JP2965529B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for recognizing speech based on a speech signal of an uttered speech by referring to a statistical language model.

[0002]

2. Description of the Related Art In a continuous speech recognition apparatus, N-gra
m, a statistical language model based on a statistical method is widely used (for example, see LRBa
hl etal., “A Maximum Likelihood Approach to Contin
continuous Speech Recognition ", IEEE Transactions on Patt
ern Analysis and Machine Intelligence, pp.179-190,1
983 ". ). In the continuous speech recognition device using N-gram, the transition probability of transition from the immediately preceding N-1 words to the next word is learned in advance using large-scale learning data, and at the time of speech recognition, By predicting the next word to be connected using the learned transition probability, the speech recognition rate is improved. Generally, as N increases, the prediction accuracy of the next word improves, but since the number of types of word chains increases, a large amount of learning data is required to obtain a reliable transition probability. Therefore, at present, N is 2 (bi-g
ram) or 3 (tri-gram). However, the word bi-gr
Analysis of the results of continuous speech recognition using "am" and "tri-gram" of words indicates that even if the local word chain within two or three words has naturalness, the whole sentence is unnatural. There are many cases where false recognition sentences are output as a result, and it is considered that more global language constraints are required.

[0003] Language models have been proposed that enable more global constraints using grammar such as context-free grammar and dependencies between words. However, considering the structure of spontaneously spoken sentences and the diversity of dependencies, it is not easy to construct rules and dependencies, and the amount of processing is enormous. On the other hand, a method of disambiguating the syntax of a sentence by an example-driven approach (hereinafter referred to as a conventional example) is described in prior art document 2 “Eiichiro Sumida et al. , IEICE Transactions (D-II), J77-
D-II, No3, pp. 557-565, March 1994 ". This conventional method is
This is a method of extracting an example from the corpus, calculating the semantic distance between the expression of the input sentence and the example according to a thesaurus, and selecting a syntax that minimizes the final semantic distance. (Prior Art Document 3, "Kurase Kura et al.," Transformation-Driven Machine Translation Utilizing Empirical Knowledge ", Transactions of Information Processing Society of Japan, Vol. 35, N.
o3, pp. 414-423, March 1994 ". ).

[0004]

However, in a speech recognition apparatus using the conventional method, for example, if an unnatural syntax is input for a learned example, the semantic distance from any example increases. However, there is a problem that the voice recognition rate is relatively low.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above problems, remove improper misrecognition results, and output a sentence that is locally and globally appropriate. An object of the present invention is to provide a speech recognition device that can obtain a high speech recognition rate.

[0006]

According to a first aspect of the present invention, there is provided a speech recognition apparatus which refers to a predetermined statistical language model on the basis of a speech signal of an uttered speech sentence composed of an input word string. And a voice recognition means for performing voice recognition processing on the voice, wherein the voice recognition means determines a degree of disqualification of the voice recognition candidate with respect to the voice recognition candidate. The function value of the ineligible sentence determination function is calculated using the ineligible sentence determination function, and when the calculated function value exceeds a predetermined threshold, the voice recognition candidate is removed and the voice recognition process is performed. It is characterized by the following.

According to a second aspect of the present invention, in the speech recognition apparatus of the first aspect, the function value of the ineligible sentence determination function is a semantic distance corresponding to an example used in the speech recognition processing. Is calculated by multiplying the calculated sum by the number of morphemes included in the speech recognition candidates to be subjected to speech recognition processing and dividing by the number of examples used in the speech recognition processing. And Further, in the speech recognition apparatus according to claim 3, in the speech recognition apparatus according to claim 1, the function value of the inappropriate sentence determination function is a sum of semantic distances corresponding to the examples used in the speech recognition processing. Calculate and calculate the average value of the semantic distance, which is the value obtained by dividing the calculated sum by the number of examples used in the voice recognition processing, and calculate the average value of the semantic distance as the target of the voice recognition processing. And a value obtained by multiplying by the number of morphemes included in the speech recognition candidate and dividing by the number of examples used in the speech recognition processing. Still further, the speech recognition device according to claim 4 is the speech recognition device according to claim 1, wherein the function value of the inappropriate sentence determination function is a sum of semantic distances corresponding to the examples used in the speech recognition processing. Is calculated, and the calculated sum is divided by the number of examples used in the above speech recognition processing to calculate the average value of the semantic distance, and the calculated average value of the semantic distance is determined by a predetermined number of morphemes. Is a value obtained by dividing by the number of examples including a plurality of predetermined morphemes in the examples used in the speech recognition processing at the stage of processing.

According to a fifth aspect of the present invention, in the speech recognition apparatus according to any one of the first to fourth aspects, the threshold value is preferably a constant value. Further, the speech recognition device according to claim 6 is the speech recognition device according to any one of claims 1 to 4, wherein the threshold value is preferably included in a partial sentence to be subjected to speech recognition processing. To be changed depending on the number of morphemes.

[0009]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a speech recognition apparatus according to an embodiment of the present invention. The speech recognition apparatus includes a microphone 1, a feature extraction unit 2, a buffer memory 3, and a hidden Markov model memory (hereinafter, referred to as an HMM memory) 5 based on input uttered speech data.
Hidden Markov model (hereinafter referred to as HM)
It is called M. ), And outputs a phoneme data by executing a phoneme verification process, and One pass DP (Vite) based on the phoneme data from the phoneme verification unit 4.
A database memory of statistical language models and examples and distances in the statistical language model memory 7 using an rbi search algorithm (referred to as a database memory).
One Pa, which executes speech recognition with reference to the example and distance database (hereinafter referred to as database) in FIG.
ss DP speech recognition unit (hereinafter referred to as speech recognition unit) 6
The speech recognition unit 6 performs an ineligible sentence determination function on the speech recognition candidate by using a predetermined ineligible sentence determination function (Equation 1 described later in detail) indicating the degree of ineligibility for the speech recognition candidate. Is calculated, and when the calculated function value exceeds a predetermined threshold value Fth, the speech recognition candidate is removed to perform speech recognition. Here, the function value of the disqualified sentence determination function is preferably calculated by calculating the sum of the semantic distances corresponding to the examples used to determine the syntax of the speech recognition candidate, and adding the calculated sum to the speech recognition. This is a value obtained by multiplying the number of morphemes included in the speech recognition candidate to be processed and dividing by the number of examples used to determine the syntax of the speech recognition candidate. Further, the threshold value Ft
h is preferably changed depending on a fixed value or the number of morphemes included in the partial sentence to be subjected to the speech recognition processing. The morpheme is a minimum unit of a character sequence having a meaning such as a stem, a prefix, and a suffix, and is a unit substantially the same as or slightly smaller than a word.

First, a description will be given of an improper sentence detection method in the voice recognition unit 6. Misrecognition in speech recognition processing using N-gram has the following features. (A) Judging from a chain of N or more words, there are grammatically and semantically inappropriate word combinations. For example, an incorrect recognition example: "Is the phone number room 2107?" (B) The sentence structure is not organized in large units. That is, the rules cannot be applied grammatically, and only local decisions can be made. For example, a misrecognition example: "Even up to three people staying in a single night ..."

In order to solve the misrecognition having the above characteristics, it is necessary to judge the consistency between words and the suitability of the syntax from a broader perspective than N-gram. on the other hand,
Example-driven speech translation method (prior art document 2 and prior art document 4 “O. Furuse et al.,“ Incremental Translatio ”
n Utilizing Constituent Boundary Patterns ", Proceed
ings of Coling '96, 1996 ". ) Uses left-to-right (left-to-right)
2) The method of determining the syntax is used. In this process, in order to resolve the ambiguity of the syntax, a method of calculating the semantic distance between the input sentence and the example using a thesaurus (synonym dictionary) and selecting a syntax corresponding to the example with a small distance is adopted. ing. The present inventor believes that the above syntax determination method has good consistency in removing erroneous recognition based on the conventional N-gram language model for the following reasons. (A) Since the above syntax determination method is an example-driven method,
Syntax that is difficult to process with conventional grammar, such as that found in conversational sentences, can be easily processed. (B) In the above syntax determination method, since the semantic distance is obtained based on the syntax, there is an ability to determine consistency between non-adjacent words. (C) Since both speech recognition and the above syntax determination method are performed from left to right (left-to-right),
Intermediate results up to a certain point in time may be determined sequentially.

Therefore, an improper sentence is detected based on the consistency of the semantic distance viewed from a global perspective and the suitability of the analyzed syntax. Specifically, the determination is made as follows. First, the inconsistency of the semantic distance in the partial sentence is determined based on the semantic distance value used in the above syntax determination method. When the sum of the semantic distances of a certain partial sentence becomes equal to or greater than a certain value, the sentence is determined to be erroneously recognized. Next, we consider the eligibility of the syntax as follows. It is assumed that a natural sentence composed of a certain number of morphemes has a coherent syntax, and that the structure of the syntax is a somewhat complicated structure. Consider the ratio (= m / R) of the number m of morphemes included in a certain partial sentence to the rule number R (or the number of examples) of the rules of the context-free grammar or the examples used for the syntax determination.
A sub-sentence without a coherent syntax does not have a hierarchical syntax structure. Therefore, the number R of used syntax rules is small and the m / R value is large for the number m of morphemes. Conversely, the more complex and hierarchical the syntax, the smaller the m / R value.
Therefore, the following equation defines an ineligible sentence determination function F _error (m).

[0014]

(Equation 1)

Here, d (r _i ) is a semantic distance or similarity distance corresponding to a plurality of examples or rules r _i , and m is included in a partial sentence of a speech recognition candidate to be subjected to speech recognition processing. R is the number of morphemes, and R is the number of examples or rules used to determine the syntax of the partial sentence of the speech recognition candidate when executing the speech recognition process. Here, the semantic distance or the similarity distance is, for example, p. This is a semantic distance between the speech recognition candidate of the input utterance speech sentence and the example, which is defined by Expression (1) of 559 and calculated using a thesaurus. In the present embodiment, the partial sentence of the speech recognition candidate is The distance to the example in the database corresponding to is searched and determined. Here, the thesaurus refers to a dictionary in which the upper-lower relationships between concepts are expressed in a tree structure, and words having the concepts are assigned to the lowest-level concepts corresponding to leaves. The semantic distance between words is defined by the semantic distance between concepts on the thesaurus, and the distance between concepts is set to a value from 0 to 1 according to the position of the smallest common superordinate concept in the thesaurus. A value of 0 means that the two concepts are the same,
A value of 1 means irrelevant. The judgment function F _error (m) is a function of the number of morphemes m, and is calculated for partial sentences of speech recognition candidates from the beginning of the sentence to the m-th morpheme. When the determination function value F _error (m) exceeds a predetermined threshold value Fth, the speech recognition unit 6 determines that the partial sentence of the speech recognition candidate is an erroneous recognition result and removes the partial sentence from the speech recognition candidate. The above equation 1 is preferably
This can be applied when m ≧ 5. The above equation (1)
When the number of rules R is 0, the function value is set to 1 and the result is determined to be an erroneous recognition result and is removed from the speech recognition candidates.

In the preferred embodiment shown in FIG. 1, the database generating unit 10 uses a predetermined similarity rule based on the examples in the example memory 11 and the word sets in the word set memory 12 to store the database. Generated and stored in the database memory 8. Tables 1 and 2 show examples of context-free grammar rules. Table 3 shows an example of the similarity rule.

[0017]

[Table 1] Example 1 Ｙ X Y ───────── My child Your company ……… ─────────

[0018]

[Table 2] Example 2 が X is Y ───────── I am a teacher ……… ─────────

[0019]

[Table 3] Similarity rule ─────────────────────────────────── (I) A set of word sets When the generated sentence is the same as the example, distance = 0. (II) When a sentence generated by a set of word sets has the same function word as the example (for example, “no” or “ga”) and has a word set of the same similar category as the example word, distance = 10 ^-5 . (III) When sentences generated by the set of word sets are words that are not in the example, distance = 0.5. ───────────────────────────────────

FIG. 2 shows an example of the word sets S1, S2, S3, and S4 in the speech recognition apparatus for Japanese processing, and the distance between the word sets when a predetermined functional word is used. In FIG. 2, for example, “you” (word set S
When 1) is “teacher” (word set S2), the distance is 10
^-5 , "child" of "you" (word set S1)
In the case of (word set S3), the distance becomes 10 ^-5 , and the “company” (word set S) of “you” (word set S1)
In the case of 4), the distance becomes 10 ^-5 . In the case of “teacher” (word set S2), “company” (word set S4),
The distance becomes 0.5.

The database generation unit 10 performs the operations shown in Tables 1 and 2
And a database generation process using the similarity rule shown in Table 3 are performed as follows. A partial sentence is generated from each set of word sets. If the partial sentence is "your company", the distance is 0, and if the partial sentence is "my school", the distance is 10 ^-5 . When the partial sentence is "child is a teacher", the distance is 0.5. Generated like this,
A database of examples of partial sentences and distances is stored in the database memory 8.

Further, the statistical language model generates, for example, a bi-gram statistical language model of the word based on the text data of the uttered voice sentence and stores it in the statistical language model memory 7. I do.

Next, the configuration and operation of the speech recognition apparatus using the statistical language model of the present embodiment will be described.

In FIG. 1, a speaker's uttered voice is input to a microphone 1, converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the phoneme matching unit 4 via the buffer memory 3.
HMM in HMM memory 5 connected to phoneme matching unit 4
Is composed of a plurality of states and arcs indicating transitions between the states. Each arc has a transition probability between states and an output probability for an input code. The phoneme matching unit 4 performs phoneme matching processing based on the input data, and outputs phoneme data to the speech recognition unit 6.

A statistical language model memory 7 for storing a statistical language model in advance is connected to the speech recognition unit 6. The speech recognition unit 6 refers to the statistical language model in the statistical language model memory 7 and the database in the database memory 8 and uses a predetermined One pass DP algorithm to move the input phoneme data from left to right. To
Processing is performed without regression to recognize a word having a higher probability of occurrence as a speech recognition candidate, and a decision function value F _error (m) is calculated for the speech recognition candidate using _Equation 1 above. Here, d (r _i ) in Equation 1 is obtained by searching a database for an example corresponding to a voice recognition candidate, and setting a distance corresponding to the searched example as a semantic distance. When the calculated determination function value F _error (m) exceeds a predetermined threshold value Fth, the speech recognition unit 6 determines that the partial sentence of the speech recognition candidate is an erroneous recognition result, and Remove from Then, the remaining speech recognition candidates are determined as speech recognition results (character string data) and output.

FIG. 3 is an operation diagram showing the operation of the speech recognition apparatus for Japanese language processing configured as described above. The input sentence, the recognition result sentence, its syntax tree and score, and the syntax analysis result sentence are shown. FIG. 4 is an operation diagram showing a syntax tree and a score thereof. FIG.
As shown in FIG. 3A, when the voice of the input sentence "My Eto school is completed" is input, as shown in FIG. 3B, "My Eto school is completed" as a recognition result sentence. Is obtained, that is, when the interjection "Eto" is mistakenly recognized as a noun "Eto". FIG. 3B shows scores between words in the recognition result sentence. Further, when the syntax is analyzed based on the recognition result sentence, as shown in FIG. 3C, a syntax tree of the syntax analysis result is obtained based on the smaller score, and the score at this time is obtained. In the case of FIG.
To the function value F of the ineligible sentence determination function
_error (m) is as follows.

[0027]

F _error (m) = (6/3) (0.5 + 0.5 + 10 ⁻⁵ ) = 2 × (1.00001) = 2.00002

In this example, the threshold value Fth for judging an improper sentence is preferably 0.6 to 0.7, and the function value = 2.00002 calculated by the above equation (2) is a threshold. Since the value exceeds the value Fth, the corresponding speech recognition candidate is removed from the speech recognition candidates. The threshold value Fth may be a constant value or may vary depending on the number of morphemes m included in the partial sentence to be subjected to the speech recognition processing.

In the speech recognition apparatus configured as described above, the feature extraction unit 2, the phoneme collation unit 4, the speech recognition unit 6
And the database generation unit 10 are configured by a computer such as a digital computer, for example, and include a buffer memory 3, an HMM memory 5, and a statistical language model memory 7.
And the database memory 8 are configured by a storage device such as a hard disk memory, for example.

Next, an example of a speech recognition device for English processing will be described. Tables 4 and 5 show examples of context-free grammar rules in English processing. For example, the similarity rule shown in Table 3 is used as it is.

[0031]

[Table 4] Example 11 {X at Y} start at 7:30 leave at 6p . m. ……………………… ────────────────

[0032]

[Table 5] Example 12 ─────────────────Z ・ X ｅthe train starts ………… …………… ─────────────────

Word sets S11 (X), S12, S13 (Z), S14 in the English language speech recognition device
FIG. 5 shows an example of (Y) and the distance between the word sets when a predetermined function word is used. In FIG. 5, for example, when “train leaves”, the distance is 10 ^−5.
, And the distance is 0 when “leave train”.
It becomes 5. In the case of “leave Kyoto”, the distance becomes 10 ⁻⁵ , and “leave at 6 p.
m. ", The distance becomes 10 ^-5 . The database generation unit 10 performs a database generation process using the examples in Tables 4 and 5 and the similarity rule in Table 3 as follows.
A sub-sentence is generated from each set of word sets, and the sub-sentence is "th
When "e train starts", the distance is 0, and the sub-sentence is "the bus leaves".
, The distance is 10 ⁻⁵ and the partial sentence is “lea
ve yacht, the distance is 0.5. The database of the example of the partial sentence and the distance thus generated is stored in the database memory 8.

FIG. 6 is an operation diagram showing the operation of the speech recognition apparatus for English processing configured as described above. The input sentence, the recognition result sentence, its syntax tree and score, the syntax analysis result sentence, FIG. 6 is an operation diagram showing the syntax tree and a score. FIG.
As shown in (a), “Thebus leaves”
Kyoto at 11 a. m. When the voice of the input sentence “” is input, as a recognition result sentence, as shown in FIG. 6B, “Thebus leaves yach”.
tat 11 a. m. Is obtained, that is, the proper noun of the place name “Kyoto” is “yach
This is the case where the noun “t” is incorrectly recognized. FIG. 6B shows scores between words in the recognition result sentence. Further, when the syntax is analyzed based on the recognition result sentence, as shown in FIG. 6C, a syntax tree of the syntax analysis result is obtained based on the smaller score, and the score at this time is obtained. In the case of FIG.
To the function value F of the ineligible sentence determination function
_error (m) is as follows.

[0035]

F _error (m) = (5/4) (10 ⁻⁵ + 0.5 + 0.5 + 10 ⁻⁵ ) = 1.25 × (1.00002) = 1.250025

In this example, since the function value = 1.250025 calculated by the above equation 3 exceeds the threshold value Fth, the corresponding speech recognition candidate is removed from the speech recognition candidates.

[0037]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present inventor conducted an experiment as follows in order to evaluate the effectiveness of a speech recognition apparatus provided with the above-described unsuitable sentence detection method. Here, the above-mentioned improper sentence determination function F
_error is, it was confirmed whether it is possible to distinguish the erroneous recognition sentence and a correct sentence in the recognition experiments using the N-gram language model. Specifically, an incorrectly qualified sentence determination function F _error is calculated for the incorrectly recognized result sentence and the correct sentence by the recognition system using the bi-gram, and an incorrectly qualified sentence determination function of the incorrectly recognized sentence and the correct answer sentence is calculated. The difference in the function value F _error was considered.
The correct sentence, the number m of morphemes is large, i.e. the longer the partial sentence, since the sentence structure also decreases ambiguity of the structure becomes complex will function value F _error is small, the distinction between misrecognition statement is likely attached I can imagine. However, considering the efficiency of the recognition process, the number of morphemes m
It is preferable to perform an improper judgment on the partial sentence of the speech recognition candidate at the stage where the is small, and remove the improper sentence from the result candidate as an erroneously recognized sentence. In order to know the number m of morphemes for obtaining a highly reliable function value F _error , a function value F _error is calculated for a partial sentence of a speech recognition candidate up to the m-th morpheme of a misrecognized or correct sentence, The change in the function value F _error when the number m of morphemes was changed was also examined. Table 6 shows the speech recognition and data conditions in the experiment.

[0038]

[Table 6] Voice recognition and data conditions ─────────────────────────────────── Task Voice guidance for travel guidance Database ─────────────────────────────────── Acoustic model Unspecified speaker HM-net, 401 state, 10 Mixture distribution ─────────────────────────────────── language model bi-gram of words 混合───────────────────────────── Voice recognition method One-pass DP, N-best search ───────── ────────────────────────── bi-gram learning data 3363 sentences, 222954 words ────────────── ─────────────── ────── Evaluation data 44 sentences included in the training data, 4 speakers ────────────────────────────── ─────

In the speech recognition processing, a bi-gram of words was used as a statistical language model, and a one-pass DP algorithm and an N-best search type speech recognition system were used. The evaluation data shown in Table 6 is used as a correct answer sentence, and the evaluation data described above is recognized by a recognition system using three types of N-grams shown in Table 6 as an erroneously recognized sentence, and the result is erroneously recognized. Sentence 94 sentences were used. FIG. 4 shows the average value and the maximum value of the function value F _error for the correct sentence, and the function value F _error for the incorrectly recognized sentence for each morpheme number m. The following can be understood from FIG. (A) For the correct answer sentence, as the morpheme number m becomes longer, both the average value and the maximum value of the function value _Ferror decrease. (B) Similarly, in the misrecognized sentence, the function value F _error tends to decrease as the number of morphemes increases, but the degree of the decrease is smaller than that of the correct answer sentence.

This means that from left to right (left-to-righ
In the (t) speech recognition processing system, there is no difference in the function value _Ferror between the correct sentence and the incorrectly recognized sentence at the beginning of the sentence where the processed morphemes are still small, and it is difficult to detect an improper sentence. As the number of morphemes increases, a difference occurs in the function value _Ferror between the correct sentence and the incorrectly-recognized sentence. Thus, by appropriately setting the threshold value Fth, it is possible to detect an improper sentence. ing. However, it can be seen that the threshold value Fth is not a constant value, and it is more effective to define the threshold value Fth as a function value using the morpheme number m as a variable. For example, when the maximum value in FIG. 4 is set to the threshold value Fth, a sentence indicating a function value F _error equal to or larger than the threshold value Fth is not processed when each morpheme number m is processed. It can be determined that the sentence is eligible. In this experiment, the ratio of sentences that could be determined to be misrecognized from the intermediate results of the sentences was 47.9% (= 45/94) of all misrecognized sentences in this experiment. The above results are summarized as follows. Two parameters, (1) the semantic distance between the input word and the example and (2) the complexity of the sentence structure represented by the rule number for the morpheme number, used for detection of the unqualified sentence are ineligible. It is a parameter effective for judging a sentence, and it has been found that the proposed improper sentence determination function F _error is effective for detecting an improper sentence. (B) The performance of improper sentence detection depends on the morpheme number m, and the detection performance increases as m increases. (C) The threshold value Fth of the ineligible sentence determination function F _error is
If it is changed depending on the morpheme number m, an unsuitable sentence can be detected more efficiently.

As described above, according to the present invention, a syntax determining method for determining a syntax while eliminating ambiguity of a syntax by using a semantic distance from an example is used.
We have invented a method of sequentially detecting ineligibility of a sentence resulting from false recognition of speech recognition using a conventional statistical language model. In this method, the semantic distance between a phrase included in a partial sentence of a recognition result and a previously learned example and the complexity of the syntax of the partial sentence of the recognition result are used as factors for determining an improper sentence. Then, as a result of the detection of an improper sentence on the result of the recognition system using the bi-gram of various words and parts of speech, a threshold value Ft for judging a misrecognized sentence and a correct sentence is obtained.
It has been found that if h is set appropriately, about half of the misrecognized sentences can be detected as improper sentences.

Therefore, the speech recognition unit 6 calculates a function value of the unqualified sentence determination function using a predetermined unqualified sentence determination function indicating the degree of disqualification of the speech recognition candidate, and When the calculated function value exceeds a predetermined threshold, the speech recognition candidate is removed and speech recognition is performed, so that an inappropriate erroneous recognition result can be removed, and the result can be locally or globally. It is possible to provide a speech recognition device that can output a qualified sentence and obtain a higher speech recognition rate than the conventional example.

In the above embodiment, the expression 1 is used as the ineligible sentence determination function. However, the present invention is not limited to this, and the ineligible sentence determination function of the following expression 4 or 5 may be used. Good.

(Equation 4)

(Equation 5)

Here, the ineligible sentence determination function F _error '
(M) calculates the sum of the semantic distances corresponding to the example used to determine the syntax of the speech recognition candidate, as compared with the ineligible sentence determination function F _error (m) in _Equation 1. An average value of semantic distances, which is a value obtained by dividing the obtained sum by the number of examples used to determine the syntax of the speech recognition candidate, is calculated. In Equation 5, M is
At the stage of processing the predetermined m morphemes, the rule of the example used to determine the syntax of the speech recognition candidate represents the number of rules including _a predetermined plurality of ma morphemes or more, where m
It is preferably a 5 or more, the m _a is preferably 3. The ineligible sentence determination function F _error ″ (m) in _Expression 5 is
Compared to the ineligible sentence determination function F _error '(m),
It is characterized in that the reciprocal of the rule number M is used instead of R). By performing speech recognition using the unsuitable sentence determination function of Equations 4 and 5, it is possible to remove inappropriate unrecognized results, and to output locally and globally appropriate sentences. A speech recognition device that can obtain a higher speech recognition rate than the example can be provided.

[0045]

As described in detail above, according to the speech recognition apparatus of the first aspect of the present invention, a predetermined statistical analysis is performed on the basis of the speech signal of the uttered speech sentence composed of an input word string. A speech recognition unit that performs speech recognition processing on the speech with reference to a language model, wherein the speech recognition unit determines a degree of disqualification of the speech recognition candidate with respect to the speech recognition candidate. Calculates the function value of the ineligible sentence determination function using a predetermined ineligible sentence determination function representing a function, and when the calculated function value exceeds a predetermined threshold value, removes the speech recognition candidate and removes the speech recognition candidate. Execute the process. Therefore, it is possible to provide a speech recognition device that can remove improper misrecognition results, can output locally and globally proper sentences, and can obtain a higher speech recognition rate than the conventional example. can do.

According to the second aspect of the present invention, in the first aspect, the function value of the ineligible sentence determination function may be a semantic value corresponding to the example used in the speech recognition process. This is a value obtained by calculating the sum of the distances, multiplying the calculated sum by the number of morphemes included in the speech recognition candidates to be subjected to the speech recognition processing, and dividing by the number of examples used in the speech recognition processing. Therefore, it is possible to easily calculate the function value of the unsuitable sentence determination function, to remove an inappropriate unrecognized result, and to output a sentence that is locally and globally appropriate. It is possible to provide a speech recognition device that can obtain a higher speech recognition rate as compared with the first embodiment.

Further, in the speech recognition apparatus according to the third aspect, in the speech recognition apparatus according to the first aspect, the function value of the inappropriate sentence determination function is a semantic corresponding to the example used in the speech recognition processing. The sum of the distances is calculated, and the calculated sum is divided by the number of examples used in the above speech recognition processing to calculate an average value of the semantic distances. This is a value obtained by multiplying the number of morphemes included in the speech recognition candidate to be processed and dividing by the number of examples used in the above speech recognition processing. Therefore, it is possible to easily calculate the function value of the unsuitable sentence determination function, to remove an inappropriate unrecognized result, and to output a sentence that is locally and globally appropriate. It is possible to provide a speech recognition device that can obtain a higher speech recognition rate as compared with the first embodiment.

According to a fourth aspect of the present invention, in the voice recognition apparatus of the first aspect, the function value of the inappropriate sentence determination function has a meaning corresponding to the example used in the voice recognition processing. Calculate the sum of the semantic distance, calculate the average of the semantic distance that is the value obtained by dividing the calculated sum by the number of examples used in the speech recognition processing, and calculate the average of the calculated semantic distance, This is a value obtained by dividing the number of examples including a plurality of predetermined morphemes in the examples used in the speech recognition processing at the stage when a predetermined number of morphemes have been processed. Therefore, it is possible to easily calculate the function value of the unsuitable sentence determination function, to remove an inappropriate unrecognized result, and to output a sentence that is locally and globally appropriate. It is possible to provide a speech recognition device that can obtain a higher speech recognition rate as compared with the first embodiment.

Further, in the speech recognition apparatus according to the fifth or sixth aspect, in the speech recognition apparatus according to any one of the first to fourth aspects, the threshold value is preferably a constant value or It is changed depending on the number of morphemes included in the partial sentence to be subjected to the speech recognition processing. Therefore, it is possible to more effectively remove the incorrect recognition result,
It is possible to provide a speech recognition device that can output a sentence that is qualified both locally and globally and that can obtain a higher speech recognition rate than the conventional example.

[Brief description of the drawings]

FIG. 1 is a block diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a relationship between a Japanese word set and a distance in the voice recognition device of FIG. 1;

FIG. 3 is an operation diagram showing an operation of the Japanese language processing of the speech recognition apparatus of FIG. 1, including an input sentence, a recognition result sentence, its syntax tree and a score, a syntax analysis result sentence, its syntax tree and a score, FIG.

4 is a simulation result of the speech recognition apparatus of FIG. 1, showing a decision function value F with respect to the number of input morphemes;
_It is a graph which shows _error .

FIG. 5 is a diagram showing a relationship between an English word set and a distance in the voice recognition device of FIG. 1;

6 is an operation diagram showing an operation of the English language processing of the speech recognition device of FIG. 1, wherein an input sentence, a recognition result sentence, its syntax tree and a score, and a syntax analysis result sentence, its syntax tree and a score are FIG.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3 ... Buffer memory, 4 ... Phoneme collation part, 5 ... Hidden Markov model memory (HMM memory), 6 ... One pass DP speech recognition part, 7 ... Statistical language model memory, 8 ... a database memory (database memory) of examples and distances, 10 ... a database generator, 11 ... example memories, 12 ... word set memories.

──────────────────────────────────────────────────続 き Continued on the front page (56) References Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J77-D-II, No. 3, March 1994, "Example-driven ambiguity resolution of English prepositional phrase endings" p. 566-572, published March 25, 2006, Japanese Society for Artificial Intelligence Language and Speech Understanding and Dialogue Processing, Vol. 1 "Method of specifying correct part of speech recognition using semantic similarity and application to speech translation method" p. 7-12, 1997 IPSJ Transactions Vol. 35, No. 3 "Conversion-driven machine translation utilizing empirical knowledge" p. 414-425 Published on March 15, 1994 IEEE Transactions on Pattern Analysis and Machine Intelli- gen, Vol. PAM I-5, No. 2, March 1983, "A Maximum Like life Approach to Continuous Speech Recognition", p. 179 −190 (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 3/00 537 G10L 3/00 531 JICST file (JOIS)

Claims (6)

(57) [Claims] 1. A speech recognition means for executing speech recognition processing on said speech by referring to a predetermined statistical language model based on a speech signal of an uttered speech sentence composed of an input word string. In the speech recognition device provided with the speech recognition unit, the speech recognition unit calculates a function value of an unqualified sentence determination function using a predetermined unqualified sentence determination function indicating a degree of disqualification of the speech recognition candidate. When the calculated function value exceeds a predetermined threshold value, the speech recognition device removes the speech recognition candidate and executes a speech recognition process. 2. The function value of the ineligible sentence determination function calculates a sum of semantic distances corresponding to the examples used in the speech recognition process, and calculates the sum to be a target of the speech recognition process for the speech recognition process. The speech recognition apparatus according to claim 1, wherein the value is obtained by multiplying the number of morphemes included in the candidate and dividing the result by the number of examples used in the speech recognition processing. 3. The function value of the unsuitable sentence determination function is obtained by calculating the sum of semantic distances corresponding to the example used in the speech recognition process, and using the calculated sum in the speech recognition process. Calculate the average value of the semantic distance, which is the value divided by the number,
The average value of the calculated semantic distances is multiplied by the number of morphemes included in the speech recognition candidates to be subjected to the speech recognition process, and is a value obtained by dividing by the number of examples used in the speech recognition process. The voice recognition device according to claim 1. 4. The function value of the ineligible sentence determination function is obtained by calculating the sum of semantic distances corresponding to the example used in the speech recognition process, and using the calculated sum in the speech recognition process. Calculate the average value of the semantic distance, which is the value divided by the number,
The average value of the calculated semantic distances is a value obtained by dividing the average value of the calculated semantic distance by the number of examples including a plurality of predetermined morphemes in the examples used in the speech recognition processing at the stage when the predetermined number of morphemes are processed. The speech recognition device according to claim 1, wherein 5. The speech recognition device according to claim 1, wherein the threshold value is a constant value. 6. The method according to claim 1, wherein the threshold value is changed depending on the number of morphemes included in the partial sentence to be subjected to the speech recognition processing. Voice recognition device.