscripts or news transcripts divided into morphemes (words such as independent words and adjunct words) in advance by morphological analysis, and sentence breaks are given by dependency analysis. . Here, the morpheme refers to the minimum character string that has no meaning if it is made finer than this, and the morpheme analysis is to analyze the sentence by breaking it down to the morpheme level. In addition, dependency analysis is a method that divides a sentence into clauses and can determine in advance for each part of speech how a focused word contained in one clause relates to a word contained in another clause. The analysis is based on gender. For example, in text data in which word breaks and phrase breaks are added to the sentence “The weather is sunny today” in Japanese, “(today / of /) (weather / ha /) (clear / is) And so on. The symbol () indicates a paragraph break, and the symbol / indicates a word break.

(Vocabulary / Language Model Storage)
The vocabulary / language model storage unit 12 stores a dictionary (vocabulary) for creating a language model (statistical language model) and a statistical language model.
The dictionary (vocabulary) is based on words (in Japanese, independent words and adjunct words), but contains two or more consecutive word strings (compound words) to improve the word recognition rate in speech recognition. Consists of. The statistical language model is a kind of probability model (connection probability) indicating the degree of connection between words in speech, which is obtained by a statistical calculation method using a large amount of speech data.
For details, a vocabulary / language model storage unit before update of a selected compound word to be described later may be represented by reference numeral 12a, and a vocabulary / language model storage unit after update of the compound word may be represented by reference numeral 12b.

(Acoustic model storage unit)
The acoustic model storage unit 13 stores an acoustic model (statistical acoustic model).
The statistical acoustic model is a kind of probability model that is obtained based on waveform patterns of a large amount of speech and indicates the relationship between waveform patterns in speech and words.

(Spoken language resource storage)
The speech language resource storage unit 14 stores speech such as broadcast (including speech signals and speech data obtained by encoding speech signals) and correct word strings associated with the speech. The voice has a predetermined period as one unit (for example, a program unit or a corner unit), and is composed of a plurality of (for example, 10,000) utterances. The correct word string is a voice converted into a character, for example, a voice transcribed or a voice that has been voice-recognized in advance and has been corrected for errors such as insertion, replacement, and omission of words. . For this reason, the correct word string does not include an error with respect to the speech. Note that the spoken language resource storage unit 14 stores, for example, N pieces of speech and correct word strings accompanying the speech. Further, it is also possible to sequentially store the voice and the correct word string obtained from the caption generation device based on voice recognition having a manual error correction function in the storage unit.

(Voice recognition unit)
The voice that is accumulated in the spoken language resource storage unit 14 is input to the voice recognition unit 21. The voice recognition unit 21 uses the vocabulary stored in the vocabulary / language model storage unit 12, the language model, and the acoustic model stored in the acoustic model storage unit 13 to utter the voice accumulated in the spoken language resource storage unit 14. Recognize and generate a plurality of correct candidate word strings as speech recognition results.
The correct answer candidate word string is a word string that is likely to be input speech, and is a word sequence that may include speech recognition errors in the speech. Specifically, the speech recognition unit 21 generates approximately 100 to 500 correct candidate words for one utterance. Thereby, for example, when the speech is composed of 10,000 utterances, correct candidate word strings that are about 10,000 × 100 to 500 = 1,000,000 to 5 million speech recognition results are generated. .
Further, the speech recognition unit 21 aligns the correct candidate word strings in accordance with the utterance times of the words constituting the correct candidate word strings. At this time, the speech recognition unit 21 adds a language score based on the language model and an acoustic score based on the acoustic model to the correct candidate word string. Note that the correct word string may be included in the generated correct word candidate word string. The speech recognition unit 21 outputs the generated correct candidate word string to the alignment unit 23.

(Correct word string alignment part)
The correct word string alignment unit 22 receives the voice accumulated in the spoken language resource storage unit 14 and the correct word string of this voice. The correct word string alignment unit 22 forms a correct word string using the speech and correct word strings, the language model stored in the vocabulary / language model storage unit 12 and the acoustic model stored in the acoustic model storage unit 13. The correct word string is aligned according to the utterance time of the word. At this time, the correct word string alignment unit 22 adds a language score based on the language model and an acoustic score based on the acoustic model to the aligned correct word string. As a result, in the aligned correct word string, information indicating what word is spoken in minutes and seconds, and when the speech recognition unit 21 recognizes the correct word string using the language model and the acoustic model. Language score and acoustic score are added. The correct word string alignment unit 22 outputs the aligned correct word strings to the alignment unit 23.

(Alignment section)
The correct candidate word string generated by the speech recognizer 21 and the correct word string aligned in accordance with the utterance time by the correct word string aligner 22 are integrated into the aligner 23 and input as a word string set. Hereinafter, the word string set is used for calculation of a reference (entropy calculation) for creating a compound word to be registered in a dictionary (vocabulary).
The alignment unit 23 forms a word string set of a language model learned by a new dictionary (vocabulary) including a compound word stored in the vocabulary / language model storage unit 12 b and an acoustic model stored in the acoustic model storage unit 13. Each word string is aligned according to the utterance time of the word. At this time, the alignment unit 23 adds an acoustic score based on the acoustic model and a language score based on the language model to each word in the aligned word string. In the initial state, since no compound word is obtained, the language model stored in the vocabulary / language model storage unit 12a is the same as the language model stored in the vocabulary / language model storage unit 12b. Therefore, in the initial state, since the speech recognition unit 21 and the correct word string alignment unit 22 in the previous stage have already been aligned, the alignment unit 23 does not perform alignment. The alignment unit 23 outputs the aligned word string set to the compound word frequency counting unit 24.

(Compound word frequency counter)
The compound word frequency counting unit 24 receives the word string set sorted by the sorting unit 23. The compound word frequency counting unit 24 counts the frequency of a pair of specific words included in the word string set. In addition, after counting, the compound word frequency counting unit 24 arranges a pair of specific words in descending order of frequency, and calculates a pair of high-frequency M words having higher ranks as compound word candidates. The compound word frequency counting unit 24 outputs the word string set and the calculated high frequency M compound word candidates to the entropy calculating unit 25.

(Entropy calculation part)
The entropy calculation unit 25 receives the high-frequency M compound word candidates calculated by the word string set and the compound word frequency counting unit 24. The entropy calculation unit 25 calculates the entropy change by calculating the entropy before and after creating the compound word using the word string set and the high frequency M compound word candidates. Here, as described above, the entropy is an index indicating the complexity of a word string calculated from a given language model or an index indicating the difficulty of speech recognition. Complexity increases and speech recognition by the language model becomes difficult. Details of the processing of the entropy calculation unit 25 will be described in a processing procedure described later. The entropy calculation unit 25 outputs the calculated change in entropy to the compound word selection unit 26.

(Compound word selection part)
The compound word selection unit 26 receives a change in entropy calculated by the entropy calculation unit 25. The compound word selection unit 26 receives the result of the entropy calculation unit 25 in the previous stage, and selects the candidate having the largest entropy change from the M compound word candidates (the entropy after the compound word is generated is the compound word). A compound word candidate that is the smallest decrease compared to the entropy before creation is selected as a compound word. The compound word selection unit 26 outputs the selected compound word to the language model update unit 27.

(Language Model Update Department)
The compound word selected from the compound word selection unit 26 is input to the language model update unit 27 (dictionary registration unit). The language model update unit 27 newly adds the compound word obtained in the previous stage to the dictionary (vocabulary) of the vocabulary / language model storage unit 12, and uses the language model learning data used for learning the language model to change the language model. Re-learning is performed, and the language model in the vocabulary / language model storage unit 12 is updated. Details of the processing of the language model update unit 27 will be described in the processing procedure described later. The language model update unit 27 outputs to the alignment unit 23 that the language model has been relearned.

Then, the alignment unit 23 performs a new alignment using the language model stored in the updated vocabulary / language model storage unit 12b in which the compound word is updated. Thereafter, the processes of the aligning unit 23, the compound word frequency counting unit 24, the entropy calculating unit 25, the compound word selecting unit 26, and the language model updating unit 27 are repeated until a desired number of compound words are obtained. After the above processing is completed, the language model stored in the vocabulary / language model storage unit 12b is used for speech recognition.
Above, description about the structure of the dictionary creation apparatus 100 which concerns on embodiment is complete | finished.

<< Processing Procedure of Dictionary Creation Device According to Embodiment >>
Below, the process sequence of the principal part (entropy calculation part 25 and the language model update part 27) of the dictionary creation apparatus 100 which concerns on embodiment is demonstrated.

(Processing procedure of entropy calculation unit)
A processing procedure of the entropy calculation unit 25 will be described with reference to FIG.
Assume that N voices are given. For the n-th speech x _n (n = 1,..., N), a word string set (correct word string w _{n, 0} ∈W _n and a plurality of correct candidate word strings w _n, which are speech recognition results) _{It is} assumed that _l ∈ W _n (l = 1,...)) is obtained. Incidentally, correct word sequence _{w n, 0} and correct candidate word sequence _{w n,} language by (1) language model for _l score _f lm _{(w n,} l), and (2) acoustic Acoustic Model score _f am (w _{n, l} | x _n ) has already been calculated in the previous stage (the speech recognition unit 21 and the correct word string alignment unit 22), and is added to each word in each word string. Further, it is assumed that high frequency M compound word candidates calculated by the compound word frequency counting unit 24 are obtained.

The loop processing in step S10 means that the high-frequency M compound word candidates that are higher ranks calculated by the compound word frequency counting unit 24 are sequentially processed (for example, in descending order of frequency).
The entropy calculation unit 25 calculates in advance the amount of change in the language score before and after creating a compound word for the m-th (initially high-frequency first) compound word candidate (step S20). The change in the score is a calculation only in the part related to the words w _a and w _b and the compound word candidate w _c constituting the compound word. A detailed flowchart of the language score calculation in step S20 is shown in FIG.

With reference to FIG. 3, the processing procedure of the language score calculation in the entropy calculation part 25 is demonstrated.
The entropy calculation unit 25 calculates the frequency of the word string that is changed by configuring the mth compound word candidate as a compound word from the learning data stored in the language model learning data storage unit 11 (step S21). . Specifically, the following frequencies (1) to (6) are obtained.
(1) compound words become the words duplicate _(w a, _{w b)} Frequency of _{C (w a,} w b)
(2) the frequency of the two pairs of the word _{w 1} is connected to the left side of _{w a (forward) ∈ {V} (w 1} , w a) C (w 1, w a)
(3) _{w a} of the right two and word _{w 2} ∈ {V} to connect to (behind) the set _(w a, _{w 2)} The frequency of _C (w _{a, w} 2)
_{(4) (w a, w} b) 3 -tuple of the word _{w 1} to be connected to the left side (front) of _{(w 1, w a, w} b) frequency _C of _{(w 1, w a, w} b)
(5) _{w b} of the right two and word _{w 2} ∈ {V} to connect to (rear) pair _(w b, _{w 2)} The frequency of _C (w _{b, w} 2)
_{(6) (w a, w} b) 3 -tuple of the word _{w 2} to be connected to the right side (rear) of the _{(w a, w b, w} 2) The frequency of _{C (w a, w b,} w 2)

  Subsequently, the entropy calculation unit 25 calculates a language score (bigram) value related to the compound word from the frequency obtained in step S21 (step S22). Specific processing is described below.

First, the bigram P ′ (w _c | w ₁ ) after obtaining the compound word w _c is expressed by the following (Equation 17).

The bigram P ′ (w ₂ | w _c ) when w _c is a history is expressed by the following (formula 18).

On the other hand, the bigram P ′ (w _a | w ₁ ) of w _a after obtaining the compound word w _c is expressed by the following (formula 19).

The bigram P ′ (w ₂ | w _a ) when w _a is a history is represented by the following (formula 20).

The bigram P ′ (w ₂ | w _b ) when w _b is a history is expressed by the following (formula 21).

  Subsequently, the entropy calculation unit 25 stores the change of the language score (bigram) when the compound word is created in the language score table storage unit 15 in which the language score table is stored (language score table creation), Reference can be made in the subsequent process (language model update S130 in FIG. 4) (step S23). Then, the process proceeds to step S30 in FIG.

The loop processing in step S30 means that processing is sequentially performed on the given N voices (voice data), and as a result, processing in the loop is performed on all voices.
The entropy calculation unit 25 calculates the difference in language score between the nth (initially first) speech before and after the word duplication is made a compound word (step S40). ). In the present embodiment, using the bigram as a language model, a word duplicate _{(w a,} w _b), and compound word candidate _{w c.}
The difference in the language score of the language model before and after creating the compound word w _c is expressed by the following (formula 22).

In (Expression 22), let f _lm (w | ·) be the following (Expression 23). Here, δ _{u, w} is a function that returns the number of word pairs (u, w) included in the word string w.

  The language score of the combined word string w can be calculated by the following (Equation 24).

  Subsequently, the entropy calculation unit 25 calculates the difference between the acoustic scores of the acoustic model before and after composing the compound word (step S50). The difference between the acoustic scores is expressed by the following (Equation 25).

Here, x _{[s, e]} is a part (a part of the voice) of the input voice that has the start time s and the end time e. σ (v) is the utterance start time of the word v (denoted as w _a , w _b etc. in Equation 25). Let τ (v) be the utterance end time. Let f _am (x _{[s, e]} | v) be the acoustic score when the word v is given. In this case, the acoustic score of the compound word string is expressed by the following (Equation 26).

  Subsequently, the entropy calculation unit 25 creates the m-th compound word from all the word strings in the word string set for the n-th speech x using the following (Expression 27) and (Expression 28). The previous entropy and the entropy after creation are calculated (step S60).

Subsequently, the entropy calculation unit 25 generates entropy H _n ^m (w | x) before creating the m-th compound word for the calculated n-th speech x and entropy H (bar) _n ^m ( The difference from w | x) is obtained by the following (formula 29) (step S70).

When the processing up to step S70 is completed for the nth speech n, “1” is added to _n, and the processing from step S40 to step S70 is performed for the next speech _{xn + 1} . The loop process in step S30 is performed for the given N sounds (all sounds). That is, the entropy H ^m (w | x) for the m-th compound word candidate is obtained by the following (Equation 30) by accumulating with N voices (all voices).

Then, by the loop processing in step S30, the entropy difference ΔH _{m with} respect to the m-th compound word candidate is integrated by N voices (all voices) to be (Equation 31) below.

When the loop processing of step S30 is completed for the given N sounds, “1” is added to m, and the processing from step S20 to step S70 is also performed for the next m + 1-th compound word candidate _{m + 1.} Do. The loop processing in step S10 is performed for the given high frequency M compound word candidates. Then, the entropy calculation unit 25 outputs the entropy difference ΔH _m for the high frequency M compound word candidates to the compound word selection unit 26.

(Processing procedure of language model update unit)
With reference to FIG. 4, the process procedure of the language model update part 27 is demonstrated.
First, the language model update unit 27 adds the compound word w _c selected by the compound word selection unit 26 to the original vocabulary {V} stored in the vocabulary / language model storage unit 12 ({V} ← {V } ∪ {w _c })) The vocabulary is updated (step S110).

  Subsequently, the language model update unit 27 replaces a pair of words before the compound word creation included in the learning data stored in the language model learning data storage unit 11 with the compound word, and updates the learning data (step S120). . Specifically, the language model update unit 27 updates word breaks (morpheme breaks) of text data that is learning data. Here, assuming that the word break is updated in Japanese, a word pair such as “independent word A + independent word B” or “independent word A + ancillary word C” included in the text data before the update is combined with “composite word”. The word “AB” or “compound word AC”.

Subsequently, the language model updating unit 27 selects the language score table (stored in the language score table storage unit 15) obtained by the language score calculation performed by the entropy calculation unit 25 in step S20 (see FIG. 3). Replace the language score (bigram) corresponding to the compound word. Further, the language model update unit 27 creates a language model by statistical processing, and updates the language model stored in the vocabulary / language model storage unit 12 (step S130). Then, the language model update unit 27 outputs to the alignment unit 23 that the language model has been updated.
Above, description about the process sequence of the dictionary creation apparatus 100 which concerns on embodiment is complete | finished.

  As described above, the dictionary creating apparatus 100 according to the present embodiment registers a compound word in the dictionary from the correct word sequence of speech and the correct candidate word sequence that is the result of speech recognition by the speech recognition unit 21. Since the reference entropy is calculated, a compound word to be registered in the dictionary can be determined in consideration of acoustic elements. Therefore, the accuracy of word prediction in the statistical language model for speech recognition is improved, and recognition errors are reduced as compared with the conventional statistical language model.

[Modification]
As mentioned above, although embodiment of this invention was described, this invention is not limited to this, It can implement in the range which does not change the meaning.

  For example, the dictionary creating apparatus 100 of the embodiment has been described assuming that Japanese is registered in the dictionary, but the present invention can also be applied to other languages such as English.

  Moreover, although the dictionary creation apparatus 100 of embodiment demonstrated the case where the bigram which is one of the n-gram models was used as a language model, it used a trigram etc., or using another language model. Also good.

11 Language Model Learning Data 12 Vocabulary / Language Model Storage Unit (Storage Unit)
13 Acoustic model storage unit (storage unit)
14 Spoken language resource storage unit 15 Language score table storage unit 21 Speech recognition unit (speech recognition means)
22 Correct word string alignment unit (correct word string aligning means)
23 Alignment part (alignment means)
24 Compound word frequency counting unit (compound word frequency counting means)
25 Entropy calculation part (entropy calculation means)
26 Compound Word Selection Unit (Compound Word Selection Unit)
27 Language Model Update Unit (Dictionary Registration Unit (Dictionary Registration Unit))
100 dictionary creation device

Claims (4)

A dictionary creation device for registering a compound word used for speech recognition in a dictionary as a new vocabulary,
A spoken language resource storage unit in which a voice and a correct word string obtained by characterizing the voice without error are stored;
A storage unit storing the dictionary, a statistical language model indicating a degree of connection between words in the speech, and a statistical acoustic model indicating a relationship between the words in the speech and a waveform pattern;
A speech recognition unit that recognizes the speech using the statistical language model and the statistical acoustic model, and generates a correct candidate word string indicating a word string of a correct answer candidate of the speech;
An alignment unit for aligning the correct word string and the correct candidate word string in accordance with an utterance time;
The frequency of each pair of words included in the correct word string and the correct candidate word string aligned by the aligning unit is counted, and the pair of words is stored in the dictionary according to the counted frequency. A compound word frequency counting unit that extracts a plurality of compound word candidates of the compound word to be registered;
In the case where each compound word candidate is defined as the compound word using the plurality of compound word candidates extracted by the compound word frequency counting unit and the correct word string and the correct candidate word string aligned in the aligning unit . An entropy calculating unit that calculates an entropy indicating that the speech recognition is difficult as the value is higher, an index indicating difficulty of speech recognition of the speech;
A compound word selection unit that selects a compound word candidate that reduces the entropy most when each compound word candidate is the compound word;
A dictionary registration unit for registering the compound word candidate selected by the compound word selection unit as a new vocabulary of the dictionary;
A dictionary creation device comprising: The entropy calculator is
Calculating the entropy using a language score obtained by the statistical language model and an acoustic score obtained by the statistical acoustic model;
The dictionary creation device according to claim 1. The spoken language resource storage unit sequentially accumulates broadcast voice and its correct word string,
The dictionary registration unit gradually registers a compound word as a new vocabulary.
The dictionary creation device according to claim 1 or claim 2, wherein A dictionary creation program that causes a computer to function as a dictionary creation device that registers a compound word used for speech recognition in a dictionary as a new vocabulary,
The computer
A spoken language resource storage unit in which speech and correct word strings obtained by characterizing the speech without error are stored; the dictionary; a statistical language model indicating the degree of connection between words in the speech; and the words in the speech And a storage unit storing a statistical acoustic model indicating the relationship between the waveform pattern and
The computer,
Speech recognition means for recognizing the speech using the statistical language model and the statistical acoustic model, and generating a correct candidate word string indicating a word string of a correct candidate for the speech;
An alignment means for aligning the correct word string and the correct candidate word string in accordance with an utterance time;
A compound word that counts the frequency of each pair of words included in the aligned correct word string and the correct candidate word string and registers the word pair in the dictionary according to the counted frequency Compound word frequency counting means for extracting as a plurality of compound word candidates,
An index indicating difficulty in speech recognition of the speech when each compound word candidate is the compound word using the extracted plurality of compound word candidates and the aligned correct word strings and the correct candidate word strings. Entropy calculating means for calculating entropy indicating that speech recognition is more difficult as the value is higher,
Compound word selection means for selecting a compound word candidate that reduces the entropy most when each compound word candidate is the compound word;
A dictionary creation program for causing the selected compound word candidate to function as dictionary registration means for registering as a new vocabulary of the dictionary.

Images