JPH11119793A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically set a threshold and the number of target active buses of cumulative likelihood for pruning by determining a threshold for pruning by a beam search method used as a speech recognition means from detected perplexity. SOLUTION: A speech input means 102 detects a speech section from time waveform, and divides the detected speech section into frames, and parameterize the frame-divided speech for recognition. A speech recognition means 104 compares standard patterns read from a dictionary with a signal parameterized in the speech input means 102 by using the beam search method, and outputs the recognition result as a character string, etc. A perplexity detecting means 105 reads out a dictionary stored in a storage means 103, and detects perplexity F which is a parameter showing recognition task complexity. A threshold value setting means 106 decides a threshold value α of pruning according to the beam search method carried out in the speech recognition means 104 from the detected perplexity.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech such as a word uttered by a human being as an input signal, comparing the speech with a standard pattern, searching for the most similar pattern, and recognizing the result. The present invention relates to a voice recognition device that outputs as "?".

[0002]

2. Description of the Related Art A conventional technique will be described with reference to FIG.

[0003] A conventional speech recognition apparatus is a speech recognition apparatus 601 for recognizing input speech and outputting a result.
The apparatus includes a voice input unit 602, a storage unit 603, a voice recognition unit 604, and a threshold setting unit 605.

The voice input means 602 detects a voice section from the time waveform, divides the detected voice section into frames,
Parameterize the framed speech for recognition.

[0005] The storage means 603 stores in advance a standardized voice standard pattern as a dictionary.

[0006] The speech recognition means 604 outputs a recognition result of a signal parameterized by the speech input means and a standard pattern read from the dictionary as a character string or the like by a DP matching method or the like. Here, the speech recognition means 604 performs speech recognition by combining the DP matching method and the like with a beam search method in order to reduce the amount of calculation.

The threshold setting means 605 sends a threshold α for pruning by the beam search method performed by the speech recognition means to the speech recognition means 604.

The speech recognition means 604 solves the optimal path search problem of comparing the input speech parameterized with the standard pattern from the first frame and calculating the cumulative likelihood. The pruning by the beam search method performed by the voice recognition unit 604 means that, in the optimal path search problem, among the paths calculated from the first frame, those having a low cumulative likelihood are stopped in the middle. It is a method. This is performed to reduce the amount of calculation in the voice recognition means. Also, this is called pruning, because it is an operation of pruning branches that increase as the frame progresses.

Here, in order to perform pruning, a threshold value of the cumulative likelihood of the path to be pruned is required. In the conventional speech recognition apparatus 601, a threshold value must be given from outside or left from outside. I needed to give the number of passes. The number of paths to be left is called a target number of active paths.

With the speech recognition apparatus 601 having the above-described configuration, it is possible to recognize the input speech and output the result as a character string or the like.

[0011]

In the conventional speech recognition apparatus, when setting a pruning threshold value by the beam search method, a threshold value of the cumulative likelihood is given from the outside, or
It was necessary to give the target number of active paths from outside.

However, the threshold value of the cumulative likelihood and the number of target active paths depend on the number of vocabularies of the recognition task, the similarity of the vocabulary registered in the dictionary, the accuracy of the standard pattern, and the like. Therefore, in order to set the threshold value of the cumulative likelihood for pruning and the number of target active paths, the relationship between the recognition task and the recognition rate must be checked in advance for each recognition task. This was a very laborious task.

The present invention has been made in view of the above problems, and has as its object to automatically set a threshold value of a cumulative likelihood for pruning and a target number of active paths.

[0014]

In order to solve the above-mentioned problems, a voice recognition device according to the present invention is a voice recognition device for recognizing an input voice and outputting a recognition result. A voice input means for detecting a voice section from the time waveform of the voice, dividing the detected voice section into frames, and converting the frame-divided voice into parameters for recognition, and a standard pattern of voice prepared in advance. Is stored as a dictionary, a signal converted into a parameter by the voice input means, and a standard pattern read from the dictionary are compared using a beam search method,
Voice recognition means for outputting a recognition result, perplexity detection means for reading a dictionary stored in the storage means and detecting perplexity representing complexity of a recognition task performed by the voice recognition means; Threshold value setting means for determining a pruning threshold value by the beam search method performed by the voice recognition means based on the perplexity detected by the plexity detection means.

The voice recognition device of the present invention is a voice recognition device for recognizing an input voice and outputting a recognition result. The voice recognition device receives a voice and detects a voice section from a time waveform of the voice. Voice input means for dividing a detected voice section into frames and parameterizing the frame-divided voice for recognition, storage means for storing a standard pattern of voice prepared in advance as a dictionary, and voice input means The signal parameterized in the above and the standard pattern read from the dictionary are compared using the beam search method, and the speech recognition means for outputting the recognition result, and the frame number currently being processed by the speech recognition means are input. A dictionary representing the complexity of a certain frame section of the recognition task performed by the speech recognition means by reading the dictionary stored in the storage means. Interval perplexity detecting means for detecting lexity, and threshold setting means for determining a pruning threshold by a beam search method performed by the voice recognition means from the perplexity detected by the section perplexity detecting means, It is characterized by having.

The voice recognition device of the present invention is a voice recognition device for recognizing an input voice and outputting a recognition result. The voice recognition device receives a voice and detects a voice section from a time waveform of the voice. Voice input means for dividing a detected voice section into frames and parameterizing the frame-divided voice for recognition, storage means for storing a standard pattern of voice prepared in advance as a dictionary, and voice input means The parameterized signal and the standard pattern read from the dictionary are compared using a beam search method, and a speech recognition unit that outputs a recognition result, and a dictionary stored in the storage unit is read, and the speech is read. A perplexity detecting means for detecting perplexity representing the complexity of the recognition task performed by the recognition means; Active path number detecting means for detecting the number of active paths, the perplexity output from the perplexity detecting means, and the number of active paths output from the real active path number detecting means. A threshold setting unit that determines a pruning threshold value in the beam search method and outputs the pruning threshold value to the voice recognition unit.

The voice recognition device of the present invention is a voice recognition device for recognizing an input voice and outputting a recognition result. The voice recognition device receives a voice and detects a voice section from a time waveform of the voice. Voice input means for dividing a detected voice section into frames and parameterizing the frame-divided voice for recognition, storage means for storing a standard pattern of voice prepared in advance as a dictionary, and voice input means The parameterized signal and the standard pattern read from the dictionary are compared using a beam search method, and a speech recognition unit that outputs a recognition result, and a dictionary stored in the storage unit is read, and the speech is read. A perplexity detecting means for detecting a perplexity representing the complexity of the recognition task performed by the recognizing means; A frame number is input, a dictionary stored in the storage unit is read, and an active path number calculation unit that calculates a calculated active path number in the frame number of the recognition task is output from the perplexity detection unit. Perplexity and the calculated number of active paths output from the number of active paths calculating means are input, determine the target number of active paths in the beam search method in the speech recognition device, and output.
Target active path number setting means, actual active path number detecting means for detecting the actual number of active paths in the speech recognition device, target active path number output from the target active path number setting means, actual active path number When the number of active paths output from the detection means is input and the number of active paths exceeds the target number of active paths, the pruning threshold is updated and a new pruning threshold is set to the speech recognition device. And output threshold setting means.

The voice recognition device of the present invention is a voice recognition device for recognizing an input voice and outputting a recognition result. The voice recognition device receives a voice and detects a voice section from a time waveform of the voice. Voice input means for dividing a detected voice section into frames and parameterizing the frame-divided voice for recognition, storage means for storing a standard pattern of voice prepared in advance as a dictionary, and voice input means The parameterized signal and the standard pattern read from the dictionary are compared using a beam search method, and a speech recognition unit that outputs a recognition result, and a dictionary stored in the storage unit is read, and the speech is read. A perplexity detecting means for detecting a perplexity representing the complexity of the recognition task performed by the recognizing means; A frame number is input, a dictionary stored in the storage unit is read, and an active path number calculation unit that calculates a calculated active path number in the frame number of the recognition task is output from the perplexity detection unit. The perplexity and the calculated number of active paths output from the number of active paths calculating means are input, and an appropriate target number of active paths in a beam search method in the voice recognition device is determined. A target active path number setting means for outputting the target active path number to the voice recognition means; a real active path number detecting means for detecting the actual number of active paths in the voice recognition means; and the real active path number detecting means And the target number of active paths output from the When the number of the active paths is larger than the target number of active paths, the comparison unit that outputs a threshold update command, and the speech recognition unit calculates the cumulative likelihood of the active path in a certain frame. An average value variance value detecting means for detecting an average value and a variance value, an average value and a variance value output from the average value variance value detecting means, and a threshold update command output from the comparing means are input A threshold setting unit that updates a pruning threshold of the voice recognition unit.

[0019]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below with reference to FIGS.

Embodiment 1 Embodiment 1 of the present invention will be described with reference to FIG.

The speech recognition apparatus 101 according to the first embodiment includes a speech input unit 102, a storage unit 103, a speech recognition unit 104, a perplexity detection unit 105, and a first threshold setting unit 106. ing.

The voice input means 102 detects a voice section from the time waveform, divides the detected voice section into frames,
Parameterize the framed speech for recognition.

The storage means 103 stores a standard pattern of voice prepared in advance as a dictionary.

The voice recognition means 104 compares the signal parameterized by the voice input means with the standard pattern read from the dictionary while using a beam search method.
Output the recognition result as a character string.

The perplexity detecting means 105 reads the dictionary stored in the storage means 103 and detects perplexity F which is a parameter representing the complexity of the recognition task. Here, the perplexity F is a constant. The perplexity detecting means 105 detects the perplexity of the entire recognition task or the perplexity of the initial frame portion of the recognition task.

The first threshold setting means 106 determines a threshold α for pruning by the beam search method performed by the voice recognition means 104 from the perplexity detected by the perplexity detection means 105. Since the perplexity is a parameter representing the complexity of the recognition task, the threshold α of the cumulative likelihood is small when the perplexity is large, and conversely, the threshold of the cumulative likelihood is small when the perplexity is small. α may be increased.

In many recognition tasks, there is a tendency that the perplexity in the initial frame portion is larger than the entire perplexity. Therefore, when the perplexity detecting means 105 detects the perplexity of the initial frame portion of the recognition task and sets the pruning threshold value, the pruning threshold value is set too narrow and the optimal path is also pruned. The danger of losing is reduced.

With the above configuration, the speech recognition apparatus of the present invention can automatically set the pruning threshold α using the perplexity F of the recognition task.

(Embodiment 2) Embodiment 2 of the present invention will be described with reference to FIG.

The speech recognition apparatus 201 according to the second embodiment includes a speech input unit 202, a storage unit 203, a speech recognition unit 204, a section perplexity detection unit 205, and a second threshold setting unit 206. doing.

The configuration of the second embodiment is almost the same as the configuration of the first embodiment, except that a section perplexity detecting means 205 is provided instead of the perplexity detecting means 105. That is.

The section perplexity detecting means 205
The frame number t currently processed by the voice recognition means
Is input, and the dictionary stored in the storage unit 203 is read. Here, the perplexity Fb of a certain limited section, for example, an initial frame portion, an intermediate frame portion, a late frame portion, etc., is detected while referring to the dictionary of the recognition task.

In general, the perplexity varies depending on the initial frame portion, intermediate frame portion, late frame portion, etc. of the vocabulary registered in the dictionary.

According to the section perplexity detecting means of the present invention, for example, it is possible to detect the perplexity Fb corresponding to each of the initial, intermediate, and late sections of the input voice, and set the pruning thresholds, respectively. It becomes possible. This allows
In the beam search method, the redundancy of calculating the cumulative likelihood of unnecessary paths and the risk of pruning remaining paths are improved.

(Embodiment 3) A speech recognition apparatus 301 according to Embodiment 3 of the present invention will be described with reference to FIG.

Voice input means 302 and storage means 303
The voice recognition means 304 and the perplexity detection means 305 are the same as in the first embodiment. The third embodiment is different from the third embodiment in that the third embodiment includes a real active path number detecting unit 306 and a third threshold setting unit 307.

The actual number of active paths detecting means 306 detects the number n (t) of active paths in the beam search method performed by the voice recognizing means 304 for each frame, and outputs the number to the third threshold setting means 307. send.

The threshold setting means 307 outputs the perplexity F output from the perplexity detecting means 305.
And the number n (t) of active paths output from the actual number-of-active-paths detecting means 306 are input, and the threshold α (t) for pruning active paths in the beam search method in the speech recognition device 304 is determined. I do.

The speech recognition apparatus 301 of this embodiment is characterized in that a perplexity F and the number of active paths n (t) are used in setting the pruning threshold value α (t). As a result, compared to the case where the pruning threshold is set only by the perplexity F, the pruning threshold α (t)
Can be automatically set to a more accurate value. That is, since the number n (t) of active paths that are actually being processed is not known, pruning is insufficient and unnecessary paths are left uncalculated. The risk of reaping is reduced.

(Embodiment 4) A speech recognition apparatus 401 according to Embodiment 4 of the present invention will be described with reference to FIG.

Voice input means 402 and storage means 403
The voice recognition unit 404, the perplexity detection unit 405, and the actual number of active paths detection unit 408 are the same as those of the third embodiment. The speech recognition apparatus 401 according to the fourth embodiment further includes an active path number calculating unit 406, a target active path number setting unit 407, and a fourth fourth threshold value setting unit 409 connected to the storage unit 403. I have it.

The number of active paths calculating means 406 receives the frame number t currently being processed from the voice recognizing means 404 and reads the standard pattern stored in the storing means 403. Here, if pruning is not performed by the voice recognition unit 404, the number of active paths M (t) for which the calculation of the accumulated likelihood is required is calculated at the frame number t.

The target number of active paths setting means 407 is
The perplexity F output from the perplexity detecting means 405 and the number of active paths calculating means 4
The calculated active path number M (t) output from step 06 is input, the target active path number N (t) suitable for the current recognition task is set, and output to the fourth threshold setting means 409. Here, the perplexity F is a constant, and the number M of calculated active paths is a value that changes for each frame.

The fourth threshold value setting means 409 calculates the actual number of active paths n (t) output from the actual number of active paths detecting means 408 and the target active path number output from the target number of active paths setting means 407. Number N
(T) and the actual number of active paths n (t)
Is larger than the target number of active paths N (t),
The pruning threshold α (t) is updated.

According to the present invention, when setting the pruning threshold value α (t), the perplexity F of the recognition task and the number of calculated active paths M (t) calculated from the dictionary of the standard pattern of the recognition task. And the target number of active paths N
(T) is set, and the pruning threshold α (t) is updated while comparing the target number of active paths N (t) with the actual number of active paths n (t) for each frame. An appropriate threshold can be automatically set.

As a result, in the beam search method, the redundancy that an unnecessary path is calculated due to insufficient pruning and the danger that the optimum path is pruned after excessive pruning are reduced. You.

(Fifth Embodiment) A speech recognition apparatus 501 according to a fifth embodiment of the present invention will be described with reference to FIG.

Voice input means 502 and storage means 503
The voice recognition means 504, the perplexity detection means 505, the active path number calculation means 506, the target active path number setting means 507, and the actual active path detection means 508 are the same as those of the fourth embodiment. It is.

The speech recognition apparatus 501 of the fifth embodiment further outputs the actual number of active paths n (t) output from the actual number of active paths detecting means 508 and the output from the target number of active paths setting means 507. The comparison unit 509 receives the target number N (t) of active paths.
And the average value A (t) and the variance value B of the cumulative likelihood, which indicate the degree of variation of the cumulative likelihood of all the active paths calculated in a certain frame by the voice recognition means 504.
Mean value variance detecting means 510 for detecting (t) and the comparator 509 output when the actual number of active paths n (t) exceeds the target number of active paths N (t). A threshold update command and an average value A of the accumulated likelihood output from the average value variance value detection means 510
And a fifth threshold value setting unit 511 to which (t) and the variance value B (t) are input.

The average value variance value detecting means 510 calculates the average value A (t) and the variance value B of the cumulative likelihood of the active path calculated by the speech recognition means 504 for each frame.
(T) is detected and output to the fifth threshold value setting means 511.

The fifth threshold value setting means 511 receives the threshold update command from the comparing means 509 and outputs the average value A (t) of the cumulative likelihood of the active path output from the average value variance value detecting means 510. A new threshold value α (t) is set based on the variance value B (t) and the current threshold value, and is sent to the voice recognition unit 504.

According to the present invention, the target active path number N (t) calculated from the perplexity F of the recognition task and the calculated active path number M (t), and the average value A (t) of the cumulative likelihood of the active paths ) And the variance value B (t), the pruning threshold value α (t) is obtained, so that a more accurate threshold value can be automatically set as compared with the above embodiment. .

As a result, in the beam search method, there is reduced the redundancy that an unnecessary path is calculated due to insufficient pruning, and the danger that the optimum path is pruned due to excessive pruning is reduced. You.

[0054]

As described above, according to the present invention, it is possible to automatically set the threshold value of the cumulative likelihood for pruning and the target number of active paths.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a speech recognition device according to a first embodiment of the present invention.

FIG. 2 is a block diagram illustrating a speech recognition device according to a second embodiment of the present invention.

FIG. 3 is a block diagram illustrating a speech recognition device according to a third embodiment of the present invention.

FIG. 4 is a block diagram showing a voice recognition device according to a fourth embodiment of the present invention.

FIG. 5 is a block diagram showing a speech recognition apparatus according to a fifth embodiment of the present invention.

FIG. 6 is a block diagram showing a conventional speech recognition apparatus.

[Explanation of symbols]

 Reference Signs List 101 voice recognition device 102 voice input means 103 storage means 104 voice recognition means 105 perplexity detection means 106 first threshold setting means 201 voice recognition device 202 voice input means 203 storage means 204 voice recognition means 205 section perplexity detection means 206 Second threshold setting means 301 Speech recognition device 302 Speech input means 303 Storage means 304 Speech recognition means 305 Perplexity detection means 306 Actual active path number detection means 307 Third threshold setting means 401 Speech recognition device 402 Speech input means 403 Storage means 404 Voice recognition means 405 Perplexity detection means 406 Active path number calculation means 407 Target active path number setting means 408 Real active path number detection means 409 Fourth threshold value setting means 501 Sound Recognition device 502 Speech input means 503 Storage means 504 Speech recognition means 505 Perplexity detection means 506 Active path number calculation means 507 Target active path number setting means 508 Real active path number detection means 509 Comparison means 510 Average value variance value detection means 511 Fifth threshold setting means

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Yoshihisa Nakato 1006 Kazuma Kadoma, Kadoma City, Osaka Matsushita Electric Industrial Co., Ltd.

Claims (6)

[Claims] 1. A speech recognition apparatus for recognizing an input speech and outputting a recognition result, wherein the speech is input, a speech section is detected from a time waveform of the speech, and the detected speech section is divided into frames. A voice input means for converting the frame-divided voice into parameters for recognition; a storage means for storing a standard pattern of voice prepared in advance as a dictionary; and a signal converted into parameters by the voice input means. When,
A voice recognition unit that compares a standard pattern read from the dictionary with a beam search method and outputs a recognition result, and reads a dictionary stored in the storage unit and performs a recognition task performed by the voice recognition unit. A perplexity detecting means for detecting a perplexity representing the depth, and a threshold setting for determining a pruning threshold by a beam search method performed by the voice recognition means from the perplexity detected by the perplexity detecting means. And a means. 2. The speech recognition apparatus according to claim 1, wherein said perplexity detection means detects a perplexity of an initial frame portion of said input speech. 3. A speech recognition apparatus for recognizing an inputted speech and outputting a recognition result, wherein the speech is inputted, a speech section is detected from a time waveform of the speech, and the detected speech section is divided into frames. A voice input unit for parameterizing the frame-divided voice for recognition; a storage unit for storing a standard pattern of voice prepared in advance as a dictionary; a signal parameterized by the voice input unit; The standard pattern read from the dictionary is compared using a beam search method, and a voice recognition unit that outputs a recognition result is input with a frame number that is currently being processed from the voice recognition unit, and is stored in the storage unit. A section for reading a stored dictionary and detecting a perplexity representing the complexity of a certain frame section of a recognition task performed by the voice recognition means. -Plexity detecting means, and threshold setting means for determining a pruning threshold by a beam search method performed by the voice recognition means from the perplexity detected by the section perplexity detecting means. Voice recognition device. 4. A speech recognition device for recognizing an inputted speech and outputting a recognition result, wherein the speech is inputted, a speech section is detected from a time waveform of the speech, and the detected speech section is divided into frames. A voice input unit for parameterizing the frame-divided voice for recognition; a storage unit for storing a standard pattern of voice prepared in advance as a dictionary; a signal parameterized by the voice input unit; A voice recognition unit that compares a standard pattern read from the dictionary with a beam search method and outputs a recognition result, and reads a dictionary stored in the storage unit and performs a recognition task performed by the voice recognition unit. A perplexity detecting means for detecting a perplexity representing the length of the path, and detecting the number of actually active paths in the voice recognition means. Active path number detecting means, perplexity output from the perplexity detecting means, and the number of active paths output from the actual active path number detecting means are input, and pruning in the beam search method is performed. And a threshold setting unit that determines the threshold value of pruning and outputs the pruning threshold value to the voice recognition unit. 5. A speech recognition apparatus for recognizing inputted speech and outputting a recognition result, comprising the steps of receiving speech, detecting a speech section from a time waveform of the speech, and dividing the detected speech section into frames. A voice input unit for parameterizing the frame-divided voice for recognition; a storage unit for storing a standard pattern of voice prepared in advance as a dictionary; a signal parameterized by the voice input unit; A voice recognition unit that compares a standard pattern read from the dictionary with a beam search method and outputs a recognition result, and reads a dictionary stored in the storage unit and performs a recognition task performed by the voice recognition unit. A perplexity detecting means for detecting a perplexity representing the frame number, and a frame number which is currently being processed is inputted from the voice recognition means. An active path number calculating means for reading a dictionary stored in a storage means and calculating a calculated active path number in the frame number of the recognition task; a perplexity output from the perplexity detecting means; Target active path number setting means, to which the calculated active path number output from the path number calculating means is input, to determine and output the target active path number in the beam search method in the speech recognition apparatus, and the speech recognition apparatus The actual active path number detecting means for detecting the actual number of active paths in, the target active path number output from the target active path number setting means, and the active path number output from the real active path number detecting means are input. , The number of active paths exceeds the target number of active paths In this case,
And a threshold setting unit that updates a pruning threshold value and outputs a new pruning threshold value to the speech recognition device. 6. A speech recognition apparatus for recognizing an input speech and outputting a recognition result, comprising the steps of: inputting a speech, detecting a speech section from a time waveform of the speech, and dividing the detected speech section into frames. A voice input unit for parameterizing the frame-divided voice for recognition; a storage unit for storing a standard pattern of voice prepared in advance as a dictionary; a signal parameterized by the voice input unit; A voice recognition unit that compares a standard pattern read from the dictionary with a beam search method and outputs a recognition result, and reads a dictionary stored in the storage unit and performs a recognition task performed by the voice recognition unit. A perplexity detecting means for detecting a perplexity representing the frame number, and a frame number which is currently being processed is inputted from the voice recognition means. An active path number calculating means for reading a dictionary stored in a storage means and calculating a calculated active path number in the frame number of the recognition task; a perplexity output from the perplexity detecting means; The calculated active path number output from the path number calculation means is input, and an appropriate target active path number in the beam search method in the speech recognition device is determined. Target active path number setting means for outputting the number of active paths, actual active path number detecting means for detecting the number of actual active paths in the voice recognition means, and the number of active paths output from the actual active path number detecting means And the target number of active paths are input, and the When the number is larger than the target number of active paths, a comparison unit that outputs a threshold update command; and the average value and variance of the cumulative likelihood of the active path calculated in a certain frame in the voice recognition unit. Average value variance value detection means for detecting the average value and variance value output from the average value variance value detection means, and when the threshold update command output from the comparison means is input, the voice A threshold setting unit that updates a pruning threshold of the recognition unit.