JP4478925B2 - Speech recognition result reliability verification apparatus, computer program, and computer - Google Patents

Description

  The present invention relates to a technique for determining acceptance / rejection of a speech recognition result, and more particularly, to a technique for performing determination using a reliability measure of a word of a speech recognition result.

  The fields where the most advanced speech recognition technology can be applied are very wide. Some of these areas have been successful. However, voice recognition technology is not yet certain. Therefore, there is a need for some scale that can monitor the voice recognition result and easily determine whether to accept or reject the recognition result. The need for such technology is expected to increase in the future as speech recognition results are applied to new and difficult fields.

  Such a measure must be easily computable and statistically meaningful. A word posterior probability (Wpp) has been proposed and well studied. Non-Patent Documents 1, 2, and 3 include studies in which word posterior probabilities are applied to word lattices / graphs or N-best lists of speech recognition results.

Stephan Ortmans, Hermann Ney, Savier Obert, “Word Graph Algorithm for Large Vocabulary Continuous Speech Recognition”, Computer Speech and Language, Vol. 43-72, January 1997 (Stefan Ortmanns, Hermann Ney, and Xavier Albert, "A Word Graph Algorithm Conforme Sp.," Long Vocable Continuum Spe. , January 1997.) Richard Schwarz and Yen Lou Chou, “N-Best Algorithm: Efficient and Accurate Procedure for Finding N Maximum Likelihood Hypotheses”, ICASSP 1990 Proceedings, 1990, Vol. 81-94 (Richard Schwartz and Yen-Lu Chow, "The N-best Algorithm, An 94 Efficient and Excise Procedure in Proceed in the Nest, Sentinel Sent. .) Frank K. Soon and En-von-Fan, “Fast Search for Discovery of N Best Sentence Hypotheses in Continuous Speech Recognition Based on Tree Trellis”, ICASSP 1991 Proceedings, 1991, Vol. 705-708 (Frank K. Soong and Eng-Fong Hang, "A Tree-Tellis Based Fast Searching for 19th in the Concealed in the Continuing. -708.)

  However, Non-Patent Documents 1 to 3 described above specifically describe how to use word posterior probabilities to easily and reliably determine acceptance / rejection of speech recognition. Absent. In application fields where the accuracy of speech recognition is not high, there are severe restrictions on accepting / rejecting judgment, and the judgment is not easy. Therefore, a technique that can easily and reliably determine acceptance / rejection of a speech recognition result is required particularly in a new application field of the speech recognition technique.

  Even in a technical field where speech recognition technology has been used, various processes can be automatically executed without human intervention if it is possible to automatically determine whether speech recognition results are accepted / rejected. For example, the speech recognition environment can be automatically adjusted based on the reliability of speech recognition, the computer can automatically train in some way, or can be used to practice language utterance.

  Therefore, an object of the present invention is to provide a speech recognition result verification apparatus that can easily and reliably accept / reject speech recognition results.

  Another object of the present invention is to provide a speech recognition result verification apparatus that can easily accept / reject each word of a word string that is a speech recognition result with high reliability.

  A speech recognition result reliability verification apparatus according to a first aspect of the present invention provides a speech recognition result representing a plurality of hypothesized word strings, each of which is output from a speech recognition decoder and is composed of words each having a word posterior probability. And a speech recognition result reliability verification device for verifying the reliability of the speech recognition result based on the word posterior probability, and for each word included in the speech recognition result, the word included in the speech recognition result A generalized word posterior probability calculating means for calculating a generalized word posterior probability based on the word posterior probability of the word, and a word posterior probability of each word included in the speech recognition result is calculated by the generalized word posterior probability calculating means. Update means for updating with the generalized word posterior probability, and based on the speech recognition result in which the word posterior probability is updated by the update means, among the plurality of hypothesis word strings, Search means for searching for a word having the maximum sum of the word posterior probabilities, and whether or not the sum of the word posterior probabilities of the hypothesized word string searched by the search means satisfies a predetermined condition And determining means for verifying the reliability of the speech recognition result.

  Preferably, the reliability verification apparatus for the speech recognition result is more likely than a threshold value determined by a predetermined criterion among the speech recognition results prior to the calculation of the generalized word probability by the generalized word a posteriori probability calculating means. Means for selecting only a word string having a high value and giving it to the generalized word posterior probability calculation means.

More preferably, each word included in the speech recognition result is further attached with information for determining a time period during the input utterance to the speech recognition decoder, and the generalized word a posteriori probability calculating means includes for each word contained, a time period overlaps with the time period of the word, and a word search means for searching for a word that matches the word in the speech recognition result, a word retrieved by the word search means Means for calculating a generalized word posterior probability of each word based on the sum of word posterior probabilities and the sum of word posterior probabilities of all words included in the speech recognition result.

  More preferably, the means for calculating the generalized word posterior probability is a sum of the word posterior probabilities of the words searched by the word search means and the sum of the word posterior probabilities of all words included in the speech recognition result. Means are included for calculating the generalized word posterior probabilities for each word by ratio.

Preferably, the generalized word posterior probability p ([w; s, t] | x ₁ ^T ) of the word w in the hypothesis word string (where s and t are the start time and end time of the time period of the word w, respectively) The following formula

Given, provided that ^{_{x 1 T = x 1, ...}} , x T is the observed speech sequence, M is the number of words included in the hypothesis of a speech recognition result, s _n and t _n, respectively, word w _Are the start and end times of the _nth word _wn , where p (x _sm ^tm | w _m ) is the acoustic likelihood, and p (w _m | w ₁ ^M ) is the language likelihood, p (x ₁ ^T ) is the acoustic observation likelihood, and α and β are predetermined constants, respectively.

  When the computer program according to the second aspect of the present invention is executed by a computer, the computer program is operated so as to realize each means of the reliability verification device for any of the above-described speech recognition results.

  A computer according to the third aspect of the present invention is programmed by the computer program described above.

[Introduction]
In the present embodiment, the problem of determining whether to accept or reject each word recognized by continuous speech recognition is solved by introducing the concept of specifying the position of the word of interest. Words other than the attention word (non-attention word) are not distinguished from each other and are treated as occupying the respective places, and the posterior probability of the attention word is calculated.

  Thus, by adopting the bisection method of attention word / non- attention word, it is possible to avoid the necessity of performing complicated processing such as character string alignment based on dynamic programming.

  First, the following concepts are introduced and explained. That is, they are (1) reduction of search space for a character string that becomes a hypothesis (candidate) for determining the position of the word of interest in the word lattice or N-best list of the speech recognition result, and (2) The relaxation of temporal constraints when grouping posterior probabilities at multiple occurrences of candidate words, and (3) appropriate weighting for contributions by acoustic and language models.

-Possibilities of character strings and words-
HMM: In the speech recognition apparatus using the (Hidden Markov Model), given the acoustic observation data ^{_{x 1 T = x 1, ...}} , for x _T, the optimal word sequence ^{_{w 1 M * = w 1 *}} , ... , W _M ^* , as shown below, is searched for a space consisting of all possible word sequences and gives the maximum posterior probability (MAP).

Here, p (x ₁ ^T | w ₁ ^M ) is the probability of the acoustic model, p (w ₁ ^M ) is the probability of the language model, and p (x ₁ ^T ) is the acoustic observation probability.

  Even the “optimal” word sequence may contain errors due to differences in training and test environments, speakers, noise, and the like. Therefore, some kind of reliability measure that is mathematically easy to handle and statistically favorable should be adopted.

The posterior probability p (w ₁ ^M | x ₁ ^T ) of the word string measures the likelihood of the recognized word string w ₁ ^M with respect to the observed sound x ₁ ^T , which corresponds to the corresponding time. Segmentation

Is calculated by assuming Here, s and t indicate the start and end times of the word w, where t _m + 1 = s _{m + 1} for _{m of} s 1 = 1, t _M = T, 1 ≦ m ≦ M−1.

  Using this, equation (2) can be rewritten as follows.

In order to measure the reliability of the entire recognized word string, it is natural to employ this word string posterior probability p (w ₁ ^M | x ₁ ^T ).

A suitable confidence measure for measuring word reliability is the word posterior probability p ([w _m ; s _m , t _m ] | x ₁ ^T ). This is calculated by summing up all posterior probabilities of word strings including a specific word.

In order to use this word posterior probability as an effective reliability measure, several problems need to be solved.

[Correction of word posterior probability]
-Number of hypotheses to consider-
In a large vocabulary continuous speech recognition apparatus (LVCSR), the search space for possible word strings is enormous. However, there is a great difference in the value of the posterior probability of each word string, and a word string with a relatively low likelihood may be trimmed. The word lattice / graph or the N-best word string list can be obtained by using only the subset of the word string hypotheses obtained in this way. In the following embodiment, it is assumed that the word lattice / graph obtained by using the subset is used.

-Temporal registration of words in a hypothesis-
Word temporal registration (registration) is represented by [w; s, t]. Even if the same word appears in different hypotheses, its position may differ slightly depending on the hypothesis. Since the ultimate goal of automatic speech recognition (ASR) is to recognize the content of words being spoken, we will relax some of the strict time constraints. Here, the period in which a certain word appears in a certain word string overlaps (overlaps) the period [s, t] of the reference word, and the word matches the reference word. Search for such words and include them in the calculation of the posterior probabilities of the reference word. As a result, equation (7) can be rewritten as follows.

-Specific gravity between acoustic likelihood and language likelihood-
In the present embodiment, the acoustic likelihood and the language likelihood are exponentially weighted by weights indicated by α and β, respectively. When this is applied to the equation (8), the following equation is obtained.

[Extract attention word]
Here, the acceptance / rejection of the attention word extracted by the word extraction method according to the present embodiment will be considered. FIG. 1 shows an example of a word lattice / graph obtained as a result of speech recognition used in the present embodiment. FIG. 2 similarly shows an N-best list of word strings obtained as a result of speech recognition (of hypothetical word strings). , A schematic example of a list of N items with a high likelihood).

  Referring to FIG. 1, the word lattice / graph 20 used in the present embodiment is different from the conventional one, and words other than the attention word (indicated by “w”) are not attached with individual word labels. Both are simply labeled “*”.

  For each occurrence of the word w, the word posterior probability can be efficiently calculated using the forward-backward algorithm. It then sums the likelihood for all of the paths through this particular word w (eg, words 30, 32, 34) and divides the sum by the sum of the likelihoods of all paths in this word lattice / graph. By normalizing, a generalized word posterior probability (hereinafter referred to as “generalized word posterior probability”) can be calculated. Further, at this time, the condition of word temporal registration (word start and end time coincidence) is relaxed. That is, the period of the word w of each path does not need to match exactly, and the total of the posterior probabilities of those that overlap in time is calculated.

  Similarly, the generalized word posterior probability of the word w can also be calculated using the N-best list as shown in FIG. Again, words other than the attention word 70 (word w) are marked with “*”, and it does not matter what the word is. Consider the case where hypotheses 50,..., 62 exist as shown in FIG. As the word w that temporally overlaps the word 70, the words 72, 74, 76, 78 of the hypotheses 54, 56, 58, and 60 can be considered. Since the word 80 of the hypothesis 62 does not overlap with the period of the word 70, it is not used for calculating the word posterior probability in this case.

  As described above, the sum of the likelihoods of the hypothesis 50, 54, 56, 58, 60 in which the word 70 and the same words 72, 74, 76, 78 that overlap with the word 70 appear in time are calculated. Then, by dividing it by the total likelihood of all hypotheses in the N-best list and normalizing it, the generalized word posterior probability of this word can be calculated. Again, the restriction on temporal registration is relaxed.

  It should be noted that when the attention word is extracted and the generalized word posterior probability is calculated as described above, word alignment is not necessary. Moreover, it is not necessary to obtain the alignment of the hypothesis by the dynamic programming method.

[Apparatus configuration according to the present embodiment]
FIG. 3 shows a block diagram of a speech machine translation apparatus 80 including a hypothesis verification apparatus 94 according to the present embodiment. Referring to FIG. 3, this speech machine translation apparatus 80 performs speech recognition of input speech s (t) 100 and outputs the recognition result as a word graph 104 such as the word lattice / graph 20 shown in FIG. And the hypothesis is verified using the generalized word posterior probability as described above for each word in the word graph 104 output from the ASR decoder 90 and the ASR decoder 90. As a result, the highest likelihood is obtained. A hypothesis verification device 94 for outputting a high word string 106 with a generalized word posterior probability and a word string 106 output from the hypothesis verification device 94 for machine translation and outputting a translation result 110 Machine translation device 92.

  The speech machine translation apparatus 80 further receives the word string 106 output from the hypothesis verification apparatus 94, and accepts or rejects the word string 106 as a recognition result based on the word posterior probabilities attached to the words of the word string. The user interface (hereinafter referred to as “user I / F”) 82 feeds back the result to the user and accepts / controls the machine translation device 92 based on the determination result. A rejection determination device 96 is included.

  FIG. 4 shows details of the hypothesis verification device 94. Referring to FIG. 4, hypothesis verification device 94 has a likelihood in word graph storage unit 120 for storing word graph 104 provided from ASR decoder 90 and in the word graph stored in word graph storage unit 120. In order to calculate a generalized word posterior probability for each word included in a word string excluding a low word string, a target word search unit for searching for the same word that overlaps with the period of the word in the word graph 122, a posterior probability calculation unit 124 for calculating the generalized word posterior probability by the calculation method described above for the word group searched by the target word search unit 122, and a posterior probability calculation unit 124 for each word. For re-assigning the generalized word posterior probabilities to each word of the word graph stored in the word graph storage unit 120 to update the word graph A path (maximum likelihood path) indicating the highest generalized word posterior probability is searched for in the rough update unit 126 and the word graph in which the generalized word posterior probability is reassigned for each word in this way, and the path is searched. And a maximum likelihood path search unit 128 for outputting the word string included in as a word string 106 together with the generalized word posterior probabilities.

[Operation]
This device operates as follows. Referring to FIG. 3, when input speech 100 is given, ASR decoder 90 performs speech recognition and outputs the result as word graph 104. Each word in the word graph 104 is given and given a word posterior probability at the time of recognition.

  Referring to FIG. 4, word graph storage unit 120 of hypothesis verification device 94 stores a subset of the word graph excluding those with low likelihood. The target word search unit 122 includes, for each word, a word group (a word that is the same as the word in question) that is a calculation target of the generalized word posterior probability in the word string of the word graph stored in the word graph storage unit 120. A word appearing on another path during a period overlapping with the appearance period of the word) is retrieved and given to the posterior probability calculation unit 124.

  As described above, the posterior probability calculation unit 124 sums up the word posterior probabilities for the word group searched by the target word search unit 122 and normalizes the sum by dividing the sum by the word posterior probabilities of all paths. The generalized word posterior probability of the target word is calculated.

  The word graph update unit 126 re-assigns the generalized word posterior probability calculated by the posterior probability calculation unit 124 for each word in the word graph.

  When the generalized word posterior probabilities are reassigned to all the words, the maximum likelihood path search unit 128 searches the word graph for the highest generalized word posterior probability in the word graph, and the result is found. The word string included in the path is given to the machine translation device 92 and the acceptance / rejection determination device 96 shown in FIG. 3 as the word string 106 together with the generalized word posterior probability.

  The acceptance / rejection determination device 96 determines whether to accept or reject the word string as a recognition result based on the word posterior probability of the word string given from the word string 106. In this case, the likelihood of this word string is compared with a predetermined threshold, and if it is greater than or equal to the threshold, it is accepted, and if it is less than the threshold, it is rejected. The acceptance / rejection determination device 96 controls the machine translation device 92 to execute machine translation for the word string 106 when accepting the recognition result. In the case of refusal, machine translation by the machine translation device 92 is stopped, and the user I / F 82 is used to inform the user that the recognition result has been rejected.

  The machine translation device 92 performs machine translation on the word string 106 in response to an instruction to start translation from the acceptance / rejection determination device 96, and outputs a translation result 110. At this time, since posterior probabilities are given to the respective words in the word string 106, translation can be performed in consideration of the posterior probabilities in translation.

[Experiment]
-Configuration of the experimental system-
An apparatus according to the above-described embodiment was set up and an experiment was performed. In this experiment, a method using an N-best list instead of a word graph was adopted. In the experiment, a Japanese basic travel expression corpus created by the applicant was used. Two sets 01 and 02 were used as test sets. These test sets consist of 510 utterances and 508 utterances, respectively. Each of these two test sets is a recording of various utterances by 10 speakers. As the ASR decoder 90, the one developed by the applicant was used.

  In this experiment, the hypothesis of the 100-best recognition result was output by the ASR decoder 90 for each utterance, and the search beam width was narrowed.

-Performance evaluation-
In the experiment, the error rate (Confidence Error Rate: CER) of the reliability measure was defined as follows using the error rejection number (FR) and the error acceptance number (FA), and the performance of the experimental system was evaluated.

-result-
If all recognition results are accepted and not rejected, the only errors that can be identified are insertion and replacement. In the experiment, this error level was used as the baseline. The baseline CER of the recognition results for the test set used in the experiment is shown as “Baseline” in Table 1 below.

A simple method for calculating word posterior probabilities is to count the number of hypotheses in which a particular word appears. Based on the ratio of this number to the total number of hypotheses, a rough value of the generalized word posterior probability of the word can be calculated. This corresponds to the case where α = β = 0 in equation (9). The results when this rough calculation method is used are shown as “reappearance rate” in Table 1. As can be seen from Table 1, using this method resulted in a CER improvement of about 2 points relative to the baseline.

  The contribution ratio of acoustic likelihood and language likelihood to generalized word posterior probabilities is not clearly understood, but in order to calculate generalized word posterior probabilities as accurately as possible, the values of α and β must be determined appropriately. Is useful. To examine the effect of various values of α and β on the generalized word posterior probability calculated by equation (9), the performance of a single threshold classifier was tested over a wide range of α and β. The results are shown by the contour diagrams of FIG. 5 (set 01) and FIG. 6 (set 02). In FIGS. 5 and 6, high performance was obtained when the combination of α and β belonging to a dark region was used.

  From this result, in both cases of set 01 and set 02, the combination of α and β exhibiting high performance exists on a substantially straight line on the graph, and the values of α and β are relatively low. It was found that there was a high performance part in a small area. Therefore, a more detailed investigation was performed on the region (α∈ [0.01, 0.2], β∈ [0.1, 1.5]) considered to be optimal. The results are shown in FIG. 7 (set 01) and FIG. 8 (set 02).

  7 and 8, the highest performance is obtained at α = 0.06 and β = 0.3 for the set 01, and at α = 0.03 and β = 0.3 for the set 02. It has been found that the highest performance can be obtained. The CER at these optimum points is shown in the bottom row of Table 1.

  FIG. 9 shows a ROC (Receiver Operating Characteristics) curve of the reappearance rate and the word posterior probability (wpp) when using the optimum α and β. FIG. 9A is for set 01 and FIG. 9B is for set 02.

It can be seen from FIG. 9 that the performance degradation is negligible when the optimal α and β obtained using one of the test sets is applied to the other test set. Table 2 shows the CER performance by this cross verification.

From Table 2, it can be seen that these parameters are very stable. Even if either α or β is slightly changed, the performance does not drop sharply.

FIG. 10 shows in more detail the behavior of the system performance in the vicinity of the optimal point for these two test sets. As can be seen from FIG. 10 (C) and (D), CER to variations in α is relatively stable. Also as can be seen from FIG. 10 (A) and (B), CER as variation of α with respect to the variation of β is not stable. However, it can be said that there is no sudden CER fluctuation and it is relatively stable.

-2-class Gaussian distribution classifier-
For comparison, the hypothesis verification method by word spotting was converted into a classification process into two classes. This classifier was trained with a given data set. The data was first voice-recognized to generate a 100-best list. Based on this 100-best list, the posterior probabilities of each word in the recognition result word string were calculated and tagged with a label indicating accuracy.

  We classified the word posterior probabilities of correct and incorrect words into two clusters and trained two Gaussian distribution models. As a training set and a test set, set 01 and set 02 were used alternately.

  Table 3 shows the obtained performance.

As can be seen from Table 3, in either case, the performance is almost the same as in the case of hypothesis classification with a single threshold shown in Table 1. However, in any case, the performance is slightly lower than in the case of Table 1.

  As described above, when the hypothesis verification device 94 according to the embodiment of the present invention is used, a generalized word posterior probability can be calculated for each word output from the ASR decoder. At that time, (1) the number of hypotheses for search can be reduced and the search space can be made smaller, so that the processing can be performed at high speed. (2) When calculating the word posterior probability of a word, the same word is selected. As a result, the stable value of the word posterior probability can be calculated. (3) In calculating the word posterior probability, the degree of contribution between the acoustic likelihood and the language likelihood is expressed as α. And β are reflected, and those values in the optimum range are specified, so that a hypothesis verification device with stable performance can be obtained.

  Each of the functional units described above in the block diagram format can be realized by computer hardware and software (computer program) executed on the computer. As this computer hardware, if it has the equipment which handles an audio | voice, what has general-purpose hardware can be used. Such software is also a piece of data that can be stored in a storage medium for distribution.

  The software need not include all the instructions necessary to realize the functions of the hypothesis verification device 94 described above. For example, a desired function can be realized by calling instructions provided in the operating system. You can do it. In other words, any function that realizes the functions of the hypothesis verification device 94 using computer hardware and software resources may be used.

  Further, the speech machine translation apparatus 80 shown in FIG. 3 can also be realized by a computer and software having a general configuration except for a microphone and a board dedicated to speech processing.

  A computer programmed with such software serves as a speech recognition result reliability verification apparatus according to the present invention.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a schematic diagram of the word lattice / graph for demonstrating the operation principle of the hypothesis verification apparatus 94 which concerns on the 1st Embodiment of this invention. It is a schematic diagram of an N-best list for explaining the principle of the hypothesis verification device 94 according to the first exemplary embodiment of the present invention. It is a block diagram of the speech machine translation apparatus 80 using the hypothesis verification apparatus 94 which concerns on the 1st Embodiment of this invention. FIG. 4 is a detailed block diagram of a hypothesis verification device 94 shown in FIG. 3. FIG. 6 is a contour plot showing the results of testing the performance of a single threshold classifier over a wide range of α and β using set 01. FIG. 6 is a contour plot showing the results of testing the performance of a single threshold classifier over a wide range of α and β using set 02. FIG. 7 is a contour plot showing the results of testing the performance of a single threshold classifier in a small range of α and β using set 01. Using a set 0 2 is a contour plot showing the results of testing the classifier performance of a single threshold value in small range of α and beta. It is a graph which compares and shows the ROC curve of a reappearance rate and the generalized word posterior probability when optimal (alpha) and (beta) are used. FIG. 10 is a graph showing the behavior of the system performance with respect to variations in α and β in the vicinity of the optimum point for the two test sets (01, 02).

Explanation of symbols

80 speech machine translation device, 90 ASR decoder, 92 machine translation device, 94 reliability scale method hypothesis verification device, 96 acceptance / rejection determination device, 120 word graph storage unit, 122 target word search unit, 124 posterior probability calculation unit, 126 Word graph update unit, 128 maximum likelihood path search unit

Claims (6)

Receiving a speech recognition result representing a plurality of hypothesized word strings each consisting of a word to which a word posterior probability is given, output from the speech recognition decoder, and verifying the reliability of the speech recognition result based on the word posterior probability A speech recognition result reliability verification device for
For each word included in the speech recognition result, generalized word posterior probability calculating means for calculating a generalized word posterior probability based on the word posterior probability of the word included in the speech recognition result;
Updating means for updating the word posterior probability of each word included in the speech recognition result with the generalized word posterior probability calculated by the generalized word posterior probability calculating means;
Based on the speech recognition result in which the word posterior probability is updated by the updating means, a search is made among the plurality of hypothesis word strings for which the sum of the word posterior probabilities of the words included in the hypothesis word string is maximized. Search means for,
By the sum of word posterior probabilities of hypotheses word string searched by the search means to determine whether to satisfy a predetermined condition, seen including a determining means for verifying the reliability of the speech recognition result ,
Each word included in the speech recognition result is further provided with information for determining a time period during the input utterance to the speech recognition decoder,
The generalized word posterior probability calculating means is:
For searching for a word that is in a time period that overlaps the time period of the word and that matches the word for each word included in the word lattice composed of the plurality of hypothesized word strings from the speech recognition result Word search means,
For each word included in the word lattice composed of the plurality of hypothetical word strings, out of the paths included in the word lattice, the sum of the likelihoods of the paths that pass through the word searched by the word search means, It said word by dividing the sum of the likelihoods of all paths included in the lattice means and the including for calculating the generalization word posterior probabilities of the words, the speech recognition result reliability verification device.   Prior to the calculation of generalized word probabilities by the generalized word posterior probability calculating means, only the word string consisting of the speech recognition results having a higher likelihood than the threshold value determined by a predetermined criterion is selected. The speech recognition result reliability verification apparatus according to claim 1, further comprising means for giving to the generalized word posterior probability calculating means. The means for calculating the generalized word posterior probability is a ratio between the sum of the word posterior probabilities of the words searched by the word search means and the sum of the word posterior probabilities of all words included in the speech recognition result. The speech recognition result reliability verification apparatus according to claim 1 , further comprising: means for calculating a generalized word posterior probability of each word. The generalized word posterior probability p ([w; s, t] | x ₁ ^T ) of the word w in the hypothesis word string (where s and t are the start time and end time of the time period of the word w, respectively) is formula
Given, provided that ^{_{x 1 T = x 1, ...}} , x T is the observed speech sequence, M is the number of words included in the hypothesis of a speech recognition result, s _n and t _n, respectively, word w _Are the start and end times of the _nth word _wn , where p (x _sm ^tm | w _m ) is the acoustic likelihood, and p (w _m | w ₁ ^M ) is the language likelihood, p (x ₁ ^T) is the acoustic observation likelihood, respectively α and β is a predetermined constant, the speech recognition result reliability verification apparatus according to any one of claims 1 to 3. A computer program that, when executed by a computer, causes the computer to operate so as to realize each means of the speech recognition result reliability verification apparatus according to any one of claims 1 to 4 . A computer programmed by the computer program according to claim 5 .