JP6026224B2 - Pattern recognition method and apparatus, pattern recognition program and recording medium therefor - Google Patents

Description

  The present invention relates to a pattern recognition method and apparatus, a pattern recognition program, and a recording medium thereof, and more particularly to a pruning threshold at which a calculation of a path with a low cumulative likelihood is canceled halfway, an index representing the complexity of the recognition process. The present invention relates to a pattern recognition method and apparatus, a pattern recognition program, and a recording medium thereof.

  Speech recognition is formulated as a process for obtaining a word sequence with the highest probability from word chain sequences (word sequences) defined in advance for time-series speech data. FIG. 5 is a block diagram showing a system configuration of a speech recognition engine based on the currently mainstream HMM (Hidden Markov Model), and through an input speech capturing unit 501, a speech segment detection unit 502 and an acoustic analysis unit 503. For the obtained acoustic features,

  (1) a grammar / language model search unit 506 for searching a word sequence constituting a sentence in units of words, and

  (2) The word sequence with the highest posterior probability is obtained by the word dictionary search unit 505 that calculates the cumulative likelihood (probability) while searching for the word that is a component of the language model.

  More specifically, the input voice capturing unit 501 extracts a waveform signal from a microphone and outputs it. The voice section detection unit 502 receives time-series waveform signals as input, sequentially calculates short-time power and the number of zero crossings, cuts out a voice section of one utterance by threshold processing, and outputs the waveform signal. The acoustic analysis unit 503 performs acoustic analysis using the waveform signal of the speech section as input, and outputs time-series data of acoustic features such as MFCC. The acoustic likelihood calculation unit 504 outputs the acoustic likelihood by using the acoustic feature amount as input and collating it with a probability density distribution in the acoustic feature amount space defined by the acoustic model 508 for each HMM state.

  The grammar / language model search unit 506 starts searching for a word sequence from the silent part of the sentence while referring to the language model 510 on the grammar as shown in FIG. 6 or the probabilistic language model as shown in FIG. When the word to be searched is determined, the word to be searched is specified to the word dictionary search unit 505, and when the word dictionary search unit 505 receives notification of the end of the search for the specified word, the word that follows at the next time Is instructed to the word dictionary search unit 505.

  In speech recognition based on a probabilistic language model, the likelihood of the probabilistic language model is added to the cumulative likelihood described below. The word dictionary search unit 505 searches the word specified by the grammar / language model search unit 506 while accumulating the acoustic likelihood for each HMM state output from the acoustic likelihood calculation unit 504 based on the tree structure dictionary 509. I do. When the search proceeds to the HMM state at the end of the word, the grammar / language model search unit 506 is notified of the end of the word search and its cumulative likelihood. Then, the word sequence having the highest posterior probability (likelihood) for the given speech feature amount sequence is output as the final recognition result. When the speech segment search process ends, the recognition result output unit 507 outputs the word sequence having the highest cumulative likelihood as the recognition result.

  In the word dictionary searched by the word dictionary search unit 505, each word is expressed as an HMM sequence. In many cases, words are decomposed into phoneme sequences according to the reading, and the HMM sequences of words are constructed by concatenating phoneme HMMs prepared for each phoneme. In this way, a straight HMM sequence is obtained for each word as shown in FIG.

  The word search is performed by the Viterbi algorithm for the HMM sequence. The hypothesis, which is one of the recognition result candidates, is entered at the beginning of the HMM sequence at the start time (next time after the end time of the previous word), and at the end of the HMM sequence at a certain time (end time). Get out of the state. The Viterbi algorithm outputs, as a probability value, the degree of coincidence between the speech data sequence and the acoustic features of the HMM sequence from the time when the hypothesis enters the HMM sequence to the time when the hypothesis enters. More precisely, this probability value is expressed as a logarithmic likelihood (L = log P), which is called “acoustic likelihood”.

  For example, in speech recognition using the probabilistic language model of FIG. 7 and the tree structure dictionary of FIG. 9, the state hypothesis that is searching for the initial letter / k / following the silence at the beginning of the sentence is “sil- The log 0.050 is added to the logarithm of the language probability 0.050 of “sil-9: 00”, which is the maximum value among the three, “9:00”, “sil-company” and “sil-shopping”. For the state hypothesis searching for / a / next to / k / beyond the first branch, log0.050 before the branch is discarded and the language probabilities of “sil-company” and “sil-shopping” are discarded. The log 0.020 logarithm of 0.020 of the maximum “sil-company” is added. Log 0.020 is discarded and log 0.010 is added to the state hypothesis that has advanced to / m / of “shopping”. Thus, as the search of the tree structure dictionary proceeds and the state hypothesis approaches the leaf, the words are limited and more accurate language probabilities are given.

  Since the word is not fixed during the recognition process, the Viterbi algorithm is executed simultaneously on the HMM sequences of various words. That is, hypotheses that store the acoustic likelihood up to that point are temporarily stored in the HMM sequence states of various words. This hypothesis is called the “state hypothesis”.

  In actual speech recognition, the number of state hypotheses is enormous in order to search for word sequences covering a huge variety. In order to save the number of state hypotheses, common partial HMM sequences are merged from the beginning between different words. This is because if the search for a common partial HMM sequence is started at the same time even between different words, the acoustic likelihood of each state of the common part is equal. By this merging, a “tree structure dictionary” as shown in FIG. 9 is formed.

  However, even if such a recognition target vocabulary is made into a tree structure dictionary, the number of state hypotheses increases explosively. Therefore, the “pruning” process that compares the likelihood of all state hypotheses at each time, leaves only the state hypothesis with high likelihood in the next time search, and discards the state hypothesis with low likelihood during the search. Is done.

  As a typical pruning technique, Non-Patent Document 1 discloses a pruning technique based on a likelihood width that leaves a state hypothesis within a certain likelihood width from the maximum likelihood in all state hypotheses. In Non-Patent Document 2, in order to leave a certain number of state hypotheses from a high likelihood state hypothesis, all state hypotheses are classified into histogram classes based on likelihood, and the cumulative number is constant from the high likelihood class. A pruning technique using a histogram that remains until a value is exceeded is disclosed. Both methods are generally used in combination.

  Furthermore, as a developed form of pruning by likelihood width and pruning by histogram, a method of dynamically changing these threshold values has been proposed.

  In Patent Document 1, as a method for adaptively controlling the search range, the distribution of the cumulative likelihood at each time, more specifically, the likelihood difference between the current time and the maximum likelihood hypothesis one hour before is used as a variable. A method for controlling the threshold for hypothetical pruning is disclosed. Patent Document 2 discloses a method for controlling the likelihood width of pruning based on the likelihood width so that the number of hypotheses after pruning does not exceed a certain number.

H. Ney, D. Mergel, A. Noll and A. Paeseler, "Data-driven search organization for continuous speech recognition," IEEE Transactions on Signal Processing, vol.40, no.2, pp.272-281, (1992 ) V. Steinbiss, B.-H.Tran and H. Ney, "Improvements in beam search," Proceedings of ICSLP 94, pp.397-400 (1994)

JP-A-6-282255 JP 2003-15683 A

  Non-Patent Document 1 and Non-Patent Document 2 attempt to reduce state hypotheses by time-invariant pruning thresholds. Patent Document 1 and Patent Document 2 change the pruning threshold in a time-varying manner, but Patent Document 1 is a control that tries to keep the number of state hypotheses at a constant value, and pruning according to the difficulty of recognition that should be originally performed. It is not processing. Originally, it is considered appropriate to loosen the pruning threshold at times when it is difficult to recognize and compare many hypotheses, and to tighten the pruning threshold at times when recognition is easy and only a small number of hypotheses can be compared.

  The object of the present invention is to solve the problems of the prior art, and to determine a pruning threshold for stopping calculation of a path with a low cumulative likelihood based on an instantaneous perplexity that represents the complexity of the recognition process. An object of the present invention is to provide a pattern recognition method and apparatus for dynamically determining, a pattern recognition program, and a recording medium thereof.

  In order to achieve the above-described object, the present invention collates feature parameters extracted from an input signal with a probabilistic state transition model in which a recognition pattern is expressed in a tree structure in time series. In the pattern recognition in which the state hypothesis is changed while repeating the likelihood calculation and the pruning of each state of the probability model with respect to the feature parameter every time, and the pattern recognition using the maximum likelihood state transition path as the recognition pattern, the following configuration is provided.

  (1) Search means for performing likelihood calculation of each state hypothesis in the search space of the state transition model based on the feature parameter of the input signal, and the language of the state M hypothesis having the highest cumulative likelihood at each time PP calculation means for approximating instantaneous perplexity based on probability, pruning threshold setting means for setting pruning threshold based on perplexity calculation result, and likelihood based on pruning threshold Pruning means for excluding the low state hypothesis from the search target.

(2) The PP calculation means calculates the average number of branches obtained from the language probabilities of the top K words among the plurality of words that each of the top M state hypotheses can reach next on the tree structure. The perplexity is approximated by.
The pattern recognition apparatus according to claim 1.

  (3) The PP calculation means approximates the evaluation sentence set prepared in advance for the task by a test set perplexity with the number of words from the sentence head calculated in advance as a variable.

  According to the present invention, the following effects are achieved.

  (1) Since hypothesis pruning thresholds are set based on instantaneous perplexity, hypothesis pruning thresholds are relaxed at times when speech recognition is difficult, while hypothesis pruning is performed at times when recognition is easy. It is possible to dynamically control the pruning threshold, such as tightening the threshold. As a result, it is possible to efficiently reduce the processing time and to speed up the recognition process while suppressing a decrease in the recognition rate.

  (2) Since the instantaneous PP is approximated based on the top K language probabilities for the top M state hypotheses with high cumulative likelihood, the perplexity calculation load can be reduced.

  (3) If the perplexity is approximately calculated based on the evaluation sentence set, the online (during voice recognition processing) operation is eliminated, and the perplexity can be approximated by the offline operation, so that the processing speed is improved.

It is a functional block diagram of 1st Embodiment of the speech recognition apparatus to which this invention is applied. It is the flowchart which showed the procedure of 1st Embodiment of voice recognition. It is a functional block diagram of 2nd Embodiment of the speech recognition apparatus to which this invention is applied. It is the flowchart which showed the procedure of 2nd Embodiment of speech recognition. It is a functional block diagram of the conventional speech recognition apparatus. It is the figure which showed an example of description grammar. It is the figure which showed an example of the probabilistic language model (bigram). It is the figure which showed an example of the linear HMM series. It is the figure which showed an example of the tree structure dictionary.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Here, a speech recognition apparatus is taken up as a typical application example of a pattern recognition apparatus using a tree structure based on a probability model. FIG. 1 is a system configuration diagram of a speech recognition apparatus according to an embodiment of the present invention, and illustrations of components unnecessary for the description of the present invention are omitted.

  The audio signal input unit 1 converts the input audio signal into a digital signal. The acoustic analysis unit 2 performs acoustic analysis on the audio digital signal to extract acoustic feature parameters, and temporarily stores them. The acoustic feature parameter is a feature vector obtained by analyzing the input speech at regular time intervals (for example, 10 ms: hereinafter referred to as a frame). Therefore, the audio signal is converted into a feature vector time series X = x1, x2,. The grammar / probabilistic language model search unit (first search unit) 3 searches the word series constituting the sentence in units of words based on the description grammar / probabilistic language model 8.

  In the tree structure dictionary search unit (second search unit) 4, the self-transition unit 401 self-transitions each state hypothesis in the search process. The LR transition unit 402 causes each state hypothesis to perform LR transition in the search process. The likelihood calculation unit 403 calculates the acoustic likelihood by collating the time series data of the acoustic feature parameters with the tree structure dictionary 9 in the self transition and the LR transition, and accumulates the cumulative likelihood in the time direction. Find the likelihood. In this embodiment, when the state series of the tree structure dictionary 9 branches into a plurality of branches due to grammatical restrictions, the tree structure dictionary search unit 4 duplicates the state hypothesis by the number of branches and advances the state hypothesis for each branch. And calculate the likelihood.

  A PP (perplexity) approximate calculation unit 404 approximately calculates the perplexity of the stochastic language model. That is, in speech recognition, in general, the larger the vocabulary to be recognized and the more alternative hypotheses, the lower the correct answer rate and the more difficult it is to recognize. On the other hand, in the case of speech recognition based on a stochastic language model, it is considered that the difficulty of recognition changes depending on the time during search. For example, when a speech recognizer accepts various questions of a speaker, the beginning of a sentence may start with a variety of words, making it difficult to recognize, and conversely, the end of the sentence is "?" It seems that it is easy to recognize because it often ends with a regular question form. The difficulty of speech recognition that changes from moment to moment can be measured by the instantaneous perplexity given by the stochastic language model.

For example, using a probabilistic language model composed of the vocabulary w ₁ ,..., W _N , the k-th word s _k from the beginning of the sentence is being searched, and the immediately preceding word is s _k−1 and the previous word is s _{k -2} is determined, the instantaneous perplexity PP (s _k ) is given by the following equation (2) using the instantaneous entropy H (s _k ) of the following equation (1): be able to.

H (sk) = -Σs _k P (s _k | s _k-1 , s _k-2 ) log ₂ P (s _k | s _k-1 , s _k-2 )… (1)

PP (sk) = 2 ^{H (sk)} … (2)

However, it is practically difficult to calculate the sum of all N words of the language probabilities for each word s _k in the speech recognition process. Also, since the preceding word s _k-1 and the preceding word s _{k- 2} also differ depending on the state hypothesis, it is not easy to calculate the perplexity PP (s _k ) for all word sequences that are alive at that time. .

Therefore, in the present embodiment, instead of considering all possibilities, the top M state hypotheses with the highest cumulative likelihood are a plurality of leaves that can be reached next on the tree structure dictionary from each word s _{k being} searched. The word corresponding to is assumed as a candidate. Then, an approximate perplexity is calculated from the language probabilities of the K words having the highest language probabilities among the next reachable words on the tree structure from each of the M state hypotheses. In this embodiment, the perplexity is approximated by the sum of the language probabilities of the top K words having the highest language probabilities. That is, entropy H (s _k ) is represented by the following equation (3).

H (s _k ) = -Σ _m (R ( _m ) Σ _{Sk, m {} P (s _{k, m} | s _{k-1, m} , s _{k-2, m} ) log ₂ P (s _{k, m} | s _{k-1, m} , s _{k-2, m} )}] / Σ _m R ( _m )… (3)

Here, R (m) represents the cumulative likelihood of the m-th state hypothesis among the top M pieces, and a weighted average is taken by the cumulative likelihood. In the above case, for each state hypothesis, the sum of the top K language probabilities is taken regardless of the location in the word dictionary, but the look-ahead value of the language probability that changes depending on the location in the word dictionary An approximation can also be made using Q (m | s _{k-1, m} , s _{k-2, m} ). That is, the entropy H (s _k ) may be represented by the following equation (4).

H (s _k ) = -Σ _m (R ( _m ) Q (m | s _{k-1, m} , s _{k-2, m} ) log ₂ Q (m | s _{k-1, m} , s _{k-2, m} )] / Σ _m R ( _m )… (4)

Returning to FIG. 1, the pruning threshold value setting unit 405 sets the pruning threshold value based on the instantaneous perplexity PP (s _k ). That is, if the pruning is based on the likelihood width, the likelihood width is set based on PP (s _k ), and if the pruning is based on the histogram, the maximum number of hypotheses that can be held is based on PP (s _k ). Is set. In the case of pruning by likelihood width, the likelihood width is given by a monotonically increasing function of instantaneous perplexity PP (s _k ), and in the case of pruning by histogram, the maximum number of hypotheses is instantaneous perplexity. It is given by a monotonically increasing function of tee PP (s _k ). The pruning unit 406 performs pruning based on the pruning threshold.

  The word hypothesis output unit 5 outputs the word hypothesis of the state hypothesis that has advanced to the end of the word. The search and the word hypothesis output by the first and second search units 3 and 4 are repeated until the input of the time series data of the acoustic feature parameters is completed. When the input of the time series data is completed, the back trace unit 6 executes the back trace to the state series having the highest cumulative likelihood among the word hypothesis sets 7 that have reached the last grammatical HMM state. The recognition result is determined.

  FIG. 2 is a flowchart showing a speech recognition procedure to which the pattern recognition method of the present invention is applied, and mainly shows the operation of the tree structure dictionary search unit (second search unit) 4.

  In step S1, one of valid state hypotheses is selected as the current calculation target. In step S2, the self-transition unit 401 performs self-transition on the current state hypothesis, and the acoustic likelihood is calculated. In step S3, the current acoustic likelihood is added to the cumulative likelihood up to now, and the cumulative likelihood is updated. In step S4, it is determined whether self-transition and likelihood calculation have been completed for all state hypotheses corresponding to the current timing. If not completed, the process returns to step S1, and the above-described processes are repeated for other state hypotheses to be transitioned at the current timing.

  When the self-transition and the likelihood calculation are completed for all the state hypotheses corresponding to the current timing, the process proceeds to step S5, and one of the valid state hypotheses corresponding to the current timing is selected again as a calculation target. In step S6, the LR transition unit 402 performs LR transition on the current state hypothesis, and the acoustic likelihood is calculated. In step S7, the current acoustic likelihood is added to the cumulative likelihood up to now, and the cumulative likelihood is updated.

  In step S8, it is determined whether or not the tree structure has branched due to the current LR transition. If it is branched, the process proceeds to step S9, where language likelihood prefetching is performed based on the probabilistic language model, and the maximum language likelihood values (prefetching values) of all reachable words are added to the cumulative likelihood. .

  In step S10, when there is a state hypothesis of self-transition at the transition destination, the cumulative likelihood is compared with the cumulative likelihood of the state hypothesis after LR transition, and the smaller one is discarded with the larger state hypothesis remaining. Is done. In step S11, it is determined whether or not the above LR transition has been completed for all state hypotheses to be transitioned at this timing. If not completed, the process returns to step S5, and the above processes are repeated for other state hypotheses that should be transitioned at the current timing.

  Thereafter, when the above-described processes are completed for all state hypotheses to be transitioned at this timing, the process proceeds to step S12, and instantaneous approximate calculation of PP is performed based on the above equations (2) and (3). That is, in this embodiment, instead of considering all the possibilities, instantaneous PP is approximated by the sum of the top K language probabilities for the top M state hypotheses with high cumulative likelihood.

  In step S13, a pruning threshold value corresponding to the instantaneous PP is set. In this embodiment, the likelihood width and the maximum number of hypotheses are set by an instantaneous monotonically increasing function of PP. In step S14, the state hypothesis is pruned based on the cumulative likelihood using the likelihood width and the maximum number of hypotheses as a pruning condition.

  In step S15, one of the state hypotheses remaining after pruning is selected. In step S16, it is determined whether or not the selected state hypothesis is a state hypothesis at the end of the word. If it is the state hypothesis at the end of the word, the process proceeds to step S17 and the word hypothesis is output. In step S18, a hypothetical state hypothesis for transitioning to the first state of the next word is set. In step S19, it is determined whether or not the above processing has been completed for all state hypotheses remaining after pruning. If not completed, the process returns to step S15, and each process is repeated while changing the state hypothesis. In step S20, it is determined whether or not there is a next frame. If there is a next frame, the process returns to step S1, and the above-described processes are repeated for the acoustic feature parameter of the next frame.

  When the above processing is completed for all frames and the search reaches the end-of-sentence frame, in step S21, the word hypotheses that have reached the last state in the grammar are sorted in the order of their cumulative likelihoods. A backtrace is performed on a plurality or only one word hypothesis having a higher degree, and a recognition result is output.

  In the first embodiment, since speech recognition processing is performed online, a system with a large vocabulary or a system with low processing capability may cause a delay in real-time processing. On the other hand, in the second embodiment described below, the processing speed can be improved by eliminating the online (during speech recognition processing) calculation and approximating with the offline calculation.

  FIG. 3 is a system configuration diagram of the speech recognition apparatus according to the second embodiment of the present invention, and FIG. 4 is a flowchart showing the operation of the present embodiment. The same reference numerals as those described above denote the same or equivalent parts. Yes.

  In the second embodiment of the present invention, roughly grasping the tendency for perplexity to decrease from the first word to the last word of a phrase (a unit that is punctuated by punctuation marks), for a plurality of evaluation sentences (test sets) prepared in advance, The test set perplexity 10 calculated off-line using the number of words from the beginning of the phrase as a parameter is given.

Here, if the lth of the evaluation sentences of all L sentences is composed of the words s _{1, l} ,…, s _{Kl, l} , the instantaneous perplexity PP (sk for the kth word from the beginning ) Can be given by the following equation (5) using the instantaneous entropy H (sk) of the following equation (6).

PP (s _k ) = 2 ^{H (k)} … (5)

H (k) =-(1 / L) Σl log ₂ P (s _{k, l} | s _k-1 , _l , s _k-2 , _l )… (6)

  In this case, for the top M state hypotheses with a high cumulative likelihood, it is possible to count how many words are being searched from the beginning of the phrase and use the average value thereof.

  In FIG. 3, the PP reference unit 407 refers to the test set perplexity 10 corresponding to the word position from the beginning of the phrase for the top M state hypotheses of the cumulative likelihood in step S12a of FIG. The perplexity is approximately calculated by obtaining the weighted average value using the degree as a weighting coefficient. In step S13 and subsequent steps, as in the first embodiment, a pruning threshold corresponding to the PP is set, and pruning of the state hypothesis based on the cumulative likelihood is performed.

  A series of pattern recognition procedures including the above search procedure can be described in a program language that can be executed by a computer, and the program is recorded on a storage medium such as a CD-ROM or a DVD and read into the computer. Thus, the pattern recognition apparatus can be configured.

  In the above-described embodiment, the present invention has been described by taking voice recognition as an example. However, the present invention can be similarly applied to other pattern recognition.

  DESCRIPTION OF SYMBOLS 1 ... Speech signal input part, 2 ... Acoustic analysis part, 3 ... Grammar / stochastic language model search part (1st search part), 4 ... Tree structure dictionary search part (2nd search part), 5 ... Word hypothesis output part, 6 ... Backtrace part, 7 ... Word hypothesis set, 8 ... Descriptive grammar / probabilistic language model, 9 ... Tree structure dictionary, 10 ... Test set PP, 401 ... Self transition part, 402 ... LR transition part, 403 ... Likelihood calculation 404, PP approximate calculation unit, 405 ... Pruning threshold setting unit, 406 ... Pruning unit

Claims (9)

The feature parameters extracted from the input signal and the probabilistic state transition model in which the recognition pattern is expressed in a tree structure are collated in time series, and the likelihood of each state of the probability model for the feature parameter for each branch of the tree structure. In a pattern recognition device that changes the state hypothesis while repeating the degree calculation and pruning, and uses the most likely state transition path as the recognition pattern,
Search means for performing likelihood calculation of each state hypothesis in the search space of the state transition model based on the feature parameter of the input signal;
PP calculation means for approximately calculating instantaneous perplexity based on the language probabilities of the top M state hypotheses with cumulative likelihood at each time,
A pruning threshold setting means for setting a pruning threshold based on the calculation result of the perplexity;
A pattern recognition apparatus comprising: a pruning unit that excludes a state hypothesis having a low likelihood from a search target based on the pruning threshold.   The PP calculation means is configured such that, among the plurality of words that each of the top M state hypotheses can be reached next on the tree structure, the language probability is estimated from the language probabilities of the top K words. The pattern recognition apparatus according to claim 1, wherein the perplexity is approximated by:   The pattern recognition apparatus according to claim 1, wherein the pruning threshold setting unit gives a likelihood width from a maximum value of the cumulative likelihood at each time by a monotonically increasing function for the perplexity. .   The pattern recognition apparatus according to claim 1, wherein the pruning threshold setting unit gives the maximum number of hypotheses that can be held at each time by a monotonically increasing function for the perplexity.   The pattern recognition apparatus according to claim 1, wherein the PP calculation unit approximates an evaluation sentence set prepared in advance for the task by a test set perplexity having a number of words from a sentence head calculated in advance as a variable. . The feature parameters extracted from the input signal and the probabilistic state transition model in which the recognition pattern is expressed in a tree structure are collated in time series, and the likelihood of each state of the probability model for the feature parameter for each branch of the tree structure. In the pattern recognition method where the state hypothesis is changed while repeating the degree calculation and pruning, and the most likely state transition path is the recognition pattern,
A procedure for performing likelihood calculation of each state hypothesis in the search space of the state transition model based on the feature parameters of the input signal;
At each time, a procedure for approximating instantaneous perplexity based on the language probabilities of the top M state hypotheses with cumulative likelihood,
A procedure for setting a pruning threshold based on the calculation result of the perplexity;
And a procedure for excluding a low-likelihood state hypothesis from a search target based on the pruning threshold.   In the procedure for calculating the perplexity, among the plurality of words that each of the top M state hypotheses can be reached next on the tree structure, the language probability is obtained from the language probabilities of the top K words. The pattern recognition method according to claim 6, wherein the perplexity is approximated by an average number of branches.   A pattern recognition program for causing a computer to execute the pattern recognition method according to claim 6 or 7. A recording medium of a pattern recognition program in which the pattern recognition program according to claim 8 is stored so as to be readable by a computer.