JP2000099089A - Search device for continuous voice recognition and search method for continuous voice recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent the omission of an optimum word train in search of the first stage and to enable searching candidates of word train without increasing search space in search of the second stage. SOLUTION: An analyzed result made by a voice analyzing means 102 is inputted, an optimum syllable train 4 obtained by an optimum solution obtaining means 2 refers to a difference model 6 in which likelihood corresponding to a syllable train of right solution and a word dictionary 7 in which standard syllable train of words is described. Then, candidates of a word train are searched, and a word train candidate 8 is outputted from a word train searching means 5.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a search device and a search method for continuous speech recognition capable of recognizing continuous speech composed of large vocabulary words and finding a correct word string candidate.

[0002]

2. Description of the Related Art In continuous speech recognition for recognizing continuous speech composed of large vocabulary words and finding candidates for word strings, a single-step search method and a multi-step search method for finding candidate word strings are available. There is a beam search method as a method of searching in one stage. In addition, as a method of searching in multiple stages, there is a method in which a word graph is created in the first stage, and a word string candidate is found in the word graph in the second stage.

[0003] The beam search method is a method of starting from an empty word string hypothesis as a hypothesis representing a partial string of a word string candidate, developing the hypothesis in synchronization with an input frame, and growing the word string. In this beam search method, as the number of possible word combinations increases with the progress of the input frame, and the number of candidates of the word string increases, the likelihood of the acoustic model and the likelihood of the language model are used. The pruning process is performed on the low hypothesis. By the pruning process, the number of hypotheses is reduced to a certain number, and the search is advanced so that the correct word string does not drop out of the candidate word strings. The likelihood is a logarithmic value of a probability that a standard syllable string is associated with an optimal syllable string.

On the other hand, the method using a word graph advances the search in two stages. First, word candidates are left in the first-stage search. For example, the search is performed by leaving only candidate words derived from the immediately preceding word. In the next search at the second stage, candidates of the word string are created by combining candidates of the word created at the first stage. At this time, the likelihood of the acoustic model and the likelihood of the language model are added to search for a candidate word string having a large likelihood. In the second search, a search using a stack decoder is used.

As a search method for searching in multiple stages, there is a method in which an optimum solution is obtained in the first stage and a search in the second stage is performed by modifying the optimum solution in the first stage.
No. 98 is disclosed. According to the present invention, in the first stage, dynamic programming (hereinafter referred to as DP: Dynami
c is called a programming method), and a recognition result is obtained by the DP method from a plurality of candidate pattern data selected in the first stage in the second stage. In this method, the optimal word string does not always match the correct word string, but is quite similar to the correct word string. However, if the pattern data of the correct word string is not found in the first row, the correct word string cannot be obtained even in the second row.

[0006]

The conventional search device for continuous speech recognition and the conventional search method for continuous speech recognition are constructed as described above. However, there is a problem that the correct answer cannot be obtained in the second stage when the error does not remain. In addition, when trying to leave a correct answer in the first row, there is a problem that the number of word candidates increases, the number of combinations of words to be considered in the processing in the second row increases, and the search space increases. Furthermore, there is a problem that the recognition accuracy is reduced because a word string candidate that is acoustically similar is searched.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems. In the first stage search, an optimal word string is prevented from being dropped, and in the second stage search, the search space is increased. It is an object of the present invention to obtain a search device for continuous speech recognition and a search method for continuous speech recognition that can search for candidates of a word string without having to search.

[0008]

A search device for continuous speech recognition according to the present invention is provided with a difference model expressing the likelihood that the optimal solution obtained in the first stage corresponds to the correct solution, and the first stage provides a difference model. A second stage search is performed by applying a difference model from the obtained optimal solution.

A search device for continuous speech recognition according to the present invention inputs an analysis result created by speech analysis means for analyzing an input speech, and controls an automaton representing a connection between syllables to find an optimal syllable string. A standard model of words and a difference model that describes the likelihood that the optimal syllable sequence obtained by the optimal solution obtaining device is input and the optimal syllable sequence obtained by the optimal solution obtaining device corresponds to the correct syllable sequence. A candidate word string is described with reference to a word dictionary describing a syllable string, and the candidate word string is output by a word string search unit.

A search apparatus for continuous speech recognition according to the present invention is characterized in that, in the difference model, a partial syllable sequence of an optimum syllable sequence, a partial syllable sequence of a correct syllable sequence, and a corresponding syllable sequence describing the likelihood thereof are described. The conversion likelihood table is used, and the word string search means searches for a candidate of a word string based on the likelihood described in the inter-syllable string conversion likelihood table.

The search apparatus for continuous speech recognition according to the present invention is characterized in that, in the difference model, a partial syllable string of an optimal syllable string, a partial syllable string of a correct syllable string, and a corresponding syllable string describing the likelihood thereof are described. A word syllable length conversion likelihood table which describes an optimal syllable string length, a syllable string length of a word dictionary, and a likelihood corresponding to the syllable string length, and a word string search means. A candidate of a word string is searched based on the likelihood described in the degree table and the word syllable length conversion likelihood table.

A search device for continuous speech recognition according to the present invention inputs an analysis result created by speech analysis means for analyzing an input speech, and controls an automaton representing a connection between syllables to find an optimal syllable string. The optimal solution obtained by the obtaining means is referred to a difference model describing the likelihood that the optimal syllable string obtained by the obtaining means corresponds to the correct syllable string, and the word dictionary describing the standard syllable string of the word. A difference model-applied word dictionary in which the optimal syllable string obtained by the acquisition means is input, and for each word in the word dictionary, a word and a syllable graph obtained by transforming a standard syllable string of the word dictionary based on the description in the difference model are described. And searches for candidates for the word string, and outputs the candidates for the word string by the word string search means.

A search device for continuous speech recognition according to the present invention inputs an analysis result created by speech analysis means for analyzing input speech, and controls an automaton representing connections between syllables to find an optimal syllable string. The optimal syllable string is input based on the optimal syllable string determined by the acquisition means, and the optimal syllable string determined by the optimal solution acquiring means is transformed based on the description of the difference model describing the likelihood corresponding to the correct syllable string. Then, a graph is created by the difference model applied syllable graph creating means, and the graph created by the difference model applied syllable graph creating means is input, and a word dictionary is described by referring to a word dictionary describing a standard syllable string of the word. In this method, a complement is searched for and a candidate of a word string is output by a word string search unit.

A search device for continuous speech recognition according to the present invention inputs an analysis result created by a speech analysis means for analyzing an input speech, and is controlled by an automaton representing a connection between syllables to form an optimum top N syllables. Is obtained by the N best solution obtaining means, and a syllable string composed of the optimum upper N syllables obtained by the N best solution obtaining means is inputted, and the optimum upper N syllables obtained by the N best solution obtaining means are obtained. Search for candidate word strings by referring to a difference model that describes the likelihood that a syllable string consisting of syllables corresponds to the correct syllable string and a word dictionary that describes a standard syllable string of words. Is output by the word string search means.

A search device for continuous speech recognition according to the present invention receives an analysis result created by a speech analysis means for analyzing an input speech, and controls the optimal N syllables controlled by an automaton representing connections between syllables. A standard model of a word and a difference model that describes the likelihood that the optimal syllable string composed of N syllables obtained by the N best solution obtaining means determines the likelihood corresponding to the correct syllable string. The syllable string described by the N best solution obtaining means is input by referring to the word dictionary describing the various syllable strings, and for each word in the word dictionary, the standard syllable of the word dictionary is described based on the word and the difference model. A candidate of a word string is searched for with reference to a difference model-applied word dictionary describing a syllable graph in which a string is transformed, and a candidate of a word string is output by a word string search unit.

A search device for continuous speech recognition according to the present invention inputs an analysis result created by a speech analysis means for analyzing an input speech, and controls an optimal N syllables controlled by an automaton representing connections between syllables. The best syllable string is obtained by the N best solution obtaining means, and the likelihood that the optimum N syllable string obtained by the N best solution obtaining means corresponds to the correct syllable string is described. Based on the description of the difference model, the optimal syllable string composed of N syllables is transformed to create a graph by the difference model applied syllable graph creation means,
By inputting the graph created by the difference model applied syllable graph creating means, referring to a word dictionary describing a standard syllable string of the word, searching for a candidate of the word string, and finding a candidate of the word string by the word string searching means. This is to output.

The search apparatus for continuous speech recognition according to the present invention provides a word syllable length conversion likelihood describing an optimal syllable string length, a syllable string length of a word dictionary, and a likelihood corresponding thereto in a difference model. Comprising a table, the word string search means comprises:
A candidate for a word string is searched for based on the likelihood of the word syllable length conversion likelihood table.

A search device for continuous speech recognition according to the present invention inputs an analysis result created by speech analysis means for analyzing an input speech, and controls an automaton representing a connection between words to find an optimal word sequence. The syllable string obtained by the syllable string conversion means is converted by the syllable string conversion means into the syllable string obtained by the syllable string conversion means. Referencing the difference model describing the likelihood corresponding to the correct syllable sequence and the word dictionary describing the standard syllable sequence of the word, searching for the candidate of the word sequence and searching for the candidate of the word sequence in the word sequence This is output by the search means.

A search device for continuous speech recognition according to the present invention inputs an analysis result created by speech analysis means for analyzing an input speech, and controls an automaton representing a connection between words to find an optimal word sequence. The difference model and the word that describe the likelihood that the optimum word string obtained by the optimum solution obtaining means is input and the optimum word string obtained by the optimum solution obtaining means corresponds to the correct word string are described. A candidate for a word string is searched for with reference to a word dictionary, and a candidate for a word string is output by a word string search unit.

A search device for continuous speech recognition according to the present invention includes a word syllable length conversion likelihood table in which a difference model describes the length of an optimal word string corresponding to a word in a word dictionary and the likelihood thereof. The word string search means is configured to search for a candidate for a word string based on the likelihood of the word syllable length conversion likelihood table.

In the search method for continuous speech recognition according to the present invention, a difference model expressing a likelihood corresponding to the optimum solution obtained in the first step and the correct answer is provided, and the difference is calculated from the optimum solution obtained in the first step. A second-stage search is performed by applying a model.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is inputted, and an optimal syllable string controlled by an automaton representing a connection between syllables is obtained. Search for candidate words in a word string by referring to a difference model describing the likelihood corresponding to the syllable string and a word dictionary describing a standard syllable string of words, and output candidate words in the word string Things.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is inputted, and an optimal syllable string controlled by an automaton representing a connection between syllables is obtained. For each word in the word dictionary, reference is made to the difference model describing the likelihood corresponding to the syllable sequence of the word and the word dictionary describing the standard syllable sequence of the word. With reference to a difference model-applied word dictionary that describes a syllable graph obtained by transforming a standard syllable string, a candidate for the word string is searched for and a candidate for the word string is output.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is input, and an optimum syllable string controlled by an automaton representing a connection between syllables is obtained. Based on the description of the difference model that describes the likelihood corresponding to the syllable sequence of, a graph was created by transforming the optimal syllable sequence, and the created graph was input to describe the standard syllable sequence of words The candidate of the word string is searched by referring to the word dictionary, and the candidate of the word string is output.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is input, and an optimal syllable string composed of upper N syllables controlled by an automaton representing a connection between syllables is obtained. Input a syllable string consisting of the optimal top N syllables, a difference model describing the likelihood that the optimal syllable string consisting of the top N syllables corresponds to the correct syllable string, and a standard syllable string of words. By referring to the described word dictionary, a candidate for a word string is searched for, and a candidate for the word string is output.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is inputted, and a syllable string composed of N optimal syllables controlled by an automaton representing a connection between syllables is obtained. A syllable string is input, and a difference model describing the likelihood that the optimal syllable string composed of N syllables corresponds to a correct syllable string and a word dictionary describing a standard syllable string of words are referred to. For each word in the word dictionary, refer to the difference model applied word dictionary that describes a word and a syllable graph obtained by transforming a standard syllable string of the word dictionary based on the description in the difference model, and search for a candidate for the word string. , The candidate of the word string is output.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is inputted, and a syllable string composed of N optimal syllables controlled by an automaton representing a connection between syllables is obtained. A graph is created by deforming an optimal syllable string of N syllables based on the description of the difference model that describes the likelihood that the syllable string of N syllables corresponds to the correct syllable string. Then, the input graph is input, a candidate word string is searched for by referring to a word dictionary describing a standard syllable string of the word, and the candidate word string is output.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is inputted, an optimum word string controlled by an automaton representing a connection between words is obtained, and this optimum word string is converted to a syllable. Converted into a sequence, this syllable sequence is referred to a difference model that describes the likelihood corresponding to the correct syllable sequence and a word dictionary that describes a standard syllable sequence of words, and searches for candidates for the word sequence. A candidate for a word string is output.

In the search method for continuous speech recognition according to the present invention, an analysis result of an input speech is inputted, and an optimal word sequence controlled by an automaton representing a connection between words is obtained. With reference to a difference model describing the likelihood corresponding to the word string and a word dictionary describing the word, a candidate for the word string is searched for, and a candidate for the word string is output.

[0030]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a configuration diagram showing a search device for continuous speech recognition according to a first embodiment of the present invention. In FIG. 1, reference numeral 101 denotes an input speech, and 102 analyzes the input speech 101 and converts it into a feature vector time series 103. Voice analysis means 2, an optimal solution obtaining means for inputting the feature vector time series 103 and obtaining an optimal syllable string 4 according to the syllable network 3;
Reference numeral 5 denotes a word string search unit that inputs the optimal syllable string 4 and refers to the difference model 6 and the word dictionary 7 to search for a word string candidate 8.

FIG. 2 is an explanatory diagram showing a syllable network in the search device for continuous speech recognition according to the first embodiment of the present invention. FIG. 3 is a diagram showing the search device for continuous speech recognition according to the first embodiment of the present invention. FIG. 3 is an explanatory diagram showing a basic HMM. The syllable network 3 is a network representation of the connection of syllables (generally words or subwords), and is composed of nodes for connecting syllables and arcs representing syllables as shown in FIG. Syllable arcs are represented by a chain of elementary HMMs as shown in FIG. To consider the effect of articulatory coupling within or between syllables, the basic HMM
Is used as a phoneme environment-dependent phoneme model.

FIG. 4 is an algorithm showing an automaton control in the search apparatus for continuous speech recognition according to the first embodiment of the present invention. FIG. 5 is a diagram showing a search method for continuous speech recognition according to the first embodiment of the present invention. FIG. 4 is an explanatory diagram illustrating an example of a dictionary. The optimal solution acquiring means 2 acquires an optimal syllable string corresponding to the feature vector time series 103 based on the automaton control one-pass DP algorithm shown in FIG.
Output as optimal syllable string 4. When the optimal syllable string 4 is input, the word string search means 5 searches for a candidate of the word string with reference to the word dictionary 7. The word dictionary 7 includes word descriptions and standard syllable string descriptions as shown in FIG.

FIG. 6 is a block diagram showing a difference model in the search device for continuous speech recognition according to the first embodiment of the present invention. FIG. 7 is a diagram showing a search device for continuous speech recognition according to the first embodiment of the present invention. It is a table | surface figure which shows the example of the conversion likelihood table between syllable strings. The difference model 6 includes a syllable string conversion likelihood table 601 as shown in FIG. As shown in FIG. 7, the inter-syllable string conversion likelihood table 601 describes the optimal syllable string corresponding to the standard syllable string, and the likelihood that the standard syllable string is converted into the optimum syllable string. This likelihood is a logarithmic value of the probability that the standard syllable string is associated with the optimal syllable string. The length of the standard syllable string and the optimum syllable string may be any value equal to or greater than zero. In the figure, the standard syllable string has a length of 1 and the optimal syllable string has a length of 1-2.

FIG. 8 is a block diagram showing an example of learning means for a difference model in the search device for continuous speech recognition according to the first embodiment of the present invention. The difference model is learned by learning means having a configuration as shown in FIG. Voice database 10
, An input speech 101 is obtained, and is converted into a feature vector time series 103 by a speech analysis unit 102. The optimal solution obtaining means 2 refers to the syllable network 3 and outputs the optimal syllable sequence 4 for the feature vector time series 103. The optimal syllable string 4 is input to the difference model learning means 9 together with the correct word string 11 and the correct syllable string 12 obtained from the speech database 10. The difference model learning means 9 performs DP matching between the optimal syllable string 4 and the correct syllable string 12, and finds a correspondence between them on the time axis. By performing this for all the speeches in the speech database 10, the likelihood that the partial syllable sequence 12 of the optimal syllable sequence 4 corresponds to the partial syllable sequence of the correct syllable sequence is obtained, and the difference model 6 is output.

Next, the operation will be described. FIG. 9 is a flowchart showing an operation procedure of the word string searching means in the search device for continuous speech recognition according to the first embodiment of the present invention. The search is performed according to the flowchart shown in FIG. 9 based on the stack decoder. In this stack decoder, a word search is started from the beginning of the optimal syllable string 4, words in the word dictionary 7 are sequentially combined, and a candidate for a word string covering from the beginning to the end of the optimal syllable string 4 is obtained. Here, a word string candidate covering the beginning to the middle of the optimal syllable string 4 is assumed to be a hypothesis. One hypothesis is that
It has a word string, an end time, and an evaluation value. The end time is the length of the optimal syllable string 4 covered by the word string of the hypothesis, and if the entire length of the optimal syllable string 4 is T, the end time is an integer value in the range of 0 to T. .

For example, if the optimal syllable string 4 is "Onse-nen-shiki-sochi", the word string of the hypothesis covering the entire optimal syllable string 4 is "speech (onse-e) recognition". )
The end time is 11.
If the word string of the hypothesis is “speech recognition, the end time of the hypothesis is 8.

Next, the operation of the stack decoder used in the first embodiment will be described. First, a hypothesis consisting of an empty word string is created and stored in the stack (step ST10).
1) It is determined whether or not the stack is empty (step ST10)
2) When the stack becomes empty, the process ends (step ST103). Next, when the stack is not empty in the judgment of step ST102, the hypothesis H0 having the largest evaluation value is extracted from the stack (step ST10).
4), let T0 be the end time of hypothesis H0. Next, hypothesis H
It is determined whether the end time T0 of 0 is equal to the optimal syllable length T (step ST111). If it is equal to the optimal syllable length T, the word string of the hypothesis is output as one of the word string candidates 8. (Step ST112), and returns to step ST104. On the other hand, if the end time T0 of the hypothesis H0 is not equal to the optimal syllable length T in step ST111, the word dictionary 7
, One word is taken out from the list and the word is set as n (step ST105). Hereinafter, the processing of steps ST106 to ST110 is performed for an arbitrary word n in the word dictionary.

In step ST106, the range from T0 + 1 as the start time to the end T is set to the end time T1 (T1: T
0 + 1 to T), and collate the optimal syllable string with the standard syllable string of the word n. In this collation, a matching likelihood is obtained from the likelihood that the standard syllable string and the optimal syllable string correspond between the partial optimal syllable string W1 and the standard syllable string W2 of the word n.

W1 = X (T0 + 1), X (T0 + 2) to X (T1) (1) W2 = Y (1), Y (2) to Y (J (n)) (2) T1: integer in the range of T0 to T J (n): standard syllable string length of word n

Next, the collation between the optimal syllable string and the standard syllable string of the word n is performed based on the flowchart of FIG. FIG.
0 is a flowchart showing a procedure for collating the optimum syllable string with the standard syllable string of the word n in the search device for continuous speech recognition according to the first embodiment of the present invention. First, an optimal syllable string and a standard syllable string of the word n are given (step ST20).
1) For each, the standard syllable string and the optimal syllable string are converted into graphs G1 and G2 composed of states and transitions (steps ST202 and ST203). Next, it is determined whether the application target of the difference model is the optimal syllable sequence or the standard syllable sequence (step ST204). When the application target of the difference model is the optimal syllable sequence, the difference model 6 is applied to the graph G1, and Is obtained (step ST205).
On the other hand, as a result of the determination in step ST204, when the application target of the difference model is a standard syllable string, the difference model 6 is applied to the graph G2 to obtain a graph G2 ′ after the change (step ST207).

Here, referring to FIG. 11, step ST20 will be described.
2 to ST208 will be described with reference to specific examples. FIG.
FIG. 1 is an explanatory diagram showing a collation operation in the search device for continuous speech recognition according to the first embodiment of the present invention. FIG.
The optimal syllable sequence is "Onsenninhisosochi", of which the syllable sequence "Ninhi" and the standard syllable sequence "Ninshiki"
Indicates when to match between At this time, step ST
The graph of the optimal syllable string becomes G1 by 202, and the graph of the standard syllable string becomes G2 by step ST203. In addition, in the syllable string conversion likelihood table 601 as the difference model 6, “dish / hi [−2.3]”, “dish / hide [−0.1]”, “hi / hi [−0. 1] ”and other syllables (referred to as X) are“ X / X [0.
0]]. Step ST204
When the application target of the difference model 6 is an optimal syllable string, the graph G1 is represented by the difference model 6 in step ST205.
G1 '"Nin (hi [-0.1] -shiki [-2.3])"
Is transformed. As a result, it is possible to collate “Ninshi” in the graph G2 with the optimal syllable string, and to search for a word string including the correct word “recognition (Ninshi)” as a candidate word string.

In step ST204, the difference model 6
When the target of application is a standard syllable string, the step ST
In 207, the graph G2 is transformed into G2 ′ “Nin ([−0.1]) − H [−2.3]” by the difference model. As a result, it is possible to collate “Ninshi” in the graph G2 with the optimal syllable string, and to search for a word string including the correct word “recognition (Ninshi)” as a candidate word string. Between the graphs changed in step ST206 or ST208 (G1 ′ and G2 or G1 and G
2 ′), the following recurrence formula is calculated to obtain the matching likelihood D (W
1, W2).

G (j, n) = 0, (j, n) {set of initial nodes} (3) G (j, n) = −）, (j, n) {initial node Other than the set｝ (4) G (j, n) = max G (i, m) + g (i → j) + g (m → n) + e (i → j, m → n), (i, m) {possible transition for node (j, n)} (5) D (W1, W2) = max G (j, n), (j, n) {set of last node} (6) where i and j are graph states on the optimal syllable string side, and m and n
Is the state of the graph on the standard syllable string side, g (i → j) and g
(M → n) are the state transition i → j and the state transition m, respectively.
→ log likelihood of n, e (i → j, m → n) is the state transition i →
j and syllable X (i → j) on the optimal syllable string side and syllable Y (m → n) on the standard syllable string side associated with state transition m → n
Here, 0 is determined when the values match, and −∞ when the values do not match.

In the flowchart of FIG. 9, it is determined whether or not the matching likelihood D (W1, W2) is higher than the threshold (step ST107). If the matching likelihood D (W1, W2) is not higher than the threshold, Step ST108 to Step ST
Step 110 is not performed. On the other hand, if the result of determination in step ST107 is that the matching likelihood D (W1, W2) is higher than the threshold, the processing of steps ST108 to ST110 is performed. In step ST108, the hypothesis H0 is copied to create the hypothesis H1, the end time of the hypothesis H1 is updated, and
1 (step ST109), the word n is added to the word string of the hypothesis H1, and the word string is grown by one word. Further, the evaluation value of the hypothesis H1 is increased by the matching likelihood D (W1, W2). Next, the hypothesis H1 is stored in the stack (step S1).
T110). The likelihood of the language model of the word string is calculated and added to the evaluation value of the hypothesis H1 together with the matching likelihood. In this case, the likelihood of the language model is calculated using the N-gram model for the word string.

As described above, according to the first embodiment, the difference model is applied from the optimal solution obtained by the optimal solution obtaining means at the first stage of the search, and the first stage of the search is performed at the second stage of the search. And a word string search unit for searching for a candidate for a word string by referring to a difference model and a word dictionary describing words, thereby preventing dropout of the optimum solution at the first stage.
At the stage, effects such as a decrease in the number of correct answers can be obtained.

Embodiment 2 FIG. 12 is an explanatory diagram showing a word dictionary in the search device for continuous speech recognition according to the second embodiment of the present invention. FIG. 13 is a diagram showing a difference model in the search device for continuous speech recognition according to the second embodiment of the present invention. FIG. 14 is a diagram showing an example of a word syllable length conversion likelihood table in the search device for continuous speech recognition according to the second embodiment of the present invention. The same reference numerals are used to indicate the same or corresponding parts, and a description thereof will be omitted. The word dictionary 7 of the second embodiment has a length J of a standard syllable string forming a word n as shown in FIG.
(N). The difference model 6 according to the second embodiment has a syllable sequence conversion likelihood table 60 as shown in FIG.
1 and a word syllable length conversion likelihood table 602. The word syllable length conversion likelihood table 602 describes the likelihood associated with the length of the standard syllable string and the length of the optimal syllable string forming a word as shown in FIG.

Next, the operation will be described. In the matching in step ST106 in FIG. 9, the likelihood of matching between the sub-optimal syllable string W1 and the standard syllable string W2 of the word n includes the matching likelihood D (W1, W2) obtained by the method described in the first embodiment. ), And the likelihood of the word syllable length is added. The likelihood of the word syllable length is determined by the optimal syllable length currently being matched (this is T1-T
0) and the length of the standard syllable string constituting word n (this is J (n) and obtained from word dictionary 7),
The likelihood is obtained by subtracting the word syllable length conversion likelihood table 602. Accordingly, when the matching is performed with the optimal syllable length largely different from the word standard syllable length, the likelihood is reduced, and the processing of steps ST108 to ST110 is performed by the determination processing of the matching likelihood and the threshold in step ST107 in FIG. Will not be.

As described above, according to the second embodiment, effects such as extreme collation can be prevented, unnecessary generation of hypotheses can be reduced, and the amount of search processing can be reduced.

Embodiment 3 FIG. 15 is a block diagram showing a search device for continuous speech recognition according to a third embodiment of the present invention. In the figure, the same reference numerals as those in the first and second embodiments denote the same or corresponding parts, and will be described. Is omitted. Prior to the recognition processing by the word string search means 5, the difference model applied word dictionary creating means 14 converts the standard syllable string of the word dictionary 7 into a graph and stores it as the difference model applied word dictionary 13. Next, in the processing of the word string search means 5, step ST106 of FIG.
When the word n is added after the end time T0 of the hypothesis H0 in FIG.
Steps 3 and ST207 for converting the standard syllable string into a graph can be omitted.

As described above, according to the third embodiment, since the result of previously converting a word standard syllable string into a graph is stored for all words, the amount of memory is increased. In the processing of No. 5, the operation of the dynamic conversion can be omitted, so that effects such as speeding up of the arithmetic processing can be obtained.

Embodiment 4 FIG. FIG. 16 is a configuration diagram showing a search device for continuous speech recognition according to a fourth embodiment of the present invention. In the figure, the same reference numerals as those in the first to third embodiments indicate the same or corresponding parts. Description is omitted. In the fourth embodiment, a difference model applied syllable graph creating unit 15 is provided.
The difference model 6 is applied to create a difference model applied input syllable graph 16. The word string search means 5 outputs the word string candidate 8 with reference to the word dictionary 7 with the input syllable graph 16 to which the difference model is applied. Therefore, since the difference model 6 is applied to the optimum syllable string, it is only necessary to apply the difference model 6 only once for one optimum voice.

As described above, according to the fourth embodiment, as in the first embodiment, the word n
Since it is not necessary to apply the difference model 6 to the optimal syllable string or the standard syllable string of the word n for each case, it is possible to obtain effects such as speeding up the arithmetic processing.

Embodiment 5 FIG. 17 is a configuration diagram showing a search device for continuous speech recognition according to a fifth embodiment of the present invention. In the figure, the same reference numerals as those in the first embodiment denote the same or corresponding parts, and a description thereof will be omitted. In the first embodiment, the difference model 6 is applied to the optimum syllable string, and the correct answer is obtained by the deformation of the optimum syllable string. However, in some cases, the correct answer cannot be obtained. However, in the fifth embodiment, the N best solutions of the syllables are obtained by using the N best solution obtaining means 21 instead of the optimum solution obtaining means 2 of the first embodiment, and the N best syllable graph 22 is obtained.
Is output.

As described above, according to the fifth embodiment, the N best candidates of the optimum syllable are obtained, and the obtained N best candidates are used as the N best syllable graph 22 to search for the word sequence candidates 8 using the difference model 6. Is performed, the number of cases in which a correct answer is not required is reduced, and effects such as an improvement in recognition rate can be obtained.

Embodiment 6 FIG. FIG. 18 is a configuration diagram showing a search device for continuous speech recognition according to a sixth embodiment of the present invention. In the figure, the same reference numerals as those in the first to fifth embodiments denote the same or corresponding parts, and thus will be described. Is omitted. In the sixth embodiment, an N best solution obtaining means 21 for outputting an N best syllable graph 22 including the N best candidates is provided in place of the optimum solution obtaining means 2 of the third embodiment. In the third embodiment, a word sequence candidate is searched from an optimal syllable sequence using a difference model applied word dictionary 13 created in advance by applying the difference model 6 to a standard syllable sequence of words in the word dictionary 7. The optimal syllable sequence was one type. For this reason, there is a possibility that the difference model applied word dictionary 13 cannot search for a correct word string.
However, according to the sixth embodiment, the N best candidates of the optimum syllables are obtained, and the search for the word sequence candidates is performed using the difference model applied word dictionary 13 as the optimum syllable graph allowing ambiguity. Is what you do.

As described above, according to the sixth embodiment, it is possible to reduce the number of cases in which a correct word string cannot be obtained and to improve the recognition rate. Compared to Embodiment 5 in which the syllable sequence on the dictionary side is modified, it is possible to obtain an effect such that a recognition result having a different tendency can be obtained.

Embodiment 7 FIG. FIG. 19 is a configuration diagram showing a search device for continuous speech recognition according to a seventh embodiment of the present invention. In the figure, the same reference numerals as those in the first to sixth embodiments denote the same or corresponding parts, and thus will be described. Is omitted. In the seventh embodiment, instead of the optimum solution obtaining means 2 of the fourth embodiment, the N best solution obtaining means 21 obtains the N best candidates of syllables and outputs the N best syllable graph 22. In the fourth embodiment, the difference model 6 is applied to one optimal syllable string 4 obtained by the optimum solution obtaining means 2 to create a difference model applied input syllable graph 16. I was searching. Therefore, there is a possibility that the word dictionary 7 to which the difference model 6 is applied cannot search for a correct word string. However, in the seventh embodiment, the N best candidates of the optimal syllable are obtained, the N best syllable graph 22 is created based on this, and the word sequence candidate 8 is applied after applying the difference model 6.
The search is performed.

As described above, according to the seventh embodiment, the number of cases in which a correct word string cannot be obtained is reduced, the recognition rate is improved, and the same N best solution obtaining means 2 is used.
Compared to the sixth embodiment using No. 1, the difference is that the syllable sequence on the optimum side is deformed, and effects such as recognition results having different tendencies can be obtained.

Embodiment 8 FIG. FIG. 20 is a block diagram showing a search apparatus for continuous speech recognition according to an eighth embodiment of the present invention. In the figure, the same reference numerals as those in the first to seventh embodiments denote the same or corresponding parts, and will be described. Is omitted. In the eighth embodiment, the optimal solution obtaining means 2
However, using the word network 17, an optimum word string 18 is obtained, and converted into a syllable string by the syllable string conversion means 19 to obtain the optimum syllable string 4, which is input to the word string search means 5. Therefore, since the optimal solution obtaining means 2 uses a word which is an acoustically longer unit instead of a syllable as a network to be referred to, the optimal word string 18 which is less affected by articulation coupling can be obtained.

As described above, according to the eighth embodiment, the optimal word string 1 that is less affected by articulation coupling
Since 8 is returned to the optimal syllable string 4, effects such as an increased possibility of finding a correct answer are obtained.

Embodiment 9 FIG. FIG. 21 is a configuration diagram showing a search device for continuous speech recognition according to Embodiment 9 of the present invention.
FIG. 22 is a block diagram showing a difference model in the search device for continuous speech recognition according to the ninth embodiment of the present invention.
FIG. 15 is a table showing a word string word-to-word conversion table in the search device for continuous speech recognition according to the ninth embodiment of the present invention. In the figure, the same reference numerals as those in the first to eighth embodiments denote the same parts. Or, since a corresponding portion is shown, the description is omitted. In the ninth embodiment, the optimal solution obtaining means 2
Is obtained by using the word network 17 to obtain an optimum word string 18 and using the word string as an input to the word string search means 5. The word string search means 5 refers to the difference model 6 and the word dictionary 7 to search for the word string. Search for candidate 8.

In the ninth embodiment, the difference model 6 is composed of a word string inter-word conversion likelihood table 603 as shown in FIG. 22, and corresponds to the partial word string column of the optimum word string 18 as shown in FIG. The column of the correct word and the likelihood are described. When the partial word string of the optimum word string 18 is given together with the word, the table is searched to obtain the likelihood.

Next, the operation will be described. FIG. 24 is a flowchart showing a word string search procedure of the word string search means in the search device for continuous speech recognition according to the ninth embodiment of the present invention, and FIG. 25 is a search for continuous speech recognition according to the ninth embodiment of the present invention. 6 is a flowchart showing a word and a partial word string of an optimum word string and a collation procedure in the apparatus. The word string search in the word string search means 5 is performed based on the flowchart of FIG. In step ST301, FIG.
5 is performed based on the flowchart of FIG. First, the optimal word string 18 and the word n are given (step ST40).
1) Match the word n with the partial word string of the optimum word string 18 to obtain the likelihood of the matching. Matching likelihood is optimal word string 1
The likelihood corresponding to the word n in part 8 is obtained by subtracting the word string inter-word conversion likelihood table 603 of the difference model 6 (step ST402).

As described above, according to the ninth embodiment, the word n and the optimum word string 18
Since the collation with the partial word string is realized by the table search, effects such as that the search for the word string candidate 8 becomes easy can be obtained.

[0065]

As described above, according to the present invention, a difference model expressing the likelihood corresponding to the optimal solution obtained at the first stage and the correct solution is provided, and the difference model is obtained from the optimal solution obtained at the first stage. Since the second stage search was performed by applying the model,
The first stage has the effect of preventing the drop of the optimal solution, and the second stage has the effect of reducing the drop of the correct solution.

According to the present invention, the analysis result prepared by the voice analysis means for analyzing the input voice is input, and the optimum syllable string controlled by the automaton representing the connection between syllables is obtained by the optimum solution obtaining means, and the optimum solution is obtained. A difference model that describes the likelihood that the optimal syllable string determined by the acquisition means corresponds to the correct syllable string, and a word dictionary that describes the standard syllable string of words , The candidate of the word string is searched for, and the candidate of the word string is output from the word string searching means. This has the effect of reducing the dropout of correct answers.

According to the present invention, in the difference model,
An inter-syllable string conversion likelihood table describing the partial syllable string of the optimal syllable string and the partial syllable string of the correct syllable string and their corresponding likelihoods, and the word string search means is described in the inter-syllable string conversion likelihood table. The candidate of the word string is searched based on the likelihood, so that the first stage can prevent the drop of the optimal solution,
In the second stage, there is an effect that the drop of correct answers can be reduced.

According to the present invention, in the difference model,
An inter-syllable string conversion likelihood table describing the partial syllable string of the optimal syllable string, the partial syllable string of the correct syllable string, and their corresponding likelihood, the length of the optimal syllable string, the length of the syllable string of the word dictionary, and the like. Has a word syllable length conversion likelihood table describing the likelihood corresponding thereto, and the word string search means uses the word string based on the likelihood described in the inter-syllable string conversion likelihood table and the word syllable length conversion likelihood table. Since it was configured to search for a candidate,
Extreme collation can be prevented, useless hypotheses are reduced, and the amount of search processing can be reduced.

According to the present invention, the analysis result prepared by the voice analysis means for analyzing the input voice is input, and the optimum syllable sequence controlled by the automaton representing the connection between syllables is obtained by the optimum solution obtaining means, and the optimum solution is obtained. The optimal syllable string obtained by the optimal solution obtaining means is referred to the difference model describing the likelihood that the optimum syllable string obtained by the obtaining means corresponds to the correct syllable string and the word dictionary describing the standard syllable string of words. For each word in the word dictionary, refer to the difference model applied word dictionary that describes the word and a syllable graph obtained by transforming the standard syllable string of the word dictionary based on the description in the difference model, Although the configuration is such that the word string candidate is searched for and the word string candidate is output from the word string search means, the amount of memory is increased. However, in the processing of the word string search means, the operation of dynamic conversion can be omitted. Fast processing There is an effect that can be.

According to the present invention, the analysis result prepared by the voice analysis means for analyzing the input voice is input, the optimum syllable sequence controlled by the automaton representing the connection between syllables is obtained by the optimum solution obtaining means, and the optimum syllable is obtained. Input a sequence and transform the optimal syllable sequence based on the description of the difference model that describes the likelihood that the optimal syllable sequence determined by the optimal solution obtaining means corresponds to the correct syllable sequence. Means, input the graph created by the difference model applied syllable graph creating means, refer to a word dictionary describing a standard syllable string of words, search for candidates of word strings, Since the complement is configured to be output from the word string search means, there is an effect that the arithmetic processing can be sped up.

According to the present invention, the analysis result created by the speech analysis means for analyzing the input speech is input, and the optimal syllable string composed of the upper N syllables controlled by the automaton representing the connection between the syllables is converted to N. The best syllable string composed of the top N syllables obtained by the best solution obtaining means and obtained by the N best solution obtaining means is inputted. By referring to the difference model describing the likelihood corresponding to the correct syllable sequence and the word dictionary describing the standard syllable sequence of the word, the candidate of the word sequence is searched, and the candidate of the word sequence is searched by the word sequence searching means. Since it is configured to output, it is possible to reduce the number of cases where a correct answer is not required and to improve the recognition rate.

According to the present invention, the analysis result created by the speech analysis means for analyzing the input speech is input, and the optimal syllable string controlled by the automaton representing the connection between syllables and composed of N syllables is converted into the N best syllables. A difference model describing the likelihood that the optimal syllable string composed of N syllables obtained by the obtaining means and obtained by the N best solution obtaining means corresponds to the correct syllable string, and a word describing the standard syllable string of the word A syllable graph obtained by inputting the optimal syllable sequence obtained by the N best solution obtaining means with reference to the dictionary, and transforming a standard syllable sequence of the word dictionary based on the description in the word and the difference model for each word in the word dictionary; Is configured to search for candidates for word strings by referring to the difference model applied word dictionary that describes the word strings, and to output candidates for word strings by means of word string search means. , Recognition rate With effect that improves, except that it modified the syllable sequence of dictionary side, there is an effect that it is possible to obtain different recognition results tend.

According to the present invention, the analysis result created by the speech analysis means for analyzing the input speech is input, and the optimal syllable string composed of N syllables controlled by the automaton representing the connection between the syllables is converted into the N best solution. The optimal syllable string obtained by the obtaining means is input, and the optimal syllable string composed of N syllables obtained by the N best solution obtaining means is described based on the description of the difference model describing the likelihood corresponding to the correct syllable string. A syllable string composed of N syllables is transformed to create a graph by the syllable graph applying means applying the difference model, and the graph created by the syllable graph applying means applying the difference model is inputted, and a standard syllable of the word is input. It is configured to search for candidates of word strings by referring to the word dictionary describing the strings and output candidates of word strings by the word string search means. Rate Causes the above, except that it modified the syllable sequence of optimal side effects such as can be obtained with different recognition results tend to obtain.

According to the present invention, in the difference model,
A word syllable length conversion likelihood table describing the length of an optimal syllable string, the length of a syllable string in a word dictionary, and the likelihood corresponding thereto; Is configured to search for candidates of a word string based on the above. Therefore, it is possible to prevent an extreme collation, reduce the generation of useless hypotheses, and reduce the amount of search processing.

According to the present invention, the analysis result created by the speech analysis means for analyzing the input speech is input, and the optimum word string controlled by the automaton representing the connection between words is obtained by the optimum solution obtaining means, and the optimum solution is obtained. The optimal word string obtained by the acquiring means is converted into a syllable string by the syllable string converting means, and the optimal syllable string obtained by the syllable string converting means is inputted, and the syllable string obtained by the syllable string converting means becomes a correct syllable string. Referring to the difference model describing the corresponding likelihood and the word dictionary describing the standard syllable string of the word, the candidate of the word string is searched, and the candidate of the word string is output by the word string search means. With this configuration, the optimal word string that is less affected by articulation coupling is returned to the optimal syllable string, so that there is an effect that the possibility of finding a correct answer can be increased.

According to the present invention, the analysis result created by the speech analysis means for analyzing the input speech is input, and the optimum word string controlled by the automaton representing the connection between words is obtained by the optimum solution obtaining means. The optimal word string obtained by the obtaining means is input, and the optimal word string obtained by the optimal solution obtaining means is referred to a difference model describing the likelihood corresponding to the correct word string and the word dictionary describing the word. Since the candidate of the string is searched and the candidate of the word string is output by the word string searching means, there is an effect that the search for the candidate of the word string can be facilitated.

According to the present invention, in the difference model,
A word syllable length conversion likelihood table describing the length and likelihood of an optimal word string corresponding to a word in the word dictionary; and a word string search unit, based on the likelihood of the word syllable length conversion likelihood table. Since it is configured to search for a candidate in a column, it is possible to prevent extreme collation, to reduce generation of useless hypotheses, and to reduce the amount of search processing.

According to the present invention, a difference model expressing the likelihood that the optimal solution obtained in the first stage corresponds to the correct solution is provided,
Since the second step is performed by applying the difference model from the optimal solution obtained in the first step, it is possible to prevent the optimal solution from dropping out in the first step, and to determine the correct answer in the second step. There is an effect that the dropout can be reduced.

According to the present invention, an analysis result of an input speech is input, and an optimal syllable string controlled by an automaton representing a connection between syllables is obtained, and this optimal syllable string corresponds to a correct syllable string. It is configured to search for candidate word strings and output candidate word strings by referring to a difference model describing the degree and a word dictionary describing a standard syllable string of words. Thus, the generation of useless hypotheses can be reduced, and the amount of search processing can be reduced.

According to the present invention, an analysis result of an input speech is input, and an optimal syllable string controlled by an automaton representing a connection between syllables is obtained, and this optimal syllable string corresponds to a correct syllable string. By referring to the difference model describing the degree and the word dictionary describing the standard syllable string of the word, for each word in the word dictionary, the standard syllable string of the word dictionary is transformed based on the description in the word and the difference model. By referring to the difference model applied word dictionary describing the syllable graph and the candidate word string, the candidate word string is searched for and the candidate word string is output, so that the memory amount increases. In the processing, since the operation of the dynamic conversion can be omitted, there is an effect that the operation process can be sped up.

According to the present invention, an analysis result of an input speech is inputted, and an optimal syllable string controlled by an automaton representing a connection between syllables is obtained, and this optimal syllable string corresponds to a correct syllable string. A graph is created by transforming the optimal syllable sequence based on the description of the difference model describing the degree, and the created graph is input, and the word sequence is referred to the word dictionary that describes the standard syllable sequence of the word. Since the candidate is searched for and the candidate of the word string is output, there is an effect that the arithmetic processing can be speeded up.

According to the present invention, an analysis result of an input speech is input, and an optimal syllable string composed of upper N syllables controlled by an automaton representing connections between syllables is obtained. Syllable string consisting of the following, and refer to the difference model that describes the likelihood that the optimal syllable string consisting of the top N syllables corresponds to the correct syllable string and the word dictionary that describes the standard syllable string of words Since the candidate of the word string is searched for and the candidate of the word string is output, it is possible to reduce the number of cases in which a correct answer is not required and to improve the recognition rate.

According to the present invention, an analysis result of an input speech is inputted, a syllable string composed of N optimum syllables controlled by an automaton representing a connection between syllables is obtained, and this optimum syllable string is inputted. With reference to the difference model describing the likelihood that the syllable string composed of the optimal N syllables corresponds to the correct syllable string and the word dictionary describing the standard syllable string of the word, for each word in this word dictionary, By referring to the difference model applied word dictionary that describes a word and a syllable graph obtained by transforming a standard syllable string of the word dictionary based on the description in the difference model, searching for a candidate for the word string, and finding a candidate for the word string. Because it was configured to output, reduce the case where the correct word string is not found,
The effect of improving the recognition rate is obtained, and the effect is that the syllable sequence on the dictionary side is different, so that recognition results with different tendencies can be obtained.

According to the present invention, an analysis result of an input speech is input, and an optimal syllable string composed of N syllables controlled by an automaton representing a connection between syllables is obtained. Based on the description of the difference model that describes the likelihood that the syllable string corresponds to the correct syllable string, a graph is created by modifying the optimal syllable string composed of N syllables, and the created graph is input. By referring to a word dictionary that describes a standard syllable string of words, searching for candidates for word strings and outputting candidates for word strings, the number of cases where correct word strings are not obtained is reduced, While improving the recognition rate,
The difference is that the syllable sequence on the optimum side is modified, and effects such as the recognition results having different tendencies can be obtained.

According to the present invention, an analysis result of an input speech is input, an optimum word string controlled by an automaton representing a connection between words is obtained, and this optimum word string is converted into a syllable string. Referencing a difference model describing the likelihood corresponding to the correct syllable sequence and a word dictionary describing the standard syllable sequence of words, searching for candidate words and outputting candidate words With this configuration, the optimal word string that is less affected by articulation coupling is returned to the optimal syllable string, so that the possibility of finding the correct answer can be increased.

According to the present invention, an analysis result of an input speech is input, and an optimal word sequence controlled by an automaton representing a connection between words is obtained. This optimal word sequence corresponds to a correct word sequence. By referring to the difference model describing the degree and the word dictionary describing the word, the candidate of the word string is searched, and the candidate of the word string is output, so that the search for the candidate of the word string can be easily performed. There is an effect that can be.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing a search device for continuous speech recognition according to a first embodiment of the present invention.

FIG. 2 is an explanatory diagram showing a syllable network in the search device for continuous speech recognition according to the first embodiment of the present invention.

FIG. 3 is an explanatory diagram showing a basic HMM in the search device for continuous speech recognition according to the first embodiment of the present invention;

FIG. 4 is an algorithm showing automaton control in the search device for continuous speech recognition according to the first embodiment of the present invention.

FIG. 5 is an explanatory diagram showing an example of a word dictionary in the search device for continuous speech recognition according to the first embodiment of the present invention.

FIG. 6 is a configuration diagram showing a difference model in the search device for continuous speech recognition according to the first embodiment of the present invention.

FIG. 7 is a table showing an example of an inter-syllable string conversion likelihood table in the search device for continuous speech recognition according to the first embodiment of the present invention;

FIG. 8 is a configuration diagram showing an example of a difference model learning means in the search device for continuous speech recognition according to the first embodiment of the present invention;

FIG. 9 is a flowchart showing an operation procedure of a word string search unit in the search device for continuous speech recognition according to the first embodiment of the present invention.

FIG. 10 is a flowchart showing a procedure for collating an optimal syllable string with a standard syllable string of word n in the search device for continuous speech recognition according to the first embodiment of the present invention.

FIG. 11 is an explanatory diagram showing a collation operation in the search device for continuous speech recognition according to the first embodiment of the present invention.

FIG. 12 is an explanatory diagram showing a word dictionary in the search device for continuous speech recognition according to the second embodiment of the present invention.

FIG. 13 is a configuration diagram showing a difference model in the search device for continuous speech recognition according to the second embodiment of the present invention.

FIG. 14 is a table showing an example of a word syllable length conversion likelihood table in the search device for continuous speech recognition according to the second embodiment of the present invention.

FIG. 15 is a configuration diagram showing a search device for continuous speech recognition according to a third embodiment of the present invention.

FIG. 16 is a configuration diagram showing a search device for continuous speech recognition according to a fourth embodiment of the present invention.

FIG. 17 is a configuration diagram showing a search device for continuous speech recognition according to a fifth embodiment of the present invention.

FIG. 18 is a configuration diagram showing a search device for continuous speech recognition according to a sixth embodiment of the present invention.

FIG. 19 is a configuration diagram showing a search device for continuous speech recognition according to a seventh embodiment of the present invention.

FIG. 20 is a configuration diagram showing a search device for continuous speech recognition according to an eighth embodiment of the present invention.

FIG. 21 is a configuration diagram showing a search device for continuous speech recognition according to a ninth embodiment of the present invention.

FIG. 22 is a configuration diagram showing a difference model in the search device for continuous speech recognition according to the ninth embodiment of the present invention.

FIG. 23 is a table showing a word string word-to-word conversion table in the search device for continuous speech recognition according to the ninth embodiment of the present invention;

FIG. 24 is a flowchart showing a word string search procedure of the word string search means in the search device for continuous speech recognition according to the ninth embodiment of the present invention.

FIG. 25 is a flowchart showing a word and a partial word string of an optimal word string and a collation procedure in the search device for continuous speech recognition according to the ninth embodiment of the present invention;

[Description of Signs] 2 Optimum solution obtaining means, 4 Optimum syllable string, 5 Word string searching means, 6 Difference model, 7 Word dictionary, 8 Word string candidate, 13 Difference model applied word dictionary, 15 Difference model applied syllable graph creation Means, 21 N best solution obtaining means,
101 input speech, 102 speech analysis means, 601 syllable sequence conversion likelihood table, 602 word syllable length conversion likelihood table.

Claims (22)

[Claims] 1. An audio analysis means for analyzing an input audio, and an analysis result created by the audio analysis means is searched for an optimal solution at a first stage, and a second stage is obtained by modifying the optimal solution at the first stage. In a search device for continuous speech recognition performing a search, a difference model expressing a likelihood corresponding to an optimal solution obtained in the first step and a correct answer is provided, and the difference model is calculated from the optimal solution obtained in the first step. A search device for continuous speech recognition, wherein the search is performed in the second stage by applying the search. 2. Speech analysis means for analyzing an input speech, an optimum solution obtaining means for inputting an analysis result created by the speech analysis means and obtaining an optimum syllable sequence controlled by an automaton representing a connection between syllables, A difference model describing the likelihood that the optimal syllable string determined by the optimal solution obtaining means corresponds to a correct syllable string; a word dictionary describing a standard syllable string of words; Word string searching means for inputting an optimal syllable string, referring to the difference model and the word dictionary, searching for a candidate for the word string, and outputting a candidate for the word string. Item 2. A search device for continuous speech recognition according to Item 1. 3. The difference model is a syllable string conversion likelihood table that describes a partial syllable string of an optimal syllable string, a partial syllable string of a correct syllable string, and their corresponding likelihoods. 3. The search device for continuous speech recognition according to claim 2, wherein the search for a candidate word string is performed based on the likelihood described in the syllable string conversion likelihood table. 4. A difference model comprises: a syllable string conversion likelihood table describing a partial syllable string of an optimal syllable string, a partial syllable string of a correct syllable string, and their corresponding likelihoods; A word syllable length conversion likelihood table that describes the lengths of syllable strings in the word dictionary and the likelihoods corresponding to the word syllable strings; the word string search means includes the inter-syllable string conversion likelihood table and the word syllable length conversion likelihood table. 3. A candidate word string is searched for based on the likelihood described in (2).
A search device for continuous speech recognition according to any of the preceding claims. 5. Speech analysis means for analyzing an input speech, an optimum solution obtaining means for inputting the analysis result created by the speech analysis means and obtaining an optimum syllable string controlled by an automaton representing a connection between syllables, A difference model describing the likelihood that the optimal syllable sequence obtained by the optimal solution obtaining means corresponds to a correct syllable sequence, a word dictionary describing a standard syllable sequence of words, the difference model and the word dictionary A difference model-applied word dictionary that describes a word and a syllable graph obtained by transforming a standard syllable string of the word dictionary based on the description of the difference model for each word in the word dictionary; And a word string searching means for inputting the optimal syllable string, referring to the difference model applied word dictionary, searching for a candidate for the word string, and outputting a candidate for the word string. 1 continuous sound Searching device for recognition. 6. Speech analysis means for analyzing input speech, optimal solution obtaining means for inputting the analysis result created by the speech analysis means and obtaining an optimal syllable string controlled by an automaton representing connections between syllables, A difference model describing the likelihood that the optimum syllable string obtained by the optimum solution obtaining means corresponds to the correct syllable string; and the optimum syllable string based on the description of the difference model by inputting the optimum syllable string. Inputting the graph created by the difference model applied syllable graph creating means, and a word dictionary describing a standard syllable string of words; 2. The search apparatus for continuous speech recognition according to claim 1, further comprising: a word string search unit that searches for a word string candidate by referring to a dictionary and outputs a candidate word string. 7. Speech analysis means for analyzing an input speech, and an analysis result created by the speech analysis means is inputted, and a syllable string composed of upper N best syllables controlled by an automaton representing connections between syllables is obtained. A N-best solution obtaining means to be obtained, a difference model describing the likelihood that the optimal syllable string composed of the top N syllables obtained by the N-best solution obtaining means corresponds to a correct syllable string, A word dictionary that describes a syllable string and a syllable string composed of the optimal top N syllables obtained by the N best solution obtaining means are input, and the difference model and the word dictionary are referenced to find a word string candidate. 2. A search device for continuous speech recognition according to claim 1, further comprising: a word string search means for searching and outputting a word string candidate. 8. A voice analysis means for analyzing an input voice, and an analysis result created by the voice analysis means is input, and a syllable string composed of N optimal syllables controlled by an automaton representing connections between syllables is obtained. N-best solution obtaining means to be obtained, a difference model describing the likelihood that the optimal syllable string composed of N syllables obtained by the N-best solution obtaining means corresponds to a correct syllable string, and a standard syllable of a word A word dictionary describing a sequence, and a syllable graph obtained by transforming a standard syllable sequence of the word dictionary based on the description of the word and the difference model for each word of the word dictionary with reference to the difference model and the word dictionary. The described difference model application word dictionary and the optimal syllable string obtained by the N best solution obtaining means are input,
2. A search device for continuous speech recognition according to claim 1, further comprising: a word string search means for searching for a word string candidate with reference to the difference model applied word dictionary and outputting the word string candidate. . 9. Speech analysis means for analyzing an input speech, and an analysis result created by the speech analysis means is inputted, and an optimal syllable string composed of N syllables controlled by an automaton representing a connection between syllables is obtained. N best solution acquiring means, a difference model describing the likelihood that the optimal N syllable string obtained by the N best solution acquiring means corresponds to a correct syllable string, and the optimal syllable string are input. And a difference model-applied syllable graph creating means for creating a graph by transforming the optimal syllable string of N syllables based on the description of the difference model, and a word dictionary describing a standard syllable string of words And word string search means for inputting a graph created by the difference model applied syllable graph creation means, referring to the word dictionary, searching for word string candidates, and outputting word string candidates. Characterized by Motomeko 1 Continuous seeker for speech recognition according. 10. A word syllable length conversion likelihood table which describes an optimal syllable string length, a syllable string length of a word dictionary, and a likelihood corresponding to each other as a difference model. The search device for continuous speech recognition according to any one of claims 5 to 9, wherein a candidate word string is searched for based on the likelihood of the syllable length conversion likelihood table. 11. Speech analysis means for analyzing input speech, optimal solution acquisition means for inputting an analysis result created by the speech analysis means and obtaining an optimal word string controlled by an automaton representing a connection between words, A syllable string conversion means for converting the optimum word string obtained by the optimum solution obtaining means into a syllable string, and a difference model describing the likelihood that the syllable string obtained by the syllable string conversion means corresponds to a correct syllable string, A word dictionary describing a standard syllable string of a word, and an optimal syllable string obtained by the syllable string conversion means are input, and the difference model and the word dictionary are referred to, a word string candidate is searched, and a word is searched. 2. The search apparatus for continuous speech recognition according to claim 1, further comprising: a word string search unit that outputs a candidate string. 12. Speech analysis means for analyzing an input speech, an optimum solution obtaining means for inputting the analysis result created by the speech analysis means and obtaining an optimal word string controlled by an automaton representing a connection between words, A difference model describing the likelihood that the optimal word sequence obtained by the optimal solution obtaining means corresponds to the correct word sequence, a word dictionary describing the words, and the optimal word sequence obtained by the optimal solution obtaining device are input. 2. The continuous speech according to claim 1, further comprising: a word string search unit that searches for a candidate of a word string by referring to the difference model and the word dictionary, and outputs a candidate of the word string. Search device for recognition. 13. A word syllable length conversion likelihood table which describes an optimal word string length corresponding to a word in a word dictionary and its likelihood as a difference model. 13. The search device for continuous speech recognition according to claim 11, wherein a search for a candidate word string is performed based on the likelihood of the likelihood table. 14. A search method for continuous speech recognition in which an analysis result of an input speech is obtained in a first step and a search in a second step is performed by modifying the optimum solution in the first step. Providing a difference model expressing the likelihood between the optimal solution obtained in the first stage and the correct solution, and performing the second stage search by applying the difference model from the optimal solution obtained in the first stage A search method for continuous speech recognition characterized by the following. 15. A difference model in which an analysis result of an input speech is input, an optimal syllable sequence controlled by an automaton representing connections between syllables is obtained, and a likelihood corresponding to the optimal syllable sequence corresponding to a correct syllable sequence is described. 15. The search method for continuous speech recognition according to claim 14, wherein a candidate word string is searched for with reference to a word dictionary describing a standard syllable string of words and a candidate word string is output. 16. A difference model in which an analysis result of an input speech is input, an optimal syllable sequence controlled by an automaton representing a connection between syllables is obtained, and a likelihood corresponding to the optimal syllable sequence corresponding to a correct syllable sequence is described. And a word dictionary describing a standard syllable string of the word, and for each word of the word dictionary, a syllable graph obtained by transforming the standard syllable string of the word dictionary based on the description of the word and the difference model. The search method for continuous speech recognition according to claim 14, wherein a candidate word sequence is searched for by referring to a difference model application word dictionary describing the word sequence and the candidate word sequence is output. 17. A difference model in which an analysis result of an input speech is input, an optimal syllable sequence controlled by an automaton representing a connection between syllables is obtained, and a likelihood corresponding to the optimal syllable sequence corresponding to a correct syllable sequence is described. Based on the description, a graph is created by transforming the above optimal syllable sequence, and the created graph is input, and a word dictionary describing standard syllable sequences of words is searched for word sequence candidates. 15. The search method for continuous speech recognition according to claim 14, wherein a candidate word string is output. 18. An analysis result of an input voice is input, and an optimal syllable string composed of upper N syllables controlled by an automaton representing a connection between syllables is obtained, and a syllable composed of these optimal upper N syllables is obtained. The sequence is input, and a difference model that describes the likelihood that the optimal syllable sequence consisting of the top N syllables corresponds to the correct syllable sequence and a word dictionary that describes a standard syllable sequence of words are referred to. 15. The search method for continuous speech recognition according to claim 14, further comprising searching for candidate strings and outputting candidate word strings. 19. An analysis result of an input voice is input, a syllable string composed of N optimum syllables controlled by an automaton representing a connection between syllables is obtained, and this optimum syllable string is input. Reference is made to a difference model describing the likelihood that a syllable string composed of syllables corresponds to a correct syllable string and a word dictionary describing a standard syllable string of words. Referring to a difference model applied word dictionary describing a syllable graph obtained by transforming a standard syllable string of the word dictionary based on the model description, searching for a word string candidate, and outputting a word string candidate. Claim 1.
5. The search method for continuous speech recognition according to 4. 20. An analysis result of an input speech is input, and an optimal syllable string composed of N syllables controlled by an automaton representing connection between syllables is obtained. A graph is created by transforming an optimal syllable string composed of N syllables based on the description of the difference model that describes the likelihood corresponding to the syllable string of the syllable string. 15. The search method for continuous speech recognition according to claim 14, wherein a candidate word string is searched for by referring to a word dictionary describing a syllable string, and the candidate word string is output. 21. An analysis result of an input voice is inputted, an optimum word string controlled by an automaton representing a connection between words is obtained, and the optimum word string is converted into a syllable string, and the syllable string is a correct syllable string. And searching for a candidate word string and outputting a candidate word string by referring to a difference model describing the likelihood corresponding to the word and a word dictionary describing a standard syllable string of the word. 15. The search method for continuous speech recognition according to 14. 22. A difference model in which an analysis result of an input speech is input, an optimum word string controlled by an automaton representing a connection between words is obtained, and a likelihood that the optimum word string corresponds to a correct word string is described. 15. The search method for continuous speech recognition according to claim 14, further comprising: searching for a candidate word string and outputting the candidate word string with reference to the word dictionary describing the word.