JP2012022069A - Speech recognition method, and device and program for the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition method, for instance, appropriate for generation of a text to be used for data mining.SOLUTION: The speech recognition method comprises: a speech recognition process; a speech document recognition reliability calculation process, a speech document removal process; and a word removal process. The speech recognition process outputs a speech recognition result, in which the reliability of word recognition is added to each word after speech recognition processing is applied to an input speech document, and the speech document recognition reliability calculation process accepts the speech recognition result as an input, and calculates and outputs speech document recognition reliability which is the recognition reliability for the whole speech document. Then the speech document removal process accepts the speech recognition result and the speech document recognition reliability as an input and removes a speech document that falls short of a prescribed threshold of speech document recognition reliability, and the word removal process removes a word that falls short of a prescribed threshold of word recognition reliability, from the speech recognition results of the speech documents that are not removed in the speech document removal process.

Description

  The present invention relates to a speech recognition method and apparatus suitable for generating text used for data mining, for example, and a program.

  A technique for statistically analyzing data collected as text data is generally referred to as text mining. Voice recognition may be used for the purpose of obtaining the text data. Misrecognition is a natural part of speech recognition. Thus, conventionally, attempts have been made to reduce this speech recognition error.

For example, there is a method of giving a recognition reliability indicating the certainty of the recognition result to the voice recognition result. In Patent Document 1, the recognition reliability of the word w _{1 with} the highest score of the N best candidates up to the top N as a result of the search of the speech recognition process is different from the score of the word w ₁ with the score of 2 or lower and w _1. The idea that the score difference from the word w ₂ is normalized by the duration of the word w ₁ is shown.

Another method is to measure the strength of association between each word in the speech recognition result, and give high recognition confidence to words that are strongly related to surrounding words, and low recognition confidence to weakly related words. There is a method of providing a degree (Non-Patent Document 1). In this method, n word sets N (w) of the word w and k words immediately before the word w and one word immediately after the word w are acquired from the speech recognition result. Then, the mutual information MI (w _i , w _j ) calculated in advance on the learning corpus is used for all combinations (w _i , w _j ) of two words included in the word set N (w). Then, the strength S (w _i , w _j ) between words is calculated. Further, the average value of the relation strength S (t, w _i ) for all the words t in the word set N (w) is calculated as the context consistency measure SC (t).

  When data mining processing is performed on a large amount of stored voice documents, text with the above-described recognition reliability is used.

JP 2005-148342 A

D. Inkpen, A. Desilets, “Semantic Similarity for Detecting Recognition Errors in Automatic Speech Transcripts,” Proceedings of HLT / EMNLP, pp.49-56, October 2005.

  When a large amount of speech documents are converted into text by speech recognition and text mining processing such as full-text search, document classification, and pattern extraction is performed, words that are not actually spoken appear in the text due to speech recognition errors. There is a problem that incorrect information is included in the text mining result. When a lot of incorrect information is included in the text mining result, it is difficult for a text mining user to obtain useful knowledge by looking at the obtained result.

  As mentioned above, it is possible to reduce the amount of erroneous information extracted by giving recognition confidence to each word of the speech recognition result and excluding words with low recognition confidence from the target of text mining processing. is there. However, when the recognition accuracy of the entire speech recognition result is low, the recognition reliability for a word tends not to accurately represent the probability of recognition. For example, it is difficult to appropriately assign a recognition reliability to a recognition error word of an audio document recorded in an environment with a large ambient noise. In addition, even when the cause is a difference in the speaker's way of speaking, the recognition reliability of the entire voice document becomes low, and an appropriate recognition reliability cannot be given to each word.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a speech recognition method, a device thereof, and a program that make it difficult for a recognition error word to be included in text data by speech recognition. .

  The speech recognition method of the present invention includes a speech recognition process, a speech document recognition reliability calculation process, a speech document removal process, and a word removal process. The speech recognition process outputs a speech recognition result with word recognition reliability assigned to each word obtained by performing speech recognition processing on the input speech document. In the voice document recognition reliability calculation process, the voice recognition result, which is the recognition reliability of the entire voice document, is calculated and output using the voice recognition result as an input. In the voice document removal process, a voice document less than a predetermined voice document recognition reliability threshold is removed by inputting the voice recognition result and the voice document recognition reliability. In the word removal process, words having a word recognition reliability less than a predetermined word recognition reliability threshold are removed from the voice recognition result of the voice document that has not been removed in the voice document removal process.

  In the speech recognition method according to the present invention, since the removal is performed in units of words after the removal in units of speech documents, it is difficult to remove in units of words whose recognition reliability is generally low. The recognition result can be appropriately removed. Therefore, for example, it is possible to effectively reduce the number of recognition error words included in text data to be subjected to text mining processing. As a result, users of text mining can obtain useful knowledge.

  It is also effective to use the speech recognition method of the present invention for unsupervised adaptation of a probability model used for speech recognition. That is, since speech recognition results with few recognition error words can be collected, the speech recognition accuracy can be improved by using the collected speech recognition results with few recognition errors for unsupervised adaptation.

The figure which shows the function structural example of the speech recognition apparatus 100 of this invention. The figure which shows the operation | movement flow of the speech recognition apparatus 100. The figure explaining N best candidate and word recognition reliability. The figure which shows the function structural example of the audio | voice document recognition reliability calculation part 20. FIG. The figure which shows the operation | movement flow of the voice document recognition reliability calculation part 20. FIG. Word w _n output by the speech recognition unit 10, shows an example of a word recognition reliability D (w _n). The figure which shows the function structural example of the speech recognition apparatus 200 of this invention. The figure which shows the operation | movement flow of the speech recognition apparatus 200. The figure which shows the function structural example of the word related degree table creation apparatus 150. FIG. The figure which shows a word set notionally. The figure which shows an example of a word related degree table. The figure which shows the function structural example of the audio | voice document recognition reliability calculation part. The figure which shows the function structural example of the speech recognition apparatus 300 of this invention. The figure which shows the function structural example of the audio | voice document recognition reliability high speed calculation part 90. FIG. The figure which shows the operation | movement flow of the voice document recognition reliability high speed calculation part. The figure which shows the example of a word set, an immediately preceding duplication flag, and an immediately following duplication flag. The figure which shows the example of the operation | movement flow of a word set acquisition step. The figure which shows the example of the operation | movement flow of a word set acoustic reliability high-speed calculation step.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows a functional configuration example of the speech recognition apparatus 100 of the present invention. The operation flow is shown in FIG. The speech recognition apparatus 100 includes a speech recognition unit 10, a speech document recognition reliability calculation unit 20, a speech document removal unit 30, a word removal unit 40, and a control unit 50. The function of each part of the speech recognition apparatus 100 is realized by reading a predetermined program into a computer composed of, for example, a ROM, a RAM, a CPU, etc., and executing the program by the CPU.

  The speech recognition unit 10 outputs a speech recognition result in which word recognition reliability is given to each word obtained by performing speech recognition processing on the input speech document (step S10). The voice recognition unit 10 analyzes the voice document into an LPC cepstrum, MFCC, and other acoustic feature parameter series in units called frames of several tens of msec by an internal acoustic analysis unit (not shown). Then, a recognition result candidate for the input speech is searched for the acoustic feature parameter series using a dictionary and a language model. As a result of the search, N best candidates up to the top N are output as a speech recognition result together with the word recognition reliability. Note that the voice document is, for example, a set of conversations exchanged between a customer and an operator at a call center, and is voice data compiled for one requirement. For example, one lecture is organized into one audio file.

  Here, N best candidates and word recognition reliability will be described with reference to FIG. Note that N best candidates and word recognition reliability are conventional techniques. The word recognition reliability is described in Patent Document 1, for example.

  The horizontal axis in FIG. 3 represents elapsed time and is represented by a frame. The vertical axis represents N best candidates in which word string candidates searched for in units of frames are arranged in descending order of scores. The score is the likelihood at the time of search.

Word recognition reliability, if the word w _** in frame t _* (* is an arbitrary integer) is different word present in the N-best candidate words w _** frame t score paired candidacy word frames in _* Given by the score difference between the next ranking score at t _* . In the example shown in FIG. 3, the word recognition of the first candidate word w ₁₁ (11 represents the first word of the first candidate) searched for in the acoustic feature parameter series of frames t _{1 to} t _4. Since the opposite words are the third candidate word w ₃₁ and the second candidate word w ₂₁ , the value obtained by dividing the total score difference (●) by the number of frames is the word recognition reliability. It becomes. For the word w ₁₃ the conflict candidate does not exist, the word recognition reliability by predetermined fixed value (○) is used. This word recognition reliability is accumulated for each candidate and becomes the word string recognition reliability.

  The voice document recognition reliability calculation unit 20 calculates and outputs a voice document recognition reliability that is the recognition reliability of the entire voice document from the word recognition reliability for each word and the word string reliability (step S20).

The voice document removal unit 30 receives the voice recognition result output from the voice recognition unit 10 and the voice document recognition reliability output from the voice document recognition reliability calculation unit 20 as input, and is less than a predetermined voice document recognition reliability threshold θ _d. Are deleted (step S30).

Word removal unit 40 removes the word of a predetermined word recognition confidence threshold θ word recognition reliability of less than _w results speech recognition of the speech documents which were not removed in the spoken document removal unit 30 (step S40). The voice document recognition reliability threshold value θ _d and the word recognition reliability threshold value θ _w may be previously provided as constants in each unit, or may be given from the outside.

  As described above, the speech recognition apparatus 100 outputs the speech recognition result in which the word unit is removed from the remaining speech document after performing the removal in the speech document unit, so that the erroneous recognition word included in the speech recognition result. Can be reduced.

  FIG. 4 shows the functional configuration of the voice document recognition reliability calculation unit 20, and the operation will be described in more detail. The operation flow is shown in FIG. The voice document recognition reliability calculation unit 20 includes a word time length acquisition unit 21, a denormalization unit 22, a voice document total time length calculation unit 23, a reliability accumulation unit 24, and a voice document recognition reliability calculation unit 25. And comprising.

  The word time length acquisition unit 21 obtains the word time length of the word output from the voice recognition unit 20 (step S21). The word time length acquisition unit 21 first initializes the total reliability D (W) of the voice document and the total time length WD of the voice document to 0 (step S50). Steps S50 to S52 are processed by the control unit 50 shown in FIG.

FIG. 6 shows an example of the word w _n and the word recognition reliability D (w _n ) output from the speech recognition unit 20. Here, an example in which an audio file is composed of one N best candidate will be described. In other words, the subscript is described with one digit. For example, the word w ₁ is the noun “today”, the word recognition reliability D (w _n ) is 9891, and its start time (wd _n sFn) and end time (wd _n eFn) are 0.00-0. It is 98 [seconds]. The value of the word recognition reliability D (w _n ) may be a negative value. In the example of FIG. 6, “Nice” corresponds to this. When the score of the first candidate word is smaller than the score of the lower candidate word, the word recognition reliability D (w _n ) is negative. In this case, it means that the reliability of the word of the first candidate is considerably low. The value of the word recognition reliability D (w _n ) represents the acoustic reliability of the speech recognition result.

The word time length acquisition means 21 acquires the word time length wd ₁ obtained by subtracting the start time from the end time of the word w ₁ as 0.98 [sec] or 98 frames when the frame time is set to 10 msec, for example. (Step S21).

The denormalization means 22 calculates a word reliability wc obtained by multiplying the word recognition reliability D (w ₁ ) = 9891 by the word time length wd ₁ (step S22). The normalization cancellation means 22 functions to cancel the normalization of the word recognition reliability that has been normalized by the number of frames.

The reliability accumulating means 24 obtains a word reliability accumulated value D (W) obtained by accumulating the normalized word reliability wc (step S24). The voice document total time length calculating means 23 is a word time length obtaining means. The voice document total time length WD obtained by accumulating the word time length wd _* acquired in 21 in the entire voice file is calculated (step S23). Processing in steps S21~S23 described above is repeated until the processing for all words w _n of the audio file, while updating the word (step S52) (no in step S51).

  The voice document recognition reliability calculation means 25 calculates the voice document recognition reliability docC of the voice document obtained by dividing the word reliability cumulative value D (W) by the voice document total time length WD (step S25). The voice document recognition reliability calculation means 25 obtains the recognition reliability per frame of the voice document. The voice document recognition reliability docC is an index representing the acoustic quality of the voice recognition result of the voice document.

The voice document removal unit 30 compares the voice document recognition reliability docC representing the acoustic reliability of the voice document with a predetermined voice document recognition reliability threshold θ _d to obtain a voice document recognition reliability docC. If it is equal to or greater than θ _d , the voice document is output as it is to the word removal unit 40. When the voice document recognition reliability docC is less than θ _d , the voice document is not output to the word removal unit 40.

The voice document recognition reliability threshold θ _d is a real value, and when set to a large value (for example, about 30000), the voice recognition accuracy of the output voice document is increased. When a small value (for example, about −30000) is set, the voice recognition accuracy of the output voice document is lowered, but the number of output voice documents is increased.

The word removal unit 40 receives the words constituting the voice document output from the voice document removal unit 30 and the word recognition reliability D (w _n ) assigned to the words, and inputs the word recognition reliability D (w _n ). If the value is less than a predetermined word recognition reliability threshold value θw, the word is replaced with a predetermined symbol indicating that it has been removed, for example, “<rejected>”, and a speech document is output.

  The speech recognition result that is finally output by the above processing is obtained by removing words with lower word recognition reliability from a speech document having a relatively high speech recognition accuracy, and is a misrecognized word included in the speech recognition result. Can be reduced. When text data for data mining is acquired using the speech recognition apparatus 100, the number of misrecognized words included in the text data can be reduced, so that a user of text mining can obtain useful knowledge. Become.

  FIG. 7 shows a functional configuration example of the speech recognition apparatus 200 of the present invention. The operation flow is shown in FIG. The speech recognition apparatus 200 is a combination of the above-described speech document recognition reliability docC and a combination of acoustic reliability and context reliability. Therefore, only the speech document recognition reliability calculation unit 70 is different from the speech recognition apparatus 200. The acoustic reliability is the same as the voice document recognition reliability docC of the voice recognition apparatus 100. Hereinafter, the voice document recognition reliability docC described in the first embodiment is referred to as acoustic reliability.

  The speech document recognition reliability calculation unit 70 of the speech recognition apparatus 200 refers to the word association degree table 60 indicating the association degree between words constituting the speech recognition result in order to obtain context reliability from the speech recognition result. The operation of the word association degree table creation device 150 that creates the word association degree table 60 will be described.

[Word relevance table creation device]
FIG. 9 shows a functional configuration example of the word association degree table creation device 150. The word association degree table creation device 150 includes a learning corpus 81, a morpheme analysis unit 82, a learning corpus word set acquisition unit 83, a word list 84, a word count unit 85, a word association degree calculation unit 86, and a table array. Part 87.

  The learning corpus 81 is a collection of voice documents on a large scale. The morpheme analysis unit 82 performs a well-known morpheme analysis process that reads a speech document from the learning corpus 81 and divides it into words, and learns with word boundaries provided with symbols representing word boundaries before and after each word, for example, “\ n” Output corpus. The morphological analysis processing is well known and is described in, for example, a reference document “Japanese Patent No. 3379643”.

The learning corpus word set acquisition unit 83 performs windowing with a window width n words and a window shift amount m words from the beginning to the end of the word-boundary learning corpus output by the morpheme analysis unit 82, and a word list included in each window The words described in 84 are combined into a word set, and a word set for each window is output. The word list 84 describes all the words that can appear in the speech recognition result, and is created in advance. FIG. 10 conceptually shows the word set. The horizontal direction is the passage of time, and the word set is indicated by N _{1 to} N _h . m is the window shift amount, and n is the window width. Adjacent word sets share a relationship of nm words.

The word count unit 85 receives the word set output from the learning corpus word set acquisition unit 83 as an input, and the number of occurrences C (w) of each word in the word set and the number of occurrences C (w _i , w _{j of} each word pair) ), Count and output the total number of word sets. The number of occurrences C (w) of the word w is the number of word sets including the word w. The number of occurrences C (w _i , w _j ) of the word pair (w _i , w _j ) is the number of word sets including both w _i and w _j .

The word relevance calculation unit 86 calculates the relevance S (w _i , w _j ) of each word pair (w _i , w _j ) using, for example, equation (1).

N is the total number of word sets, C (w) is the number of single occurrences of word w, and C (w _i , w _j ) is the number of co-occurrence of words w _i and w _j . When the value of the relevance S (w _i , w _j ) is large, it means that the relevance between these words is high.

The table arrangement unit 87 arranges the relevance S (w _i , w _j ) calculated from the words w _i and w _j in a table format that can be referred to. FIG. 11 shows an example of the word association degree table 60. The uppermost column and the leftmost column are words w _{1 to} w _N , and the relevance S (w _i , w _j ) of each word is arranged in a column where each row and each column intersect.

  FIG. 12 shows a functional configuration example of the voice document recognition reliability calculation unit 70. The voice document recognition reliability calculation unit 70 includes a recognition result word set acquisition unit 71, a word set acoustic reliability calculation unit 72, a word set reliability calculation unit 73, and a reliability integration unit 74.

The recognition result word set acquisition unit 71 moves the words w _k and the word recognition reliability C (w _k ) of the speech recognition result by moving a predetermined number n from the head by m, which is smaller than n. Divide into word sets. That is, n words are acquired from the head of the speech recognition result of the speech document, and the obtained word set is set to N ₁ . Then, to get the n words of m words from the beginning of the speech recognition results, resulting word set to N _2. Next, n words are acquired from the 2m-th word from the beginning of the speech recognition result, and the obtained word set is defined as N ₃ . The above processing is repeated until the end of the speech document is reached, and h word sets N _k (N _k : N _{1 to} N _h ) are acquired. k is a variable that identifies a focused word set and word using n in the first embodiment.

The word set acoustic reliability calculation means 72 acquires the word recognition reliability D (w) and the word time length wd _k of all the words included in each word set, and the speech document recognition reliability described in the first embodiment. The acoustic reliability CA (N _k ) is calculated for each word set N _k by performing the same processing as the calculation unit 20.

The word set context reliability calculation means 73 calculates the word association strength S (w _i , w _j ) for all combinations of two words (w _i , w _j ) included in each word set. The average value is calculated as the context reliability CL (N _k ) of each word set.

The reliability integration unit 74 calculates and outputs an average value of h acoustic reliability CA (N _k ) and context reliability CL (N _k ) as a voice document recognition reliability. In this way, by calculating the recognition reliability of the voice document using the acoustic reliability CA (N _k ) and the context reliability CL (N _k ), the accuracy is higher than the voice document recognition reliability of the first embodiment. The voice document recognition reliability can be obtained.

FIG. 13 shows an example of the functional configuration of the speech recognition apparatus 300 of the present invention in which the processing for calculating the speech document recognition reliability is saved. The speech recognition apparatus 300 is different from the speech recognition apparatus 200 in that it includes a speech document recognition reliability high speed calculation unit 90. The voice document recognition reliability high-speed calculation unit 90 calculates the acoustic reliability at high speed by omitting the reliability addition processing overlapping between the word sets _Nk .

  FIG. 14 shows a functional configuration example of the voice document recognition reliability high-speed calculation unit 90. The operation flow is shown in FIG. The speech document recognition reliability high-speed calculation unit 90 is different from the word set acoustic reliability calculation unit 72 of the speech recognition apparatus 200 in that it includes a word set acquisition unit 91 and a word set acoustic reliability high-speed calculation unit 92.

  The word set obtaining unit 91 obtains a word set by assigning each word with two flags, the immediately preceding duplicate flag and the immediately following duplicate flag used for the purpose of reducing the calculation amount of the acoustic reliability. The immediately preceding duplicate flag is a true / false value indicating whether or not the word is also included in the immediately preceding word set. The immediately following duplication flag is a true / false value indicating whether or not the word is also included in the immediately following word set.

The word set acquisition unit 91 divides the word w _k of the speech recognition result input from the speech recognition unit 10 into h word sets while moving a predetermined number n from the beginning, m less than n. (Step S91). When the word set is divided into word sets, the word set acquisition unit 91 sets the immediately preceding duplicate flag BF and the immediately following duplicate flag AF of the first to mth words to be added to the first word set to false, and the (m + 1) th to nth The word immediately preceding duplicate flag BF is false and the immediately following duplicate flag AF is true, and the first to (N-1) -mth word immediately preceding duplicate flags to be added to the Nth word set are true and the immediately following duplicate flag is false. , (N−1) · m + 1 to n + (N−1) · m-th word are set to false, and the next duplicate flag is set to true.

  The word set acoustic reliability high-speed calculation unit 92 includes an overlapping section storage unit 920, and a value obtained by canceling time normalization by multiplying the word recognition reliability of a true word by the word recognition reliability of the true word and the duration of the word. And the duration length thereof are stored in the overlapping section storage unit 920. Then, the value obtained by multiplying the word recognition reliability of the true word by the duration length of the word and canceling the time normalization, the duration length, and the overlap interval memory The acoustic reliability of the word set is calculated from the value stored in the unit 920 (step S92).

FIG. 16 shows an example of the word sets N _{1 to} N _h and the immediately preceding duplicate flag BF and the immediately following duplicate flag AF given to each word set. The horizontal direction in FIG. 16 is the elapsed time. The word sets that overlap in the horizontal direction are shown shifted in the vertical direction.

The m words from the beginning of the word set N _1, the front and rear because the word does not overlap with the words in the set immediately before overlap flag BF and after overlap flag AF is (BF: 0: 0, AF ). Since the words m + 1 to n from the beginning of the word set N ₁ overlap with the words in the immediately following word set (N ₂ ), the immediately preceding duplicate flag BF and the immediately following duplicate flag AF are (BF: 0, AF: 1 ). Hereinafter, the notation of BF: and AF: may be omitted.

Since the 1st to m-th words in the second word set N ₂ overlap only with the words in the previous word set (N ₁ ) (1,0), the m + 1-th to n-th words are immediately before and after (1, 1), the (n + 1) th to n + mth words overlap with only the words of the immediately following word set (0, 1). The third and subsequent word sets have the same relationship as the second word set.

In the example of FIG. 15, since m is set to m <n / 2, there is a state (1, 1). However, when m = n / 2, the immediately preceding duplicate flag BF and the immediately following duplicate flag AF have N ₁ ( (0, 0), (0, 1)), and after N ₂ are ((1, 0), (0, 1)). That is, the word set acquisition unit 91 sets the immediately preceding duplicate flag BF of the first to (N−1) · mth words to be true and the immediately following duplicate flag AF to be false, and (N−1) · m + 1 to n + (N -1)-The preceding duplicate flag BF of the mth word is false and the immediately following duplicate flag AF is true.

  FIG. 16 shows an operation flow of the word set acquisition unit 91 for acquiring the word set shown in FIG. FIG. 16 shows an example in which the window shift amount m is m <n / 2.

When spoken document recognition reliability calculating unit 90 starts the process, the word set obtaining unit 91 variable w _k, and the count value i of each N _k = N ₁ specifying the variable N _k and words w _n for specifying a word set , W _k = w ₁ , i = 1 is initialized (step S91a). Then, 0 last overlap flag BF word w _k to word w ₁ to w _m from the beginning of the speech recognition result to m-th (false), the immediately overlap flag AF is set to 0 (false) (step S91b~ S91d). Further, the immediately preceding duplication flag BF of words w _{k + 1 to} w _{n is set} to 0 (false) and the immediately following duplication flag AF is set to 1 (true) (steps S91e to S91g), and the words w ₁ to w are set. Up to _n are acquired as the _first word set N ₁ .

Next, the second word set N ₂ is obtained with the word set N _k as N _k = N _{k + 1} (step S91h). Here, it is assumed that the variable w _k representing the word of interest is w _k = w _{mi that} is moved mi from the _first word w ₁ (step S91i). Here, since i = 1, w _k = w _m .

Thus, the word w _{m + m} , that is, the immediately preceding overlap flag BF from the first word to 2m words is set to 1 (true), and the immediately after overlap flag AF is set to 0 (false) (steps S91j to S91m). Then, the immediately preceding duplication flag BF is set to 1 (true) and the immediately following duplication flag AF is set to 1 (true) for 2m + 1 to n + m (i-1) words from the first word (steps S91n to S91n). S91p). Further, the immediately preceding overlap flag BF is set to 0 (false) and the immediately following overlap flag AF is set to 1 (true) for the n + m (i-1) +1 to n + mi words from the first word ( Steps S91q to S91s). By such processing, the words w _{m to} w _{n + m} are acquired as the word set N ₂ with the immediately preceding overlapping flag BF and the immediately following overlapping flag AF of the word set as shown in FIG.

If the n + mi + 1-th word from the head exists, the count value i is incremented (step S91u), and the variable N _{k representing the} word set is also incremented to N _k = N _{k + 1} (step S91v). The process returns to step S91i.

  The above processing (steps S91i to S91v) is repeated until the last word of the speech recognition result is reached (no in step S91t). As a result, the words of the speech recognition result are divided into h word sets while moving a predetermined number n from the beginning by m, which is smaller than n. Then, as shown in FIG. 15, the immediately preceding overlap flag BF and the immediately following overlap flag AF are assigned to each word set.

  The word set acoustic reliability high-speed calculation means 92 includes an overlapping section storage unit 920, and does not calculate the word reliability of a word whose true overlap flag AF is true twice. The operation will be described with reference to the operation flow of the word set acoustic reliability high speed calculation means 92 shown in FIG.

Word set sound reliability fast computation means 92, first the variable w _k for specifying the variable N _k and words w _n for specifying a word set is initialized to _{N k = N 1, w k} = w 1 , respectively ( Step S92a). Then, the word set acoustic reliability fast calculation means 92 multiplies the word recognition reliability of the words (N ₁ , w ₁ ) to (N ₁ , w _n ) by the duration time of the word to cancel the time normalization. Get words reliability (step S92b) and, accumulating the word confidence wc _n (step S92c). At the same time, the word duration time wd _n is also accumulated (step S92d). This process is repeated until the last word w _n of the word set N ₁ is reached (step S92i).

Then, since the duplication flag AF immediately after the word from the _m-th word w _m to the last word w _n of the word set N ₁ is 1 (true), the word trust for the words w _{m to} w _n degrees wc _n is stored in BackDupDM (step S92f). Also, duration wd _n words w _m to w _n are stored in BackDupWD (Step S92G). BackDupDM and BackDupWD store the word reliability wc _k and the duration time wd _k in a data structure called a queue, for example.

When the calculation of the word reliability up to the last word w _n is finished (yes in step S92i), the acoustic reliability of the word set N ₁ and the cumulative value S (wc _k ) of the word reliability wc _k are set as the duration length. Calculation is performed by dividing by the cumulative value S (wd _k ) (step S92i).

N-m-th from the second head of the word set N _2, i.e. the top of n-th word wm~wn from the word is already computed so overlaps with the immediately preceding word set N _1. Therefore, the calculated word reliability stored in BackDupDM is copied to PreDM (step S92m), and the duration length stored in BackDupWD is copied to PreWD (step S92n). Then, duplicate flag BF immediately after the word set N ₁ does not calculate the word reliability is 0 word w _{n + 1} ~w _{n + mi} to release the normalized word reliability and duration of the (false) The accumulation is calculated (steps S92p and S92q). Note that the denormalization step is omitted for convenience of drawing.

The word reliability wc _k and the duration time wd _k of the words wn _{+ 1 to} wn _{+ mi} whose duplicate flag BF is 0 (false) immediately after being newly calculated are stored in BackDupDM and BackDupWD (step S92r). . Since BackDupDM and BackDupWD are structures in which data is erased in the oldest order, the word reliability wc _k and duration length wd _{k of the} latest words w _{2m to} w _{n + m} are stored. For the nth and subsequent words (w _{n + 1} ) from the first word in the speech recognition result, only the word with the previous duplication flag BF = 0 is denormalized and the cumulative duration is calculated. The calculated values are stored in queues (BackDupDM, BackDupWD) (see FIG. 16).

Last word w _{n + mi} to process word set N ₂ is completed (step yes in S92t), newly computed sum word confidence stored acoustic confidence word set N ₂ to BackDupDM Calculate the acoustic reliability of the word set N ₂ by dividing the cumulative value including the word reliability by the value obtained by adding the newly calculated duration to the sum of the durations stored in BackDupWD ( Step S92u).

Then, the count value i is incremented (step S92x), the word set N _{k of} interest is incremented, and the same processing is repeated for the word sets after N ₃ until there are no words in the speech recognition result (yes in step S92w). ). In this way, the word set already calculated in the immediately preceding word set N _k and the duration length are copied and calculated, and thus are not calculated twice.

That is, the word set acoustic reliability high-speed calculation unit 92 multiplies the word recognition reliability D (w _k ) of a word whose true overlap flag BF is true by the overlap interval storage unit 920 by the duration duration wd _k of the word. Stores the value wc _k with the time normalization canceled and its duration wd _k, and sets the word w _k to the word recognition reliability D (w _k ) of the word with the immediately preceding duplicate flag AF being false and the immediately following duplicate flag BF being true. The acoustic reliability of the word set w _k is calculated from the value wc _{k obtained} by canceling the time normalization by multiplying the continuous time length, the continuous time length wd _k, and the value stored in the overlapping section storage unit 920. Therefore, the acoustic reliability can be calculated at a higher speed than the speech recognition apparatuses according to the first and second embodiments.

  Although an example in which the acoustic reliability is calculated at high speed using the immediately preceding overlap flag BF and the immediately following overlap flag AF has been described, the processing method is not limited to the above-described example. For example, in addition to the example shown in FIG. 18, the total word reliability of overlapping sections and the total duration time of overlapping sections may be stored. By doing so, it is possible to reduce processing for calculating the cumulative value of the word reliability and the duration length stored in the queue every time the acoustic reliability is calculated (step S92u).

  When the processing means in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (5)

A speech recognition process for outputting a speech recognition result to which word recognition reliability is given for each word obtained by performing speech recognition processing on an input speech document;
A speech recognition reliability calculation process for calculating and outputting a speech document recognition reliability, which is a recognition reliability of the entire speech document, using the speech recognition result as an input;
An audio document removal process of removing an audio document having a threshold value less than a predetermined audio document recognition reliability threshold by inputting the audio recognition result and the audio document recognition reliability;
A word removal process for removing words having a word recognition reliability less than a predetermined word recognition reliability threshold from the voice recognition result of the voice document that has not been removed in the voice document removal process;
A speech recognition method including: The speech recognition method according to claim 1,
The above speech recognition reliability calculation process is as follows:
A word set acquisition step of dividing the word of the speech recognition result into a word set while moving a predetermined number n from the beginning by m, which is a number smaller than n;
Multiply the word recognition reliability by the duration of the word and cancel the time normalization, add the sum, and divide the sum by the sum of the durations of all words in the word set. A word set sound reliability calculation step to be obtained as a set sound reliability;
By referring to a word association degree table representing the degree of association between two words among all the combinations between words included in the speech recognition result, an average value of association strengths of the combinations of two words included in the word set is obtained. A word set context reliability calculation step to obtain as word set context reliability,
A reliability integration step of obtaining a value obtained by averaging the word set acoustic reliability and the word set context reliability of the entire voice document as a voice document recognition reliability;
A speech recognition method comprising: The speech recognition method according to claim 1,
The above speech recognition reliability calculation process is as follows:
The word of the speech recognition result is divided into a word set while moving a predetermined number n from the beginning by m less than n, and immediately before the first to mth words to be added to the first word set The duplicate flag and the immediately following duplicate flag are set to false, the immediately preceding duplicate flag of the first to N · mth words to be added to the Nth word set is true and the immediately following duplicate flag is false, and the N · m + 1 to n + N A word set acquisition step in which the immediately preceding duplicate flag of the mth word is false and the immediately following duplicate flag is true;
The overlapping section storage step for storing the value obtained by multiplying the word recognition reliability of the word for which the duplication flag is true by the duration length of the word and canceling the time normalization, and the duration length, and
A value obtained by multiplying the word recognition reliability of the word whose previous duplicate flag is false and the true duplicate flag is true by the duration length of the word and canceling the time normalization, the duration duration, and the overlapping section storing step A word set acoustic reliability fast calculation step for calculating the acoustic reliability of the word set from the value stored in
A speech recognition method comprising: A speech recognition unit that outputs a speech recognition result to which word recognition reliability is given for each word obtained by performing speech recognition processing on the input speech document;
A speech recognition reliability calculation unit that calculates and outputs a speech document recognition reliability that is a recognition reliability of the entire speech document by using the speech recognition result as an input;
A voice document removal unit that removes a voice document that is less than a predetermined voice document recognition reliability threshold by inputting the voice recognition result and the voice document recognition reliability;
A word removal unit that removes words having a word recognition reliability less than a predetermined word recognition reliability threshold from the voice recognition result of the voice document that has not been removed by the voice document removal unit;
A speech recognition apparatus comprising:   A speech recognition method program for causing a computer to execute the speech recognition method according to claim 1.

Images