JPH1097275A - Large-vocabulary speech recognition system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain the large-vocabulary speech recognition system which can perform real-time operation with inexpensive hardware constitution even for a very large vocabulary by composing a parser of a forward and a backward operation part and making a viterbi search driven under restriction conditions of tri-phoneme units considering phoneme environment. SOLUTION: This system is equipped with the parser 5 including the forward operation part 6 and backward operation part 7. The forward operation part 6 makes a viterbi search under restriction conditions of phoneme units considering the phoneme environment and the backward operation part 7 develops a hypothesis by using a viterbi search while referring to a tree structure dictionary of language models 10 considering the phoneme environment. Then the order of development is determined on a best.first basis by using A* algorithm making good use of the sum of the score of the forward operation result and the score of the operation result of the backward viterbi search made in phoneme units, and outputted as a word candidate for a recognition result in the order of received hypotheses and once a specific number of word candidates are found, the backward operation is ended.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a large vocabulary speech recognition apparatus, and in particular, to reduce the processing amount when performing speech recognition using an HMM (Hidden Markov Model) for each subword. The present invention relates to a large vocabulary speech recognition device.

[0002]

2. Description of the Related Art FIG. 6 is a block diagram showing a conventional speech recognition apparatus. In FIG. 6, a speech waveform is provided to a sound analysis unit 1 and is converted into a sound parameter which is a feature of the speech waveform using a linear prediction analysis or the like. These acoustic parameters are given to a parser 2 which is a parsing algorithm.

The parser 2 uses two models, an acoustic model 3 and a language model 4, to analyze acoustic parameters. The acoustic model 3 is for determining what parameters each phoneme is to be, and the language model 4 is for determining vocabulary and grammatical information for determining in what order the phonemes are arranged into meaningful sentences and words. It is for doing.
The parser 3 combines the two models to find a word or sentence that best matches the input.

In the acoustic model 3, the sub-word model HM
M is generally used, and in particular, a phoneme environment-dependent continuous distribution HMM is often used because it can accurately represent an unspecified speaker. In Japanese, for example, HMNet (Hi) described in JP-A-07-175494 is used.
dden Markov Netwook) has a good recognition rate. Here, a subword is a unit of expression that can express language speech with high accuracy and efficiency, and includes phonemes and syllables.

[0005] A speech recognition system using an HMM has a large processing amount. Looking at this by element, generally, the likelihood calculation and search processing (parser section) of the HMM are two major factors.
The likelihood calculation of the HMM is positioned as a pre-process of the search in the parser 2. In online speech recognition, real-time operation is desirable, and the large amount of processing not only directly affects the price but also imposes a load on other tasks, so reducing the processing amount is an important issue. In recent years, speech recognition using a subword HMM has become common, and research on speeding up the speech recognition has been increasing.

[0006] In continuous speech recognition, a sequence permitted by a given grammar (language model) is collated with input speech,
The phoneme sequence having the highest matching score is set as the recognition result. However, if all the phoneme sequences permitted by the grammar are collated with the input speech, a large amount of calculation is required. Reducing the number of times of collation as much as possible and performing only necessary collation is the key to the high-speed search process. As one method for that purpose, a method of searching for an accurate N-best candidate at high speed using the A ^* algorithm has been proposed.

FIG. 7 is a diagram for explaining the A ^* algorithm. In FIG. 7, when an arbitrary point on the graph is n, the estimated value of the optimal road cost from the starting point S to n is g ^* (n), and the optimal road cost from n to the target point is Let the estimated value be h ^* (n). If there is no road, g ^* (n) or h ^* (n) will be infinite. An estimate f ^* (n) of the cost of the optimal road through n is given by:

F ^* (n) = g ^* (n) + h ^* (n) (1) If the above equation is used as an evaluation function and the estimated cost h ^* is a lower bound of the true cost h, (h ^* ( n) ≦ h ^* (n))
Is called the A ^* algorithm.

[0009] In the A ^* algorithm shown in Figure 7 h ^*
(N) is shown beside the nodes, and f ^* for each node is shown in parentheses. The list changes as follows. (S (7) →
(A (8) B (9)) → (D (8) B (9) C (1
0)) → (B (9) C (10) H (10) I (10))
→ (D (7) C (10) E (10) H (10) I (1
0)) → (H (9) I (9) C (10) E (10)) →
(I (9) G1 (10) C (10) E (10) L (1
1)) → (G2 (9) G1 (10) C (10) E (1
0) L (11)). Next, G2 is taken out of the open and ends. The solution is S → B → D → I → G2.

On the other hand, a bundle search has been proposed as a method of sharing / approximate the collation calculation. In this method, only one matching calculation is required for each word, so that a high-speed search can be performed even with a complicated grammar. However, the amount of calculation depends on the number of words.

In the conventional discrete word recognition method, a score is calculated by performing a Viterbi search for each word included in the recognized vocabulary for the input speech. Therefore, the number of processes in the parser unit is proportional to the number of words. When the number of vocabularies is very large, the same applies to discrete word recognition as in the case of continuous speech recognition. Best-firsts for word-by-word matching
Although a method using t-type Viterbi search has been proposed, the reduction of the search space has been achieved, but the actual amount of processing has not been significantly reduced due to the large amount of calculation of the heuristic function. In addition, it can be said that creation of a heuristic function is an effective method for a discrete HMM.

For an unspecified speaker, it is necessary to use a mixed continuous HMM of a more accurate acoustic model. It is difficult to design a heuristic function for this mixed continuous HMM. Methods using preselection have long been proposed. However, due to the unavoidable drawback of preliminary selection mistakes, they are not often adopted recently. There is also a report that the amount of calculation could be reduced to 1/24 with a slight decrease in the recognition rate. However, this experiment is specific speaker recognition. When an unspecified speaker is targeted, errors in preliminary selection increase, and it is expected that the recognition rate will decrease.

"Large Vocabulary W" published by Chen et al.
ord Recognition Based on Tree-trellis Search, ”Pr
oc. ICASSP-94, pp.II-137-II-140 (1994)
We propose a method to speed up large vocabulary word recognition in Chinese using Soong's continuous speech recognition method. In the forward search, a syllable-free network is used. In the backward search, the A ^* algorithm is used with reference to a word dictionary converted into a syllable tree, and high-speed vocabulary recognition can be performed. Chinese has 412 syllables, so subsyllables (subs
yllable) A syllable is efficiently constructed by sharing the state of the HMM.

The phoneme environment considers only syllables, and only the connection conditions between vowels in syllables are environment-dependent. That is, syllables are not considered (= environmentally independent), the syllables are created as arcs in the tree structure dictionary, and syllables are directly connected. In this paper, the phoneme environment independent type (Table 4)
Recognition experiments of both the syllable phoneme environment-dependent type (Table 5) are being attempted. Of course, a higher recognition rate is obtained by using the syllable phoneme environment dependent type. However, this experiment is also specific speaker recognition. In the case of an unspecified speaker, degradation in recognition performance is expected because the intersyllable environment is not considered.

[0015]

In the conventional discrete word recognition method, a score is calculated by performing a Viterbi search for each word included in a recognized vocabulary for an input speech. Therefore, the processing amount of the parser 2 is proportional to the number of words. When the number of vocabularies is very large, the same applies to discrete word recognition as in the case of continuous speech recognition. When applied to information retrieval and the like, the number of vocabulary words can be several thousand words or more. For a system that takes 0.1 second for 100 words,
It takes 1 second for 1,000 words and 10 seconds for 10,000 words. This method is not suitable for online speech recognition where real-time operation is desired.

Since the Chen discrete word recognition method described above is intended for a specific speaker, good recognition performance was obtained because the acoustic model was originally highly accurate without considering the intersyllable environment. However, when this method is applied to an unspecified speaker, there is a problem that the A ^* algorithm does not operate well. This is because the accuracy of the acoustic model is low, the accuracy of the forward search is low, and the probability of failure in the backward search is high.

A similar experiment was attempted in Japanese, "Comparison of full search method, beam search method, and A ^* search method in isolated word recognition," Proc. Of the Acoustical Society of Japan, 2-5-10, pp.77 -
78 (1996.3). In this document,
An arbitrary syllable chain is used for estimating the heuristic function x ^* (t) for A ^* search, and the phoneme environment is not taken into account. However, good results have not been obtained compared with the conventional method.

The preselection scheme has not been used much recently. This is because the preliminary selection unit performs rough recognition / classification by a matching method with a small processing amount. Therefore, an error (= preliminary selection error) often occurs. This is because there is an unavoidable disadvantage that this error cannot be recovered later and causes a reduction in the recognition rate.

Therefore, a main object of the present invention is to provide a feature that the processing amount is hardly influenced by the number of vocabularies by adopting a search method in which the search space is efficiently narrowed in the parser. Another object of the present invention is to provide a large vocabulary speech recognition device capable of real-time operation with an inexpensive hardware configuration.

[0020]

The invention according to claim 1 is
In a speech recognition apparatus using a phoneme environment-dependent phoneme hidden Markov model, an input means for inputting speech, a feature vector extraction means for analyzing the input speech for each short-time frame, and extracting a feature vector, A dictionary creating means for expressing a recognition vocabulary obtained by adding a silent model before and after the beginning of the word based on the extracted feature vector as a phoneme environment-dependent phoneme sequence and converting the phonemes into a tree-structured dictionary having arcs as phonemes; , A parser means including a forward operation unit and a backward operation unit, the forward operation unit performs a Viterbi search driven under a constraint condition of a phoneme unit in consideration of a phoneme environment, and the backward operation unit includes a tree in consideration of a phoneme environment. The hypothesis is developed using Viterbi search while referring to the structure dictionary, and the result of the forward Viterbi search is calculated based on the score of the forward calculation result and phoneme unit. The order in which developed using the A ^* algorithm using the sum of the scores determined to best-first, and outputs it as a word candidate of the recognition result in order of received hypotheses, word candidates a predetermined number is Motomema In this case, the backward calculation is terminated.

According to a second aspect of the present invention, the forward operation unit of the first aspect performs a Viterbi search driven under a constraint condition of a state unit of a hidden Markov model in consideration of a phoneme environment.

According to the third aspect of the present invention, the backward operation unit of the first aspect limits the collation range of the Viterbi search for each phoneme by a predetermined method based on a predetermined duration for each phoneme.

According to the fourth aspect of the present invention, the backward operation unit of the first aspect executes the Viterbi search in the development of the hypothesis for each phoneme using the A ^* algorithm.

[0024]

FIG. 1 is a block diagram showing an embodiment of the present invention. In FIG. 1, sound input from a microphone (not shown) is converted into a digital signal by an A / D converter, and input to the acoustic analysis unit 1. The acoustic analysis unit 1 analyzes the input speech for each frame and extracts acoustic parameters. The acoustic parameters include, for example, LPC cepstrum, differential LPC cepstrum, and differential power.

In this embodiment, HMNet will be described as a phoneme environment-dependent HMM of an acoustic model. Each phoneme is represented as a triphone to make it phonemic environment dependent, and each state may be shared with other triad phoneme states. Therefore, a phoneme list connected before and after for each phoneme, a phoneme list connected before and after for each state,
A state list connected before and after each state is described for each state to form one network. Table 1 shows HMn
The example of various information regarding the state of et was shown. A forward search operation and a backward search operation are realized using these pieces of connection information.

[0026]

[Table 1]

The parser 5 in FIG. 1 is composed of two operation units, a forward operation unit 6 and a backward operation unit 7. The forward operation unit 6 first calculates the likelihood for each state of HMNet for each frame. This calculation result is referred to by the forward calculation unit 6 and the backward calculation unit 7 as a likelihood table. Next, a Viterbi search is performed, which is driven under a constraint condition in units of three phonemes in consideration of a phoneme environment.

FIG. 2 is a diagram showing an example of state / phoneme connection information. In FIG. 2, one phoneme is composed of three to four states. Although the total number of triads in Japanese is three hundred and several hundred, when HMNet used in the embodiment of the present invention is used, the state is shared, so that triads in different state series (three different phonemes) There are several hundred types. Therefore, in the embodiment of the present invention, the states are arranged for each of the three sets of phonemes differently as shown in FIG. Accordingly, among the three sets of phonemes, those having the same HMM state series are omitted from the calculation.

In FIG. 2A, the time is changed from t-1 to t.
When transitioning to, of the allowed transitions,
The state related to the state a is indicated by a solid arrow. Heuristic function h of a certain phoneme p at a certain time t
^* _p (t) is calculated as a cumulative score for all different phonemes for each frame as in the following equation (2). In the state j at the time t in FIG. 2A, there are paths from the state i and the state j on the immediately preceding frame t-1, and these paths are traced from the t-1 frame to the state of the t frame. Among the cumulative scores that reach j, the largest one is the cumulative score of state j at time t.

[0030]

(Equation 1)

Here, j is the state number of the final state of the phoneme p, and is defined as h ^* _p (t) ≡h ^* _jp (t) for simplicity. b _j (t) is the symbol output probability at time t in state j and is stored in the likelihood table. a _ij is a state transition probability from the state i to the state j.
C _p (j) means a set of states belonging to the phoneme p among the states that can transition to the state j, and is obtained by referring to the state / phoneme connection information of the sound / language model 8 shown in FIG.
v _ip (t−1) is a cumulative score of the state i at time t−1, and is recursively obtained by Viterbi search.

FIG. 2B shows the case where j is the initial state number of the phoneme p. In this case, as shown in Expression (3), the largest one of the self-loop and the phoneme group that can be connected to the phoneme p is selected.

[0033]

(Equation 2)

Here, I (p) means a set of phonemes that can be connected to phoneme p and their final state numbers, and is obtained by referring to phoneme connection information shown in Table 2 described later.

It is also effective in the forward operation unit 7 to perform a Viterbi search in which the driving is performed under the constraint condition of the state unit of the HMM in consideration of the phoneme environment. In this case, h
^* The accuracy of (t) is slightly inferior to the above method. However, the degree to which calculations need not be performed due to state sharing increases. In the end, since it is sufficient to calculate only the number of states, there is an advantage that the amount of calculation required for the forward search is reduced to 1/10 or less. The forward calculation method may be determined according to the application.
The heuristic function h ^* _p (t) of a certain phoneme p at a certain time t is calculated for every state for each frame as shown in Expression (4), and is calculated using the state number j of the final state of the phoneme p. ^* _j (t) is typically represented.

[0036]

(Equation 3)

Here, S (j) means a set of all the states that can transition to the state j, and is obtained by referring to the connection information for the state. v _i (t−1) is the cumulative score of state i at time t−1, and when j is the state number of the initial state of the phoneme p, the same as in equation (2).

Next, Table 2 shows a table of phoneme connection information in the set I (p) of connection information between phonemes and initial state numbers.

[0039]

[Table 2]

This table is referred to for the initial state of the phoneme p. Considering linguistic restrictions, restrictions are placed on the connection between specific phonemes. For example, in Japanese, it is considered that consonants are not connected to each other, so that a route from “m” to “h” is not provided. Phonetic notation follows the Hepburn Roman spelling. However,
“Q” indicates a prompting sound, “N” indicates a repellent sound, “y” indicates a relentless sound, and “−” indicates no sound. In Table 2, it means that the final state of the phoneme HMM on the left can be connected to the initial state of the phoneme HMM on the right.

The backward operation unit 7 refers to the tree structure dictionary of the language model 10 using phonemes as arcs. This tree structure dictionary is created in advance from the recognized vocabulary list. Parser 5
Is performed before the word, the speech section of each word is extracted. Therefore, for a section determined to be a voice section, a certain margin, that is, a surrounding environment sound section is often set before the beginning of the word and after the end of the word. To accommodate this margin, a silence model of HMM is added before and after each recognized vocabulary. The silence model is a model that has been trained on ambient noise that contains no voice, and does not always indicate true silence with a waveform of zero.

FIG. 3 is a diagram showing an example of a part of a tree structure dictionary, and FIG. 4 is a diagram showing a tree structure dictionary when a phoneme environment is not considered for comparison. 3 and 4,
The numbers indicate the node numbers of the dictionary, and the alphabets indicate the phonemes of the arc. The phonemes are developed under the connection information restriction between phonemes as exemplified in Table 2 described above, and the surrounding phoneme environments are considered. Therefore, branching is increasing. While four arcs extend from the node 328, as shown in FIG. 4, if the phoneme environment is not taken into account, two arcs occur as in the node 76. Since the direction of operation is opposite to time, a branch extends from the end to the beginning.
In order to make the backward operation reference faster, each arc is assigned a state number of HMNet in advance.

The hypotheses are developed along the tree structure dictionary shown in FIG. 3, and the development order is determined as best-first using the A ^* algorithm. That is, the node of the partial hypothesis having the highest evaluation value f ^* _p (t) in the following equation (5) is expanded and the process proceeds. Here, p represents a phoneme of the tip arc of the hypothesis, and t represents a frame number in a collation range R (p, t ₀ ) determined by a method described later.

The evaluation value f ^* (t) is the forward operation result score h ^* (t), the as shown in equation (6) t∈R (p,
t ₀ ) is represented by the sum of the scores g _p (t) of the calculation results of the backward Viterbi search executed for each phoneme. Since this sum is in accordance with the tree structure dictionary, the before and after phoneme environment information given to h ^* (t) and g _p (t) is reflected, and the phoneme environment is considered. For simplicity of processing, the connection point is limited to t'1 which gives the maximum value of f ^* (t). In other words, the sum of hypotheses is reduced by representing the same hypothesis by one phoneme sequence only at the different connection points. Viterbi search of the starting point is put with this t ₀ makes this t 'when you then expand the arc to connect to this hypothesis.

[0045]

(Equation 4)

Here, i is the state number of the initial state of the phoneme p, and is defined as g _p (t) ≡g _ip (t). C ^*
_p (i) is a set in which the direction of the state connection of C _p (j) is reversed. As described in the best-first hypothesis growing algorithm described later, there are a single hypothesis and a group hypothesis. For the single hypothesis, p is the only phoneme of the tip arc. For the group hypothesis, p indicates the phoneme of the selected arc that gives the highest score in the group. t ₀ indicates the previous connection point t ′ of the hypothesis.

The matching range R (p, t ₀ ) of the Viterbi search for each phoneme p in the backward calculation is the end point from the normal start point t ₀ (= the previous connection point) of the hypothesis, that is, the beginning of the input voice. Up to t = 1. Therefore, the matching range becomes very wide near the end of the word, which leads to an increase in the amount of calculation. Phonemes have their own duration.
In general, vowels, sound repellents, and consonants tend to be long, and consonants tend to be short. Then, using the speech data labeled in phoneme units, the average duration length μ _p and the variance σ ²
_{Find p} . If there is a large amount of voice data, the average duration and variance may be calculated for each triad of phonemes for accuracy. As a usage, a section of “average value μ _p ± α × standard deviation σ _p ” from the previous connection point t ₀ is targeted. For example, in this embodiment, as shown in the following equation (7), “average value μ _p + 3 × standard deviation σ _p ” from t _0.
The section that has been traced back is set as the collation range R (p, t ₀ )

[0048]

(Equation 5)

It has been confirmed by experiments that this collation range restriction not only reduces the amount of calculation but also prevents unnecessary time-axis matching often occurring in Viterbi search, thereby contributing to an improvement in the recognition rate. Have been.

FIG. 5 is a flowchart showing an algorithm for growing a hypothesis at best-first. This flowchart conforms to Soong described above. In FIG. 6, the total score is the evaluation value f ^* _p (t) of the equation (5), the root node means a hypothesis that has not yet been developed at all, and the single pass is a hypothesis. Indicates one hypothesis and the group path indicates a hypothesis having a plurality of hypotheses, and N indicates the number N of N-best candidates.

The development of the hypothesis proceeds by expanding the top entry of the stack (= the best partial hypothesis) by one arc and dividing it into two hypotheses (the best single path and the remaining group). Since the development target is always the best partial hypothesis, the development order is best-first.

More specifically, with reference to FIG.
A root node is placed on the stack, and initialization is performed.
Next, the top entry of the stack is taken out, and it is determined whether or not the best partial hypothesis is a single pass and not a group pass. If it is not a single pass, the best partial hypothesis is split into two hypotheses (the best single pass and the remaining groups), and a total score is calculated for these two hypotheses,
One hypothesis is returned to the stack and sorted based on the total score. In the case of a single pass, it is determined whether or not the best partial hypothesis has reached the terminal node. If the best partial hypothesis has not reached the terminal node, the group hypothesis is divided into two hypotheses, and the total score of these two hypotheses is calculated. Calculate and put these two hypotheses back on the stack and sort based on the total score. When the best partial hypothesis reaches the terminal node, the hypothesis is output and the accepted number counter is incremented. If the received number counter is not equal to N, the top entry of the stack is taken again. If the received counter is equal to N, the process is terminated.

As described above, by sequentially developing hypotheses in the best-first manner, N-best word candidates as recognition results are obtained in the order of the received hypotheses. In other words, since the candidates with the highest scores are accepted in order, the first, second, third
Word candidates are output in the order of order,. When a predetermined number of word candidates (for example, up to the tenth place for ten words) are obtained, the backward calculation is terminated, and the N-be
Only the arc related to the st word candidate is executing the Viterbi search by backward calculation.

The accuracy of the heuristic function h ^* (t) greatly affects the search efficiency of the backward search (= the number of hypothesis developments). If h ^* (t) is equal to the true value h (t), the development proceeds ideally, and there is no need to develop useless hypotheses. At this time, the processing amount does not depend on the number of recognized vocabularies, that is, the size of the tree structure dictionary. h
^* (T) is the acceptability of the A ^* algorithm: h ^* (t)
Although the relationship of ≧ h (t) holds, h ^* (t) = h (t) does not hold because a weak grammar is used.

Therefore, in practice, there is some useless hypothesis development, and a search is also made for a peripheral arc of a hypothesis that is close to the correct answer, so that the processing amount slightly depends on the number of recognized words. As a result, the processing amount is hardly influenced by the number of words. Compared with the conventional method of executing the Viterbi search one word at a time, the search space can be dramatically reduced as the vocabulary increases. Therefore, this embodiment can be said to be a recognition method suitable for a large vocabulary. It has been confirmed by experiments that the processing amount of the parser 5 can be reduced to 1/40 when 20,3000 words are used as the recognition vocabulary.

The hypotheses that are being developed are stored on the stack. The stack must be rearranged (sorted) for each expansion operation. The stack size may theoretically be the same as the number N of N-best candidates under an ideal environment.
However, since it is actually affected by the number of recognized vocabulary words and the performance of the acoustic model, it is necessary to set a value with a margin.
In an actual environment experiment, a size of several hundreds is desired. For example, in this embodiment, the vocabulary is 1,000 when the vocabulary is 20,000 words, and 500 when the vocabulary is 5,000 words. Therefore, stack sorting causes an increase in the processing amount. When the hypothesis is returned to the stack in order to speed up the processing, this becomes a factor for increasing the binary tree search processing amount. In order to speed up the processing, when returning a hypothesis to the stack, a place to insert is determined using a binary tree search. This eliminates the need to sort all data in the stack. Stack replacement is performed by pointer operation, and actual data on the stack is not moved. However, depending on the processing system, it may be more effective to make the stack a list structure than to operate the pointer.

Note that it is also possible to use the A ^* algorithm instead of the Viterbi search for each phoneme in the backward calculation. In this case, the heuristic function that has already been obtained by the forward operation can be used.
However, since the A ^* algorithm is activated many times and has an overhead such as a stack operation, whether or not the processing speed is improved depends on conditions of a processing system to be mounted such as a memory access speed.

In this embodiment, phonemes are used as subword units, but syllables can be used. There are about 110 types of Japanese syllables, and the number of syllables is more than 10,000 in consideration of the phoneme environment. Therefore, the processing amount of the forward calculation is large, but the accuracy of the forward search is improved. Is performed more efficiently.

In the above embodiment, words are picked up as recognition targets. However, the vocabulary of the dictionary is not limited to words, and if a phrase is regarded as one word and a tree-structured dictionary is created, the phrase Recognition is also feasible. In Japanese, since the ending expression can be shared by a tree structure in expressions such as particles, it is possible to search efficiently.

[0060]

As described above, according to the present invention, the recognized vocabulary is represented by a phoneme environment-dependent phoneme sequence, and these phonemes are converted into a tree-structured dictionary having arcs. Perform the Viterbi search driven under the constraint condition of the phoneme unit considered, develop the hypothesis using the Viterbi search while referring to the tree structure dictionary considering the phoneme environment in the backward operation unit, and calculate the score of the forward operation result and The best order to develop using the A ^* algorithm that uses the sum of the scores of the results of backward Viterbi search performed on a phoneme-by-phoneme basis
−first, and output them as word candidates of the recognition result in the order of the accepted hypotheses. When a predetermined number of word candidates are obtained, the backward operation is terminated and the words of the recognition candidates are output. By taking advantage of the feature that the processing amount is not proportional to the number of vocabularies, it is possible to realize a speech recognition device capable of real-time operation with an inexpensive hardware configuration even for very large vocabularies. For example, when 20,000 words are used as the recognition vocabulary, according to the present invention, of the two major elements of the processing amount of speech recognition, the likelihood calculation of the HMM and the parser, the latter is about 1/4.
It can be reduced to zero.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of a restriction on a connection between states of an HMM.

FIG. 3 is a diagram showing an example of a part of a tree-structured dictionary considering a phoneme environment.

FIG. 4 is a diagram showing an example of a part of a tree-structured dictionary without considering a phoneme environment.

FIG. 5 is a flowchart illustrating an algorithm for growing a hypothesis in best-first.

FIG. 6 is a block diagram showing a configuration of a conventional general word speech recognition device.

FIG. 7 is a diagram for explaining an A ^* algorithm.

[Explanation of symbols]

 Reference Signs List 1 sound analysis unit 5 parser 6 forward operation unit 7 backward operation unit 8 sound / language model 9 sound model 10 language model

Claims (4)

[Claims] 1. A speech recognition apparatus using a phoneme environment-dependent phoneme hidden Markov model, comprising: input means for inputting speech; analyzing speech input from the input means for each short-time frame; Based on the feature vector extracted by the feature vector extraction means, a recognition vocabulary to which a silence model is added before and after the beginning of a word is represented by a phoneme environment-dependent phoneme sequence,
Dictionary creation means for converting the phonemes into a tree-structured dictionary having arcs, and parser means including a forward operation unit and a backward operation unit, wherein the forward operation unit is provided with phoneme unit constraints in consideration of a phoneme environment. The Viterbi search driven below was performed, and the backward operation unit developed a hypothesis using Viterbi search while referring to a tree structure dictionary in consideration of the phoneme environment, and executed the forward operation result score and phoneme unit. The order of development using the A ^* algorithm that uses the sum of the scores of the operation results of backward Viterbi search is best-first.
A large vocabulary speech recognition device, characterized in that the hypotheses are output in the order of the accepted hypotheses as word candidates as recognition results, and when a predetermined number of word candidates are obtained, the backward calculation is terminated. 2. The method according to claim 1, wherein the forward calculation unit performs a Viterbi search driven under a constraint condition of a state unit of a hidden Markov model in consideration of a phoneme environment.
Large vocabulary speech recognition device. 3. The method according to claim 1, wherein the backward calculation unit limits a collation range of the Viterbi search for each phoneme in a predetermined manner based on a predetermined duration for each phoneme. Large vocabulary speech recognition device. 4. The large vocabulary speech recognition apparatus according to claim 1, wherein said backward operation unit executes a Viterbi search in developing a hypothesis in phoneme units using an A ^* algorithm.