JP5008078B2 - 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 pattern recognition method and apparatus for performing pattern recognition such as speech recognition using a probabilistic state transition model typified by an HMM. And a pattern recognition program and a recording medium thereof.

  In speech recognition, the word string closest to the input speech signal is determined based on the similarity with the word expressed as the state series. The HMM (Hidden Markov Model) is one of probability models suitable for expressing words and phonemes constituting the words, and each state of the HMM has a state transition probability and an output probability density function. Hereinafter, a conventional speech recognition method will be described by taking the case of using the HMM as an example.

  In the speech recognition apparatus, the recognition process proceeds according to a grammar in which a set of recognizable sentences is described as a network in units of words and a word dictionary in which readings of words constituting the sentence (phoneme strings) are described. FIG. 9 is a diagram showing an example of a grammar, and here, a case where four voices “Ito Ito”, “Ito Itoi”, “Imai Ito”, “Ito Doi” are identified is described as an example. .

  The grammar shown in FIG. 9 is a state transition diagram in which the state “1” indicated by the circled number 1 is the beginning (start of sentence) and the state “5” is the end (end of sentence), and is a word associated with an arrow. To transition between states. Each word constituting the grammar is expressed as an HMM state sequence according to its reading (phoneme string), and a set of words included in the word dictionary is expanded as a tree structure dictionary as shown in FIG.

  In the tree structure dictionary, each word is decomposed into a phoneme string, and if it is the word “Itoi”, it is expanded into four phoneme “i”, “t”, “o”, and “i” strings. Each phoneme is usually composed of about three states (HMM states). The tree structure dictionary is a state transition diagram in which a branch expands toward the right by merging partial state sequences that are common from the beginning among words expressed as HMM state sequences. In the tree structure dictionary of FIG. 10, the HMM state series corresponding to “I” at the head of the word is merged with three words “Ito”, “Itoi”, and “Imai”, and “Ito” and “Itoi” are further merged. The HMM state series corresponding to “Ito” is merged. In addition, the HMM state sequence corresponding to “d” at the beginning of the word is merged between “Doi” and “Is”. “Sil” in the figure represents a silent period (silence).

  In the speech recognition process, a token called a state hypothesis transitions from the left to the right in the tree structure dictionary from the HMM state at the beginning of the word in the tree structure dictionary shown in FIG. 10 in accordance with the grammatical restrictions shown in FIG. When the state hypothesis reaches the HMM state at the end of the word, a history called a word hypothesis is left and the state transitions to the transition destination state of the word in the grammar of FIG. If the transition destination state is not the end of the sentence, the tree structure dictionary is similarly searched from the next time according to grammatical constraints.

  While the state hypothesis transitions from the left to the right in the HMM state sequence in the tree structure dictionary, a word-likeness score (cumulative likelihood) is calculated for the input speech. Each HMM state constituting the tree structure dictionary has a probability distribution (output probability density function) that outputs likelihood with respect to the input of acoustic feature parameters. Also, transition probabilities (state transition probabilities) are defined for transitions between HMM states. The cumulative likelihood is calculated by accumulating these probabilities in the time direction.

  This cumulative likelihood is determined in the word hypothesis together with the index of the preceding word hypothesis when the state hypothesis reaches the terminal state of each word and leaves a history called the word hypothesis for the recognition result determination process described later. Stored.

  Each hypothesis that has transitioned to each HMM state at every fixed period that analyzes the audio signal to obtain acoustic feature parameters further transitions to its own HMM state (self-transition) and to the right-hand HMM state (LR transition) is repeated at the same time. At this time, if the cumulative likelihood in which the state j exists in the t-th frame is αj (t), the cumulative likelihood αj (t) is expressed by the following equation (1). Here, αij is a transition probability from the state i to the state j, and bj (ot) is a probability that the state j outputs the acoustic feature quantity ot. Self-transition is considered as the case of i = j in the following equation (1).

  Search space for self-transition and LR transition when searching for a word sequence consisting of N HMM states for a speech signal composed of T frames, that is, when a state hypothesis transitions through an HMM state sequence FIG. 11 shows (trellis). The trellis space is a lattice graph showing a possible state sequence with the horizontal axis as the observation sequence and the vertical axis as the state, and each state sequence is a line segment representing a point at each time (a circle). It is represented by a broken line connected with

  As shown in FIG. 11, there are many paths that reach the state j at the timing of the t-th frame, but since speech recognition is aimed at finding the most likely path (maximum likelihood path), In each HMM state, a Viterbi search that leaves a state hypothesis with a high cumulative likelihood is performed according to the following equation (2).

  Since the speech recognition process needs to search all word chains allowed by the grammar, a number of state hypotheses transition to their own HMM state at the same time (in FIG. 11, self-transition to the right side). And a transition to another adjacent HMM state (LR transition to the lower right neighbor in FIG. 11), the amount of calculation becomes enormous. In order to suppress this increase in the amount of calculation, “pruning” is generally performed during the Viterbi search, in which a state hypothesis with a low probability is excluded from the search space.

  After that, when the search process ends due to some condition for determining the end of speech, such as power reduction, from the word hypotheses that reach the end of the sentence at the end time, the word that precedes the one with the highest cumulative likelihood from the beginning of speech A sequence (word history up to this word hypothesis) is derived, and this is a final recognition result candidate. This operation is called backtrace.

  In the above-mentioned pruning, it is required to efficiently reduce the number of state hypotheses while always leaving the state hypothesis having the maximum cumulative likelihood when the search for the entire utterance is completed. Conventional pruning leaves the state hypothesis with a high cumulative likelihood in each frame of the Viterbi search, and removes the state hypothesis with a low cumulative likelihood from the search target for the next frame and the following two methods. Often.

  (1) Pruning by beam width

  In this pruning method, the state hypothesis that the likelihood is within a certain range is left as a search space at the next time from the maximum likelihood at the time being processed, and the state hypothesis with a likelihood that is less than a certain range is Excluded from the time search space. That is, the likelihood of time t and state j is compared with the maximum likelihood in all state hypotheses at the same time. When the following equation (3) is satisfied, the state j is left in the search space at the next time, and when the following equation (4) is satisfied, the state j is excluded from the search space at the next time. θpruning is a pruning threshold and is a positive real number and is called a beam width.

  (2) Pruning with histogram

  In the above-described pruning with the beam width, a large number of state hypotheses may remain in the range of the beam width, and the processing time may be increased more than expected. On the other hand, in this pruning method, the state hypotheses to be left are limited by the number. However, in order to strictly limit the number by the number, it is necessary to sort the number of state hypotheses in the order of score (cumulative likelihood), but pruning is performed using a histogram in order to avoid sorting costs.

  In other words, an appropriate category is set in advance for the difference from the maximum log likelihood in all state hypotheses, a histogram is created by recording which category the all state hypothesis corresponds to, and finally the cumulative likelihood. The frequency is accumulated in order from the highest category. Then, the state hypotheses up to the section where the cumulative frequency exceeds a certain number are left as search targets for the next frame, and the state hypotheses of the subsequent sections are excluded. The number threshold is called the maximum allowable hypothesis number.

  Patent Documents 1 to 5 disclose a technique (prior art A) for adaptively setting the beam width and the maximum allowable number of hypotheses, and Patent Document 6 discloses the diversity of state hypotheses left by pruning. A technique (prior art B) for preventing the loss of data is disclosed.

  Patent Document 1 discloses a technique for estimating a fluctuation range of rank in an unsearched portion by a probability calculation unit using an acoustic model with a small calculation amount, and dynamically setting a beam width and the maximum allowable number of hypotheses based on an estimated value. It is disclosed. Patent Document 2 discloses a technique for dynamically setting the beam width and the maximum allowable number of hypotheses by calculating perplexity, which is an index of the number of grammatical options.

  Patent Document 3 discloses a technique for counting the number of hypotheses remaining with the current beam width and adaptively setting the subsequent beam widths according to the number of remaining hypotheses. Patent Document 4 discloses a technique for dynamically determining a beam width by a neural network using a current search position (depth) in a grammar as a parameter.

Patent Document 5 discloses a technique for counting the number of hypotheses in speech recognition based on DP matching and adaptively setting a threshold for the distance after that according to the number of remaining hypotheses. Patent Document 6 discloses a technique for performing pruning for each acoustic model when a search process is performed using a plurality of acoustic models (that is, a plurality of tree structure dictionaries) together.
JP 2001-75596 A Japanese Patent Application Laid-Open No. 11-119793 Japanese Patent Laid-Open No. 10-153999 JP-A-6-282295 JP-A-5-35292 JP 2005-10464 A

  The purpose of pruning is to efficiently search the wide search space of the tree structure dictionary with few state hypotheses so that the path with the maximum likelihood can be finally found. In the conventional pruning method described above, state hypotheses may remain concentrated in some areas of the tree structure dictionary, and state hypotheses may not remain in other wide areas. This allows efficient search. I can't say that.

  Even if the improved technique belonging to the above-mentioned conventional technique A is applied, the above problem is not fundamentally solved. In the technique belonging to the conventional technique B, the pruning in the case of using a plurality of acoustic models is made efficient, but the pruning in the single tree structure dictionary cannot be made efficient.

  The object of the present invention is to solve the above-mentioned problems of the prior art, and even if a single tree structure dictionary is used, the state hypothesis can be obtained without lowering the recognition rate by utilizing the structural features of the tree structure dictionary. It is an object to provide a pattern recognition method and apparatus, a pattern recognition program, and a recording medium thereof that can efficiently prune.

  In order to achieve the above-described object, the present invention collates a feature parameter extracted from an input signal with a probabilistic state transition model in which a recognition pattern is expressed in a tree structure, and obtains a probability model for the feature parameter. The pattern recognition apparatus that changes the state hypothesis while calculating the likelihood of each state and uses the most likely state transition path as a recognition pattern is characterized by the following measures.

  (1) Search means for performing likelihood calculation in the search space of the state transition model based on feature parameters of the input signal, area dividing means for virtually dividing the search space into a plurality of areas, and the same time Distribution means for allocating the state hypothesis to be transferred to any one of the areas, and pruning means for excluding a low-likelihood state hypothesis from the search target in the same area for each of the divided areas. Features.

  (2) It further includes pruning condition setting means for setting a pruning condition for each divided area.

  According to the feature (1) described above, the search space of the state transition model is divided into a plurality of regions, and pruning is performed independently for each region, so that one region and the other region are independent. Thus, a state hypothesis having a higher cumulative likelihood can be left. Therefore, the cumulative likelihood of each state hypothesis belonging to one region is relatively higher than the cumulative likelihood of each state hypothesis belonging to another one region, and one region and another region are not divided. The state hypothesis of the other region whose cumulative likelihood is not higher when pruning all together can be left by pruning for each region.

  According to the feature (2) described above, the search space of the state transition model is divided into, for example, a word determination area where the word is fixed and a word unconfirmed area where the word is not fixed, and pruning the word determination area The condition is made stricter than the pruning condition of the word undetermined area, or the search space of the state transition model is divided into multiple areas based on the total number of words branching from each state, and the area with fewer total branching words If the pruning condition is tightened, the state hypothesis can be pruned more efficiently without reducing the recognition rate.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the best embodiment of the present invention will be described in detail with reference to the drawings. Here, first, the outline of the present invention will be described using the tree structure dictionary of place names beginning with the phoneme string “a” shown in FIG.

  Of the place names that begin with the phoneme sequence "ai", place names that start with the phoneme sequence "aio" are limited to "aioi", so if the state hypothesis passes the search space of the phoneme sequence "aio", the word (pattern) Is fixed. Similarly, there are three place names that begin with the phoneme string “aika”: “Akikacho”, “Aikawa”, and “Aikan Umetsubo”, but place names that start with the phoneme string “aikam” are limited to “Akikacho”. Therefore, if the state hypothesis passes the search space for the phoneme sequence “aikam”, the word is determined.

  Here, when pruning is performed for all N1 state hypotheses that have self-transitioned or LR-transitioned at a certain time, the pruning is performed for all N1 state hypotheses in the conventional technique, and the beam Only the top n1 state hypotheses, which are determined by the width and the maximum allowable number of hypotheses, remain. At this time, in order to leave the possibility that the state hypothesis reaches many words, it is desirable that the state hypothesis remains in the top n1 without omission over a wide range of the search space. However, in reality, there are cases where only the state hypothesis of the region where the word is determined occupies the upper rank, and in such a case, the possibility of reaching many words is lost at once.

  On the other hand, the search space is divided into, for example, a region where a word is fixed (word fixed region) and a region where a word is not fixed (word unconfirmed region), and pruning separately in each region. The above-mentioned technical problems can be solved by leaving a higher level hypothesis.

  That is, among all N1 state hypotheses that have self-transitioned or LR-transitioned at the same time, pruning is performed only on all N2 state hypotheses in the word determination area, and the top n2 state hypotheses remain. On the other hand, pruning is performed on all N3 state hypotheses in the word unconfirmed region, and the top n3 state hypotheses are left. In this way, since state hypotheses are not unevenly distributed only in a partial region of the search space, the total number of state hypotheses can be reduced without reducing the recognition rate.

  Furthermore, for the word fixed area, all the state hypotheses that exist in a single row will reach a single word, so even if the pruning condition is tightened and some of the state hypotheses are subject to pruning, There is a high possibility that a word hypothesis having a likelihood close to the maximum likelihood is obtained.

  In the present invention, based on such considerations, the search space is divided into a plurality of regions, and for each region, the likelihood is calculated between state hypotheses that transition at the same time within the same region. By pruning hypotheses, the total number of state hypotheses can be reduced without lowering the recognition rate.

  FIG. 2 is a block diagram showing the configuration of the main part of a speech recognition apparatus to which the pattern recognition method of the present invention is applied.

  The audio signal input unit 11 converts the input audio signal into a digital signal. The acoustic analysis unit 12 acoustically analyzes the audio digital signal to extract acoustic feature parameters, and stores them in the parameter storage unit 13. 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 sequence of feature vectors X = x1, x2,.

  The search database 14 stores in advance a search grammar and a tree structure dictionary (probabilistic state transition model of a phoneme model). An acoustic model created in advance is registered in the acoustic model database 17. The search unit 15 compares the time series data of the acoustic feature parameters with the search grammar, the tree structure dictionary, and the acoustic model to calculate the acoustic likelihood, and accumulates the likelihood in the time direction to calculate the cumulative likelihood. A likelihood calculation unit 151 for obtaining a degree, a self-transition unit 152 for self-transitioning each state hypothesis in the search process, an LR transition unit 153 for LR transition, and a branch for excluding a state hypothesis having a low likelihood in the search process from the search target A trimming unit 154 and a word hypothesis output unit 155 that outputs a word hypothesis of a state hypothesis advanced to the end of the word are included.

  When the state series of the tree structure dictionary branches into a plurality of branches due to grammatical constraints, the search unit 15 duplicates the state hypotheses by the number of branches, and advances the state hypothesis for each branch to calculate the likelihood. The recognition result determination unit 16 sorts all state hypotheses that have reached the last HMM state in the grammar based on the cumulative likelihood, and performs a backtrace on the state series with the highest cumulative likelihood to determine the recognition result To do.

  FIG. 3 is a functional block diagram showing the configuration of the main part of the pruning unit 154. The region dividing unit 154a that divides the search space of the state transition model into a plurality of regions and the state hypothesis that changes at the same time are shown. A state hypothesis allocating unit 154b that distributes to one of the divided regions, a pruning condition setting unit 154c that sets a pruning condition for each of the divided regions, and the divided regions within the same region. And a region-specific pruning unit 154d that excludes a low-likelihood state hypothesis from the search target.

  The region-specific pruning unit 154d may repeat pruning for each region with a single pruning unit as long as processing within a practical time is possible, or by providing a plurality of pruning units. The pruning may be performed in parallel.

  FIG. 4 is a flowchart showing the procedure of speech recognition to which the pattern recognition method of the present invention is applied, and mainly shows the operation of the search unit 15.

Here, the recognition target speech signal input to the speech signal input unit 11 is acoustically analyzed for each frame by the acoustic analysis unit 12, and the acoustic feature parameter of each frame is accumulated in the parameter storage unit 13. start. Further, it is assumed that the search space is divided into a “word determined region” and a “word undefined region” by the region dividing unit 154a as shown in FIG.

  In step S1, as a search process initialization, a hypothetical state hypothesis that transitions to the top state of the top word of the grammar is set. In step S2, the acoustic feature parameters stored in the parameter storage unit 13 are captured in the normal order from the first frame of the recorded speech section. In step S3, one of the valid state hypotheses is selected as the current calculation target. In step S4, self-transition is performed, and the likelihood is calculated and updated. In step S5, 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 S3, and the above-described processes are repeated for other state hypotheses to be transitioned at this 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 S6, and one of the valid state hypotheses corresponding to the current timing is selected again as a calculation target. In step S7, each state hypothesis is LR transitioned. In step S8 Vit e rbi search is performed. In step S9, it is determined whether or not the above-described LR transition and Viterbi search have been completed for all state hypotheses to be transitioned at the current timing. If not completed, the process returns to step S6, and the above-described processes are repeated for other state hypotheses that should transition at the current timing. Thereafter, when the above-described processes are completed for all the state hypotheses to be transitioned at the current timing, the process proceeds to step S10, and an area-specific pruning process is performed in which pruning is performed independently for each area.

  FIG. 5 is a flowchart showing the procedure of the area-specific pruning process. In step S101, one of valid state hypotheses corresponding to the current timing is selected as a calculation target. In step S102, the selected current state hypothesis is distributed to any region. In step S103, it is determined whether or not the distribution has been completed for all state hypotheses, and if not completed, the process returns to step S101 and the above-described distribution process is repeated.

  After that, when all the state hypotheses are distributed to any region, in step S104, one of the divided regions is selected as the current pruning region. In step S105, pruning conditions are set based on the cumulative likelihood of all state hypotheses in the region. In step S106, one of the state hypotheses in the region is selected as the current pruning target. In step S107, pruning is performed to leave only the state hypotheses whose cumulative likelihood is higher than the pruning condition and exclude other state hypotheses from the next search.

  In this embodiment, each likelihood αj (t) at time t and state j is compared with the maximum likelihood αmax (t) in all state hypotheses for each region of the search space, and the following equation (7) is obtained. The satisfied state hypothesis is left in the search space at the next time, and the state hypothesis satisfying the following equation (8) is excluded from the search space at the next time. θpruning is the beam width, and the pruning condition setting unit 154c may set the same beam width for each region or may set different beam widths.

  In addition, if the search space is divided into the word decision area and the word undetermined area as described above, the recognition rate is improved by making the pruning condition of the word decision area stricter than the pruning condition of the word undetermined area. The state hypothesis can be efficiently pruned without lowering.

  In step S108, it is determined whether or not the above-described processing has been completed for all state hypotheses in the current region. If not, the processing returns to step S106 and the above-described processing is repeated while changing the state hypothesis. It is. In step S109, it is determined whether or not the above-described processing has been completed for all the regions. If not, the processing returns to step S104, and the above-described processing is repeated while changing the regions.

  Returning to FIG. 4, in step S11, one of the state hypotheses remaining after pruning is selected. In step S12, it is determined whether or not the selected state hypothesis is a state hypothesis at the end of the word, and if it is a state hypothesis at the end of the word, the process proceeds to step S13 and the word hypothesis is output. In step S14, a hypothetical state hypothesis is set for transitioning to the head state of the next word on the grammar. In step S15, it is determined whether or not the above-described processing has been completed for all state hypotheses remaining after pruning. If not completed, the process returns to step S11, and the processes described above are repeated while changing the state hypothesis.

  In step S16, it is determined whether or not there is a next frame. If there is a next frame, the process returns to step S2 to acquire the acoustic feature parameter of the next frame and the above-described processes are repeated.

  When the above-described processing is completed for all frames and the search reaches the end-of-sentence frame, in step S17, all state hypotheses that have reached the last HMM state in the grammar are displayed in ascending order of their cumulative likelihoods. Sorted. In step S18, a backtrace is performed on a plurality of or only state hypotheses with the highest cumulative likelihood, and a recognition result is output.

  In the above-described embodiment, the present invention has been described by taking an example in which the state hypothesis is pruned based on the beam width. However, the present invention is not limited to this, and the cumulative likelihood of the state hypothesis is not limited thereto. The present invention can be similarly applied to a case in which degrees are histogrammed and pruned based on the maximum allowable number of hypotheses.

In the above-described embodiment, the present invention has been described by taking speech recognition as an example. However, the present invention can be similarly applied to other pattern recognition such as search for protein sequences in genome analysis.

  Further, in the above-described embodiment, the search space is divided into the word determination region and the word unconfirmed region, and pruning is performed independently for each region. However, the present invention is limited to this. Instead, the search space may be divided into a plurality of regions based on one or a combination of the following rules.

  (1) As shown in FIG. 6, the search space is divided into a plurality of regions in accordance with the distance from the root of the tree structure dictionary. At this time, if the pruning condition is made different for each region, it is desirable to narrow the pruning condition by narrowing the beam width and the maximum allowable number of hypotheses in a region farther from the root of the tree structure dictionary. In the illustrated example, the beam width and the maximum allowable number of hypotheses are first region> second region> third region, and the pruning condition of the third region is the strictest.

  (2) As shown in FIG. 7, the search space is divided for the total number of patterns that branch after each HMM state. At this time, if the pruning conditions are made different for each region, it is desirable to narrow the pruning condition by narrowing the beam width in the region where the total number of patterns is smaller. In the illustrated example, the total number of patterns branched from the first region is 7, the total number of patterns branched from the second region is 3, and the total number of patterns branched from the third region is 4. Therefore, the beam width and the maximum allowable number of hypotheses are first region> third region> second region, and the pruning condition of the second region becomes the strictest.

  (3) As shown in FIG. 8, the search space is divided into a plurality of regions based on the root type of the tree structure dictionary (here, the phonemes at the beginning of words). At this time, if the pruning condition is varied for each region, it is desirable to set the beam width for each group with each group as the same region. Such a dividing method is effective when the total number of patterns branched thereafter is specific according to the content of the root of the tree structure dictionary.

It is the figure which showed an example of the tree structure dictionary for demonstrating the outline | summary of this invention. It is the block diagram which showed the structure of the principal part of the speech recognition apparatus to which this invention is applied. It is a functional block diagram of the search part shown in FIG. It is the flowchart which showed the procedure of the speech recognition to which this invention is applied. It is the flowchart which showed the procedure of the pruning process according to area | region. It is a figure for demonstrating the modification (the 1) of this invention. It is a figure for demonstrating the modification (the 2) of this invention. It is a figure for demonstrating the modification (the 3) of this invention. It is the figure which showed an example of grammar. It is the figure which showed an example of the tree structure dictionary. It is the figure which showed an example of the space (trellis) of a self transition and LR transition.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Audio | voice signal input part, 12 ... Acoustic analysis part, 13 ... Parameter memory | storage part, 14 ... Search database, 15 ... Search part, 16 ... Recognition result determination part, 151 ... Likelihood calculation part, 152 ... Self-transition part, 153 LR transition unit 154 Pruning unit 154a Area dividing unit 154b State hypothesis sorting unit 154c Pruning condition setting unit 154d Area pruning unit 155 Word hypothesis output unit

Claims (7)

The feature parameters extracted from the input signal are compared with the probabilistic state transition model whose recognition pattern is expressed in a tree structure, and the state hypothesis is changed while calculating the likelihood of each state of the probability model for the feature parameter. In the pattern recognition device that uses the most likely state transition path as a recognition pattern,
Search means for performing likelihood calculation in a search space of the state transition model based on a feature parameter of an input signal;
A region dividing means for virtually dividing the search space into a word determination region where a word is fixed and a word unconfirmed region where the word is not fixed ;
State hypothesis allocating means for allocating the state hypothesis that transitions at the same time to any of the divided regions;
A pattern recognition apparatus, comprising: a pruning unit for each region that excludes, from the search target, a state hypothesis having a low likelihood in the same region for each of the divided regions.   The pattern recognition apparatus according to claim 1, further comprising a pruning condition setting unit that sets a pruning condition for each of the divided areas. The pruning condition setting means, a pattern recognition apparatus according pruning conditions of the word fixing region to claim 2, characterized in that stricter than pruning conditions of said word undetermined area. Acoustic analysis means for extracting acoustic feature parameters from the audio signal in units of frames;
Pattern recognition apparatus according to any one of claims 1 to 3, characterized in that the speech recognition based on the acoustic feature parameter. The computer collates the feature parameter extracted from the input signal with a probabilistic state transition model in which the recognition model is expressed in a tree structure, and calculates the likelihood of each state of the probability model with respect to the feature parameter In a pattern recognition method in which a hypothesis is transitioned and the most likely state transition path is a recognition pattern,
A computer executing likelihood calculation in a search space of the state transition model based on a feature parameter of an input signal;
The computer virtually divides the search space into a word determination area where a word is fixed and a word undetermined area where a word is not fixed ;
A procedure in which the computer distributes the state hypothesis that transitions at the same time to any one of the areas;
A pattern recognition method , wherein the computer includes a region pruning procedure for excluding a state hypothesis having a low likelihood in the same region from a search target for each of the divided regions. A pattern recognition program for causing a computer to execute the pattern recognition method according to claim 5 . A recording medium in which the pattern recognition program according to claim 6 is stored so as to be readable by a computer.

Images