JP2001092495A - Continuous speech recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To surely obtain a solution and also shorten a processing time by performing a 1st path retrieval using a coarse model to create a word lattice, and performing a 2nd path retrieval from the end on the word lattice using a high accurate model. SOLUTION: By the 2nd path retrieval, a most delayed one among retrievals of all assumptions, namely, the most delayed one N7 in head word boundary time of the assumption, is selected; the assumption is extended and developed by one word to obtain a score; pruning is performed; the retrieval selects a most delayed one again; and the similar operations are repeated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous speech recognition method for finding a hypothesis closest to an input speech through a plurality of search stages from hypotheses of many word strings that can be generated by a prescribed grammar or connection relation. About.

[0002]

2. Description of the Related Art First, an example of a conventional continuous speech recognition method will be described with reference to FIG. In the figure, an input speech 11 is converted into a vector series of feature parameter vector data in an analysis processing unit 12,
In step 3, the acoustic model 15 corresponding to the hypothesis (hereinafter simply referred to as a hypothesis) of the word string permitted by the grammar / language model 16 is compared with the vector data time series of the feature parameter. The score, which is the evaluation value of the matching result of this hypothesis, is composed of an acoustic score indicating the acoustic closeness between the input speech and the hypothesis and a language score indicating the probability of the existence of the hypothesis, and the hypothesis having the highest score is the recognition result. It is output as 14.

[0003] Cepstrum analysis is often used as signal processing in the analysis processing unit 12, and the characteristic parameter is MFCC (Mel Frequency Cepstral C).
oefficient), ΔMFCC, logarithmic power, and the like. As the acoustic model 15, a Hidden Markov Model (Hidden Markov Model) modeled on the basis of probability and statistical theory is used.
, Hereinafter referred to as HMM). Normally, an HMM is created for each phoneme (phoneme model), but at present, when creating an HMM for a certain phoneme, the phonemes connected before and after the phoneme are also taken into account (considering the phoneme environment).
HMM called MM has become mainstream. This HMM
Details are disclosed in, for example, a document (edited by the Institute of Electronics, Information and Communication Engineers, edited by Seiichi Nakagawa, “Speech Recognition by Stochastic Model”).

[0004] The grammar / language model 16 defines a connection between words for defining a sentence to be recognized, and uses a word network with words as branches, a language probability model, or the like. In the case of continuous speech recognition, the grammar is shown in FIG.
Many words take the form of a word network in which any word can be connected to any word. By adopting such a format, it is possible to generate a hypothesis of an arbitrary word string within the range of words registered in the word network. As the language probability model, the existence probability of a single word and the probability of two or more words being linked are used. The model representing the existence probability of a single word is the word 1-gram, the model representing the two-chain probability and the three-chain probability of the word is the word 2-gr, respectively.
am, called the word 3-gram. By using the probability model of this language, generation of a hypothesis that cannot exist as a language (here, Japanese) can be suppressed. The details of the probability model of this language are disclosed in, for example, Seiichi Nakagawa, "Speech Recognition by Probability Model", edited by the Institute of Electronics, Information and Communication Engineers.

[0005] The search processing unit 13 collates an acoustic model corresponding to a hypothesis on a word network indicating a connection relationship between words specified by a grammar with a time series of feature parameter vector data, and indicates an acoustic likelihood. At the same time as obtaining the score, the language score is obtained from the language model corresponding to the hypothesis, and the score of the hypothesis consisting of the acoustic score and the language score is obtained for each hypothesis from the beginning to the end of the input continuous speech, and the hypothesis of the largest score, That is, the hypothesis closest to the input speech is output as the recognition result. In continuous speech recognition, the number of hypotheses that can be generated by grammar is enormous, and in order to obtain a recognition result with high speed and high accuracy, a hypothesis search is performed in multiple stages, and a multipath search that narrows down candidate hypotheses in stages is performed. Laws are often enforced. For details of the multipath search, see, for example, R. Schwartz, L. Nguyen, and John M.
akhoul: “Multiple-pass SearchStrategies”, in Au
tomatic Speech and Speaker Recognition Advanced To
pics, pp.429-456, Kluwer Academic Publishers (199
6). Etc.

Here, the most general multi-path search for narrowing hypotheses in two stages will be described with reference to FIG. In the first-stage search (first-pass search) 21, a coarse language model in the grammar / language model 16, for example, the word 2-gram, is selected from a huge number of hypotheses that can be generated by a word network as shown in FIG. Or a coarse acoustic model in the acoustic model 15, for example, considering only the phonemic environment in a word tr
Using a model with low computational cost such as iphone HMM,
Narrow candidate hypotheses close to the input voice at high speed.

The first path search 21 often employs a method called a time-synchronous beam search. In the time-synchronous beam search, the matching between the input speech and the hypothesis can be performed simultaneously for all the hypotheses that can be generated by the word network in FIG. Is significantly increased with the passage of time, and it is difficult to complete this processing in a practical processing time. Therefore, by terminating (pruning) the search for a hypothesis that is unlikely to be a recognition result during the search, the search is intended to be completed in a realistic processing time. The criteria for pruning the hypotheses in the time-synchronous beam search are, among all the hypotheses, the method of leaving m hypotheses in order from the largest score to the largest score, cutting off the other hypotheses, and the scores of all the hypotheses. There is a method in which a score obtained by subtracting a constant value θ from the largest score among the scores is used as a threshold, only hypotheses having a score equal to or higher than the threshold are left, and hypotheses having a score lower than the threshold are pruned. Here, the parameters m and θ that determine the criteria for pruning the hypothesis are called beam widths. In the time-synchronous beam search, hypotheses that are estimated to have no possibility from the size of each hypothesis at the same time are pruned, so there is little possibility that a hypothesis that provides a correct solution will be pruned. However, in time-synchronized beam search, hypotheses are pruned during the search, so that the hypothesis with the highest score is not always obtained as a recognition result. It is. For details of time-synchronous beam search, see, for example, R. Haeb-Umbach and H. Ney: “Improvem
ents in beam search for 10000-word continuous-spee
ch recognition ”, IEEE Trans. Speech and Audio Pro
cessing, Vol.2, No.2, pp.353-356 (1994). Etc.

The result of the first path search is obtained as an intermediate expression such as a trellis form or a word lattice. Here, it is assumed that a word network which compactly expresses a word connection called a word lattice as shown in FIG. I do. The word lattice stores a word boundary time as a result of the first pass search and a score of the hypothesis up to that time at that time. For more information on the word lattice, for example,
S. Ortmanns and H. Ney: “A word graph algorithm for
large vocaburary continuous speech recognition
”, Computer Speech and Language, Vol. 11, No. 1, p
p.43-72 (1997). Etc.

As shown in FIG. 8, in a second stage search (second path search) 22, a high-precision acoustic model and a grammar in the acoustic model 15 are obtained on the word lattice 23 obtained as a result of the first path search 21. The recalculation of the hypothesis score using the high-precision language model in the / language model 16 is performed for each analysis frame, and the final recognition result 14 is obtained. Methods often used as the second path search 22 include N-best rescoring and A ^* search.

[0010] N-best rescoring is an N-bes ordering by scores from a search using a coarse model.
The scores of a plurality of (N high-score) sentence candidates called st sentence candidates are replaced by scores obtained by searching using a high-precision model, and the order of sentence candidates is changed in descending order. When N-best rescoring is used for the second pass search 22, first the first
Based on the score of the path search, N sentence candidates (N-best sentence candidates) are created from the word lattice 23 in descending order of the score, and the score based on the coarse language model such as the word 2-gram is converted to the word 3-gram or the like. The score is recalculated by replacing the score with a higher accuracy language model score, and the order of the sentence candidates is changed in descending order of the recalculated score.
The N-best rescoring is simple in implementation, and can reliably obtain a recognition result. Details of N-best rescoring are described, for example, in L. Nguyen, R. Schwartz, Y.
Zhao, and G. Zavaliagkos: “Is N-best dead?”, Pr
oc.DARPA Speech and Natural Language Workshop, pp.
411-414 (1994). Etc.

In the A ^* search, the hypothesis n having the highest score defined by the following equation (1) is expanded preferentially (best
-first search). f _n (t) = g _n (t) + h _n (t) (1) Here, t is the time (frame number), and g _n (t) is the score of the section that has already been searched, that is, in FIG. Word boundary time (also called node) N0-N1-N2-N3
−N4−N5, which is the score of the hypothesis linking h _n (t)
Is an estimated score (heuristic) of an unsearched section from the word boundary time N5 to the beginning. That is,
f _n (t) is an estimated score for all sections of hypothesis n, and using f _n (t) as the score of hypothesis n is
This means that scores of all hypotheses are obtained for all sections from the start to the end. This also makes it possible to compare hypotheses with different degrees of search progress (different time lengths). In order to obtain the solution with the highest score (optimal solution) in the A ^* search, the value of _hn (t) must be greater than its true value (assumed) (A
^* Eligibility) is known. Further, the more efficient the search is, the closer h _n (t) is to its true value. When the A ^* search is used for the second path search 22, as shown in FIG. 10, a hypothesis development is performed on the word lattice in the word unit from the end of the sentence in the opposite direction to the first path search 21. At this time, g
_n (t) is a language model and an acoustic model, each of which has higher accuracy than the language model and the acoustic model used in the first pass search, such as a word 3-gram, and a triphone that considers each phoneme environment within and between words. It is obtained by recalculating using HMM.

In the second path search score, the first path search score stored in the word lattice 23 can be used for h _n (t). In FIG. 10, N0-N
1-N2-N3-N6, N0-N1-N2-N3-N4
-N5, N0-N1-N2-N7, N0-N1-N8,
Although there are N0-N11-N9, the N0-N11-N10 total of six hypotheses, if from this f _n (t) is maximum one (e.g. now, N0-N1-N2-N3 -N4-N5 )
To expand this. Details of the A ^* search are, for example,
It is disclosed in Corona Publishing, Nils.J.Nilsson, Shuhei Goda, Izu Masuda, and "Artificial Intelligence: A System of Problem Solving".

[0013]

SUMMARY OF THE INVENTION By the way, N-bes
For the t rescoring, the word lattice is converted to N-best
When creating sentence candidates, a large number of similar candidates that differ by only one word appear, so in order to obtain sufficient recognition accuracy, it is necessary to perform rescoring on relatively many sentence candidates. Although recalculation of the acoustic score using a more accurate acoustic model is possible, it is not efficient. On the other hand, in the A ^* search, the result of the first
Although there is an advantage that it can be used as a heuristic in the path search, since the models used in the first path search and the second path search are different, h _n (t) close to the true value is not always obtained. Depending on the case, the efficiency of the search may be reduced. It is sufficient that h _n (t) is close to the true value and the best hypothesis expansion of f _n (t) can be performed successfully. However, if h _n (t) is far from the true value, the number of hypotheses increases extremely, Real-time recognition becomes difficult.

The present invention has been made in view of the above-mentioned problems in the N-best rescoring and the A ^* search, and utilizes the results of a path search using a coarse model such as the A ^* search. , Then perform efficient path search,
Another object of the present invention is to provide a continuous speech recognition method that can always obtain a solution, such as time-synchronized beam search and N-best rescoring.

[0015]

According to the present invention, a hypothesis whose search is most delayed in a search using a high-precision model using a word network (word lattice) obtained by a search using a coarse model is described. Repeated execution of preferential deployment. By doing so, the lengths of the hypotheses being developed are almost the same. Therefore, pruning is performed during hypothesis development, an efficient search becomes possible, and a solution is always obtained.

In a search using a high-precision model, a score is calculated by giving each word boundary time stored in the previously obtained word lattice a width of 5 ms or more or 1 frame or more. I do.

[0017]

Embodiments of the present invention will be described below. In this embodiment, for example, as shown in FIG. 8, a first pass search is performed on a vector data sequence of input feature parameters using a coarse acoustic model and a coarse language model, and a word lattice 23 is generated. Word lattice 2
3, a second pass search is performed using the high-accuracy acoustic model and the high-accuracy language model.

The feature of this embodiment lies in the second path search method. In the second path search, as in the conventional A ^* search, the hypothesis is developed in word units from the end of the sentence (end of the input speech) in the opposite direction to the first path search. At this time, in the present invention, the word-based hypothesis development is preferentially developed from the one with the slowest search (Shortest-first search). For example, as shown in FIG. 1, a hypothesis consisting of word boundary times (nodes) N0-N1-N3-N8, N0
A hypothesis consisting of -N1-N3-N9, N0-N1-N3-
Hypothesis consisting of N7, ..., N0-N1-N4-N5-N1
In the state where the seven hypotheses of No. 3 are developed, the hypotheses whose search is the slowest are the head nodes N8, N9, N7, N10, N11, N12, N13 of each hypothesis.
The node N7 at the time t whose time is the latest is selected. However, each time is indicated as the end time advances with respect to the start point of the input voice. The hypothesis is developed for the node N7 selected in this way. For example, suppose that nodes connected from the word lattice 23 (FIG. 8) to the start end with respect to the node N7 are N14, N15, and N16, and the estimated score (heuristic) of the unsearched section from the node N14 to the start end is h _n 1 (T) Similarly, the node N1
5, heuristic unsearched section extending from N16 to start each h _n 2 (t), assumed to be h _n 3 (t).

A temporary connection from node N0 to N14 via N7.
Theory score g_n1 (t) is calculated for each analysis frame.
This g_n1 (t) and h_nSum f with 1 (t)_n1
(T), that is, scores in all sections of the hypothesis are obtained.
Similarly, the score g of the hypothesis reaching the node N15_n2
(T) and the score f in all the sections_n2 (t)
Hypothesis score g leading to node N16_n3 (t) and that
Score f in all sections of _n3 (t) is obtained.

As described above, the hypothesis that extends the hypothesis by one word is developed from the leading node N7 of the hypothesis that is the latest, and the hypothesis that is the latest is selected and the hypothesis is determined by one for each extension of the word. Hypothesis development is performed to extend words. In this way, the hypotheses are developed while the temporal lengths of the respective hypotheses are almost the same as in the time-synchronized beam search. Therefore, pruning based on the score becomes possible,
In this embodiment, pruning is performed while developing a hypothesis. This pruning can use one or both of the two approaches. One of them is to calculate the hypothesis score g _ni (t) (for example, i = 1, 2, 3) for each analysis frame, which is obtained when extending the hypothesis. When the calculation for each analysis frame is completed, The score obtained by subtracting the constant value θ from the highest value of the scores g _n (t) of all the hypotheses at that time is set as a threshold, and the hypothesis of a score below the threshold is discontinued and removed therefrom.

For example, as shown in FIG. 2, the envelope of the highest value of the score g _n (t) obtained by calculation for each analysis frame is represented by a curve 31, and a curve 32 having a score smaller by θ than the curve 31. Then, those in which the score g _n (t) is below the curve 32 during the calculation of the hypothesis expansion are excluded.
Only hypotheses whose scores fall between curves 31 and 32 are left. FIG. 2 shows the case where the score calculation direction is set such that the score becomes smaller as the hypothesis is extended.

Another method of pruning is to extract m total interval scores f _n (t) of all hypotheses in descending order each time one hypothesis is extended and expanded for one word. And the smaller hypotheses are removed.
FIG. 3 shows the procedure for developing the above-described hypotheses. First, the one with the latest time is selected from the head node group N = {n1,..., Nx} of all the hypotheses (S1). The node group {ni1,..., Niy} developed from the node ni is extracted (S
2). One node nij in order from the extracted node
(J = 1,..., Y), the score g of the hypothesis
The calculation of _n (t) (nij) is started (S3). During the calculation of each g _n (t) (nij), the highest score is obtained from the calculation result for each analysis frame, and θ is subtracted therefrom. When the calculated score falls below the threshold value (S4), the calculation is stopped and the expansion to the node nij is stopped, that is, the hypothesis to expand to the node is pruned. To step S7 (S12).

During the calculation, the score does not fall below the threshold value.
When the core calculation ends (S5), the node nij starts.
Not at the end (S6), and at all of the extracted nodes nij
If the calculation is not completed (S7), step S
3 and start calculating the score for the next node nij.
Start. Calculate hypothetical scores for all nij and finish
Then (S7), the previously selected ni is read from the head node group N.
Erase and add all nij to head node group N (S
8). If the number of hypotheses in this state is m or less (S
9) Returning to step S1, again from the top node group
Also selects the node whose time is late and performs the same processing.
U. On the other hand, if the number of hypotheses is not less than m,
Interval score f_n(T) = g _n(T) + h_nLarge (t)
Take m pieces in order from the thing, leave only that hypothesis,
The hypothesis is removed (S10). With this removal, the removal
The first node of the removed hypothesis is also removed from the first node group N.
You. After this pruning, the process returns to step S1.

If nij is the starting point in step S6, the total section score f _n (t) = g _{n of the} hypothesis obtained at that time is obtained.
(T) (nij) is stored for the hypothesis, and the process proceeds to step S7 (S12). This nij is not added to the head node group N again (search for nij is completed). The above processing is performed until there are no more top nodes to be selected in step S1, and when there are no more top nodes to be selected, hypotheses with the largest or the largest number of stored hypothesis scores are output as recognition results.

Since different models are used in the first path search and the second path search, there is a possibility that the word boundaries may be shifted between the first path search and the second path search even for the same hypothesis. Therefore, in this embodiment, instead of using the word boundary time of the first path search stored in the word lattice as it is as the word boundary time of the second path search, the second path search is performed while allowing a shift of several frames before and after. Do.

That is, for example, as shown in FIG. 4A, the word boundary time stored in the word lattice is t1 between the words A and B, and t2 between the words B and C. At this time, FIG.
B, between word A and word B, not only t1 but also t1
1-Δ and t1 + Δ are also boundary times, and between the words B and C, not only t2 but also t2-Δ and t2 + Δ are boundary times. The calculation of the score at this time starts from time t2 + Δ, and the value Δg (t2 + Δ, t2) when time t2 is reached
Is further stored, and the value Δg (t2 + Δ, t2-Δ) when the calculation is continued and the time reaches t2-Δ is stored. The calculation is further continued and the value g (when the time t1 + Δ is reached) is stored. t2 +
Δ, t1 + Δ), and the calculation is further continued to proceed to t1
The value g (t2 + Δ, t1) at the time t is stored, the calculation is further continued, and the value g (t2 +
Δ, t1-Δ), and each score g when the hypothesis is extended from t2 + Δ, t2, t2-Δ to t1 + Δ, respectively.
(T) + Δ, t1 + Δ) and g (t2 + Δ, t1 + Δ) −
Δg (t2 + Δ, t2) and g (t2 + Δ, t1 + Δ)
−Δg (t2 + Δ, t2−Δ), the largest one among the three at time t1 + Δ is defined as t2 + Δ, t2, t2
Each score g when the hypothesis is extended from -Δ to t1
(T2 + Δ, t1) and g (t2 + Δ, t1) −Δg
(T2 + Δ, t2) and g (t2 + Δ, t1) −Δg
(T2 + Δ, t2-Δ), the maximum score is the score at time t1, and each score g (t2) when the hypothesis is extended from t2 + Δ, t2, t2-Δ to t1-Δ, respectively.
+ Δ, t1−Δ) and g (t2 + Δ, t1−Δ) −Δg
(T2 + Δ, t2) and g (t2 + Δ, t1-Δ) −Δ
The largest one of the three g (t2 + Δ, t2-Δ) is set as the score at time t1-Δ.

Note that Δ is set to one analysis frame or more to about 5 milliseconds or more. However, when the value of Δ is increased, the amount of calculation increases, so it is set to several frames to several tens of milliseconds or less. In the above, f _n
(T) = g _n (t) + h _n (t) was used.
_n (t) is given a weight α close to 1 and f _n (t) =
g _n (t) + αh _n (t) may be used as the whole section score to improve the accuracy. In order to obtain α, a score h is calculated for an appropriate word string using a coarse model used for the first pass search, and a score g is calculated for the word string using a high-precision model used for the second pass search. Then α
= G / h to calculate the weight α.

In the above description, the present invention is applied to the second path search, but can also be applied to a case where recognition is performed by a three-step search. In short, the present invention can be applied to the case where a path search is performed using a coarse model to create a word lattice, and a path search is performed on the word lattice using a highly accurate model. Subsequently, a comparative continuous speech recognition experiment using the above-described N-best rescoring and the search according to the present invention (hereinafter referred to as time asynchronous beam search) is applied to the large vocabulary continuous utterance recognition system developed by the present inventors. The results will be described. For the large vocabulary continuous speech recognition system,
IEICE Technical Report SP96-102, Yoshiaki Noda, Shoichi Matsunaga, Shigeki Sagayama, "A Study on Approximate Calculation Method in Large Vocabulary Continuous Speech Recognition Using Word Graph" (1
997). The acoustic model is a triphone HMM with a total number of 2000 states and a mixture number of 8 using 6,700 sentences as learning data from one month of a news program.
The feature amount is a total of 39 dimensions of 12 dimensions of MFCC and its first and second order regression coefficients, and log power and its first and second order regression coefficients. The language model is 50 for news program manuscripts for four years.
The word 2-gram and the word 3-gram learned in the transcription of the news program audio for one month for all sentences.
For the evaluation set, 50 sentences (total words: 1800, average utterance length: 12 seconds) were selected from the news program for 5 days. The first
The word error rate of the hypothesis (optimum solution) having the highest score among the hypotheses included in the word lattice obtained as a result of the path search was 9.51%.

FIG. 5A shows an experimental result of the N-best rescoring and the time asynchronous beam search which does not allow the above-mentioned word boundary time shift. From this, it can be seen that in the time asynchronous beam search, a solution can be obtained faster and more accurately than in the N-best rescoring. Subsequently, the effect of allowing a deviation of several msec was investigated in consideration of the deviation of the word boundary time in the time asynchronous beam search. The experiment was performed by changing the allowable shift from 10 to 50 msec based on the case where the shift shown in FIG. 5A is not allowed (0 msec). FIG. 5B shows the results.
Shown in From this, it can be seen that a higher accuracy solution can be obtained by allowing a displacement of about 20 msec, and a higher accuracy solution can be obtained by allowing the displacement than by not allowing the displacement. In this experiment, A ^* search was also evaluated. However, in the fourth sentence, a solution could not be obtained even after about 30 minutes. However, according to the present invention, all the solutions were obtained within a practical time, and it was confirmed that the present invention was superior to the A ^* search.

[0030]

As described above, according to the present invention, hypothesis development and pruning are performed such that the lengths of the hypotheses being developed are as uniform as possible, such as a time-synchronous beam search in which a stable solution can be obtained. By doing so, a solution can always be obtained. In addition, using the word boundary time and score information stored in the word lattice as a result of a path search using a coarse model such as A ^* search, and using the word boundary time stored in the word lattice with high accuracy Rather than using the word boundary time as it is in the path search using the model as it is, by allowing a shift of several frames, it is possible to obtain an effect that a final solution can be obtained with high accuracy, efficiently, and stably.

[Brief description of the drawings]

FIG. 1 is a hypothesis development diagram for explaining that a hypothesis whose search is delayed most, which is a main part of the present invention, is preferentially developed.

FIG. 2 is a diagram illustrating pruning by a score beam.

FIG. 3 is a flowchart showing an example of a processing procedure for preferentially developing a hypothesis whose search is delayed, which is a main part of the present invention, and for pruning;

FIG. 4 is an explanatory diagram for allowing a deviation of a word boundary time according to the present invention.

FIG. 5 is a view showing experimental results showing the effect of the present invention.

FIG. 6 is a diagram showing a general functional configuration of a speech recognition process.

FIG. 7 is a diagram showing a word network permitted by the grammar.

FIG. 8 is a diagram showing a functional configuration of a continuous speech recognition process based on a multipath search.

FIG. 9 is a view showing an example of a word lattice generated by a first pass search in FIG. 8;

FIG. 10 is a diagram showing a state of hypothesis development in a conventional A ^* search.

Continuation of the front page (72) Inventor Shoichi Matsunaga 2-3-1, Otemachi, Chiyoda-ku, Tokyo F-term in Nippon Telegraph and Telephone Corporation (reference) 5D015 AA01 BB01 HH23 LL03

Claims (5)

[Claims] 1. An acoustic model for obtaining an acoustic score indicating an acoustic closeness between a word and an input voice, and a language model for obtaining a grammar defining a connection relationship between words or a language score indicating an easy connection thereof. A search is performed on the continuously uttered input speech using a coarse acoustic model and a coarse language model, and a word network is created by narrowing down hypotheses that are close to the input speech from hypotheses of word strings allowed by the grammar. Then, using a higher-accuracy acoustic model and a higher-precision language model than the search, searching on the word network and converting the input speech from the hypothesis of a word string allowed in the word network to the input speech In a continuous speech recognition method in which a nearer one is narrowed down and a hypothesis of one or a plurality of word strings closest to the input speech is finally recognized as a recognition result, Search using the Le and language model,
For each word sequence hypothesis expansion, the hypothesis of the word sequence whose search is delayed is selected and performed, and for each word sequence hypothesis expansion, the condition deviated from the predetermined condition based on the obtained word string score A continuous speech recognition method characterized by terminating the development of a word string hypothesis. 2. Terminating the development of the hypothesis of the word string is performed by scoring a score g as a searched section to be performed when the hypothesis of the word string is developed.
_2. The continuous speech recognition method according to claim 1, wherein, if the calculation result of _n (t) or the score thereof becomes less than or equal to the threshold value during the calculation, the development of the hypothesis of the word string is terminated. 3. The terminating hypothesis expansion of the word string is 1
When the hypothesis development of the word string for one node (one word boundary on the word network) is completed, the hypothesis score f _n (t) of each word string is changed to the score g _{n of the} searched section.
(T) and the sum of the scores h _n (t) obtained in the previous search in the unsearched section, and the hypothesis f of this word string
3. The continuous speech recognition method according to claim 1, wherein the hypothesis expansion of word strings other than the m word string hypotheses is terminated in ascending order of _n (t). 4. The score h of the unsearched section_n(T)
Multiplied by weight α and f _n(T) = g_n(T) + αh
_nThe continuous sound according to claim 3, wherein (t) is set.
Voice recognition method. 5. The method according to claim 1, wherein the score calculation at the time of developing the hypothesis of the word string is performed within a range shifted from the word boundary time stored in the word network by about 5 ms or more. 5. The continuous speech recognition method according to any one of claims 1 to 4.