JP3473704B2 - Voice recognition device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus suitable for recognizing speech.

[0002]

2. Description of the Related Art Conventional speech recognition devices are roughly classified into
From a voice uttered by a speaker, for example, a speech recognition unit that word-spots a word (vocabulary), and a word candidate string obtained as a result of word spotting in the speech recognition unit,
Using the prepared syntax information, for example,
It is composed of a parsing unit for parsing sentence by sentence.

In the speech recognition apparatus having such a configuration, when the analysis unit requests the speech recognition unit for a word to be spotted, the speech recognition unit performs word spotting on the voice uttered by the speaker.
It is output to the analysis unit. In the analysis unit, the word candidate string obtained as a result of the word spotting in the speech recognition unit is syntactically analyzed in sentence units using the syntax information.

Then, based on the result of the syntactic analysis, a correct sentence (sentence) is obtained from the word candidate string as the word spotting result by removing the incorrect word candidate.

[0005]

By the way, in the conventional voice recognition apparatus, when the number of words for which the word spotting processing is requested by the voice recognition unit is larger than that by the analysis unit, the voice recognition unit is required to finish the speech by the speaker. , There were times when I couldn't do the spotting process for all that word.

Therefore, in this case, there is a problem that the response to the input voice is delayed.

Therefore, there is a method in which the speech recognition unit suspends the spotting processing of some of the words for which the word spotting processing request has been made by the analysis unit.

However, according to this method, the spotting process of words required by the analysis unit is not performed, and the response to the input voice may be delayed.

The present invention has been made in view of such a situation, and improves the response speed of the apparatus.

[0010]

A speech recognition apparatus according to the present invention requests a word spotting processing section 4 as a recognition means for recognizing a speech and the word spotting processing section 4 to perform speech recognition processing of a word in a speech. In addition, the word spotting processing unit 4 is provided with a syntactic analysis unit 5 as an analyzing means for analyzing the recognition result and understanding the speech, and the syntactic analysis unit 5 adds a priority of speech recognition processing to the word, The word spotting processing unit 4 supplies the word spotting processing unit 4, and the word spotting processing unit 4 supplies the word
Based on word priority or your own speech recognition throughput.
Modify Zui, based on the word priority, and performs a speech recognition process words from in the speech.

The speech recognition apparatus can further include a speech section detection unit 3 as a detection means for detecting the speech section of the speech, and the word spotting processing unit 4 can detect the speech section during and after the speech section. Based on the priority of the word, the word to be recognized in the voice can be changed.

Further, this speech recognition apparatus can cause the syntax analysis unit 5 to change the priority of words based on the analysis result of the recognition result of the word spotting processing unit 4.

[0013]

[0014]

In the speech recognition apparatus having the above-described structure, the word spotting processing unit 4 causes the words of the words to be speech-recognized
The number is the priority of the word added by the syntax analysis unit 5 or
Changes based on its own speech recognition processing capability, and
Based on the priority of the word, the speech recognition process of the word in the voice is performed. Therefore, it is possible to speed up the response process to the input voice.

In the word spotting processing unit 4, the word spotting processing unit 4 is based on the priority of the word during the voice section and after the end of the voice section.
In the case where the word to be recognized by voice can be changed from the voice, for example, after the end of the voice section, it is possible to perform the voice recognition process only for the word with the highest priority. Can be improved.

In the case where the syntactic analysis unit 5 can change the priority of the word based on the analysis result of the recognition result of the word spotting processing unit 4, the response processing to the inputted voice is further speeded up. Can be planned.

[0017]

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram showing the configuration of an embodiment of an external device controller to which the voice recognition device of the present invention is applied. In this external device controller, an external device (not shown) such as an AV device connected to the external device operating unit 6 can be operated by voice.

That is, the voice input unit 1 converts the input voice into a voice signal as an electric signal, further A / D-converts it, and outputs it to the voice analysis unit 2. The voice analysis unit 2 extracts, for each frame, voice feature parameters such as a linear prediction coefficient from the voice signal from the voice input unit 1 and outputs them to the voice section detection unit 3 and the word spotting processing unit 4 in time series. To do.

The voice section detection unit 3 detects the frame at which the utterance started and the frame at which the utterance ended, that is, the voice section, based on the characteristic parameter from the voice analysis unit 2,
It is output to the word spotting processing unit 4 and the syntax analysis unit 5.

When the speech section detection unit 3 detects a frame at which speech is started, the word spotting processing unit 4 outputs each word output from the syntax analysis unit 5 together with the priority of the word to the input speech. The word spotting is sequentially performed based on the priority of the word, and the word spotting result is output to the syntax analysis unit 5.

That is, the word spotting processing unit 4 is
When the speech section detection unit 3 detects a frame at which speech is started, first, the characteristic parameters of the speech from the speech analysis unit 2 are sequentially stored in a built-in input buffer (not shown) and stored in the input buffer. Set the characteristic parameter of the voice
Read frame by frame. Then, the word spotting processing unit 4 sets the standard pattern of each word output by the syntax analysis unit 5 together with the priority of the word as the input pattern sequentially read from the built-in input buffer based on the priority of each word. Match the characteristic parameters,
When the score obtained as a result is equal to or higher than a predetermined threshold value, the score and the start point and end point of the matched section are output to the syntax analysis unit 5.

Here, the score means the likelihood of the spotted word, and the larger the value, the more likely the spotted word is.

The syntactic analysis unit 5 establishes a hypothesis (sentence hypothesis) of the input voice before the speech section detection unit 3 detects the frame in which the utterance has started, and at the initial stage of the analysis of the voice. Add the highest priority to all required words,
It outputs to the word spotting processing unit 4 and requests the word spotting processing.

Furthermore, after the speech section detection unit 3 detects the frame in which the utterance has started, the syntax analysis unit 5 causes the word spotting processing unit 4 to make a score, a start point, and a score for each word for which the word spotting processing is requested. When the end point is supplied, the input voice is analyzed based on the score or the section length from the start point to the end point and the syntax information registered in advance in the built-in syntax dictionary (not shown), and the analysis is performed. Create a new text hypothesis based on the results.

Then, the syntactic analysis unit 5 adds priorities to the words necessary for analyzing the new sentence hypothesis,
It is output to the word spotting processing unit 4.

In this case, the priority of the word is added to the word based on the method of searching (selecting) the sentence hypothesis in the syntax analysis unit 5. That is, in the syntax analysis unit 5,
For example, when syntactic analysis is performed by the best-first method, a higher priority is added to a word associated with a sentence hypothesis having a higher certainty factor.

After the speech section detection unit 3 detects the frame in which the utterance has ended, the syntactic analysis unit 5 makes an erroneous (incorrectly) error from the word candidate sequence as the word spotting result based on its own syntactic analysis result. When a correct sentence (sentence) is obtained excluding the word candidates that are considered to be present, it is output to the external device operation unit 6.

The external device operation unit 6 operates the external device connected thereto in accordance with the content of the sentence (sentence) output from the syntax analysis unit 5.

Next, the operation will be described. First, in the voice input unit 1, the input voice is converted into a voice signal as an electric signal and A / D converted, and the voice analysis unit 2
Is output to. In the voice analysis unit 2, the voice feature parameter for each frame is extracted from the voice signal from the voice input unit 1 and output to the voice section detection unit 3 and the word spotting processing unit 4.

The voice section detection unit 3 detects the frame in which the utterance has started and the frame in which the utterance has ended, that is, the voice section, based on the characteristic parameters of the voice from the voice analysis unit 2, and the word spotting processing unit 4 and It is output to the syntax analysis unit 5.

At the same time, the syntactic analysis unit 5 sets a sentence hypothesis of the input voice before the speech section detection unit 3 detects the frame in which the speech is started, and at the initial stage of the analysis of the voice. The highest priority is added to all necessary words, and the words are output to the word spotting processing unit 4.

In the word spotting processing unit 4, the word spotting of each word output from the syntax analysis unit 5 together with the priority of the word is performed on the input voice, for example, according to the flowchart shown in FIG. .

That is, in the word spotting processing unit 4, first in step S1, a variable indicating the number of all the words output from the syntax analyzing unit 5 (the word spotting processing request (search request) has been issued from the syntax analyzing unit 5). W
Is initialized by setting 0 to step S
Go to 2. In step S2, the number of all the words output from the syntax analysis unit 5 (the word spotting processing request (search request) from the syntax analysis unit 5) is set in the variable W, and the process proceeds to step S3, and the voice section detection unit In 3, it is determined whether or not the frame in which speech has started is detected.

In step S3, the voice section detector 3
When it is determined that the frame in which the utterance is started is not detected in step S2, the process returns to step S2, and the steps in step S2 and step S3 are performed until it is determined in step S3 that the frame in which the utterance is started is detected. The process of S3 is repeated.

In step S3, the voice section detector 3
When it is determined that the frame in which the utterance is started is detected, that is, when the input of the voice to the voice input unit 1 is started, the process proceeds to step S4, and the word spotting processing unit 4 sets the number N of words to be word-spotted. In the variable N shown, the value stored in the variable W (the word requested to be processed by the word spotting processing unit 4 from the syntactic analysis unit 5 before the frame in which speech is started is detected by the voice section detection unit 3) (Several W) is set, and the process proceeds to step S5.

In step S5, the voice feature parameters output from the voice analysis unit 2 and already stored in the input buffer incorporated in the word spotting processing unit 4 are read out for one frame, and the process proceeds to step S6.
It is determined whether or not data (characteristic parameter of the voice output from the voice analysis unit 2) remains in the input buffer built in the word spotting processing unit 4. Step S
When it is determined that the data remains in the input buffer incorporated in the word spotting processing unit 4 in 6,
In step S7, the word spotting processing unit 4 decrements the variable N indicating the number N of words to be subjected to word spotting by 1, and then proceeds to step S8.

On the other hand, if it is determined in step S6 that no data remains in the input buffer built in the word spotting processing section 4, then step S7 is skipped and step S8 follows. Among the standard patterns of the words output together with the priority, the standard patterns of N words in the descending order of priority are sequentially matched with the characteristic parameters as the input patterns read in time series from the built-in input buffer. If the resulting score is greater than or equal to a predetermined threshold, the score,
The start point and the end point are supplied to and stored in an output buffer (not shown) built in the word spotting processing unit 4.

That is, in step S8, among the words output by the syntax analysis unit 5 together with the priority of the words, the N high-priority words are word-spotted.

After the word spotting process in step S8, the process proceeds to step S9, and it is determined whether or not data (speech characteristic parameter output from the voice analysis unit 2) remains in the input buffer incorporated in the word spotting processing unit 4. To be judged. If it is determined in step S9 that no data remains in the input buffer built in the word spotting processing unit 4, the process proceeds to step S10.
The variable N indicating the number N of words subjected to word spotting processing by the word spotting processing unit 4 is incremented by 1, and the process proceeds to step S11.

On the other hand, in step S9, when it is determined that the data remains in the input buffer built in the word spotting processing section 4, step S10 is skipped, the process proceeds to step S11, and the syntactic analysis section 5 creates a new one. If there is a word spotting processing request (search request), the number of requested words is added to the variable W, and the process proceeds to step S12.

In step S12, if the output buffer incorporated in the word spotting processing unit 4 stores the score, the start point, and the end point as the word spotting result, the score is supplied to the syntax analysis unit 5, and the process proceeds to step S13. The voice section detection unit 3 determines whether or not the frame in which the utterance has ended is detected, and whether or not the input buffer incorporated in the word spotting processing unit 4 is empty.

If it is determined in step S13 that the frame in which the utterance has ended is not detected by the voice section detector 3 or the input buffer incorporated in the word spotting processor 4 is not empty, step S13 Returning to S5, the processing from step S5 is repeated again.

In step S13, it is determined that the speech section detection unit 3 has detected the frame for which speech has ended,
If it is determined that the internal buffer of the word spotting processing unit 4 is empty, step S2 shown in FIG.
In step 1, the number of words having the highest priority among the words for which word parsing processing has been requested by the syntax analysis unit 5 has not been performed. Is set to step S2
Go to 2.

In step S22, of the words for which the word parsing process has been requested by the syntactic analysis unit 5 for the entire input speech, the word spotting process has not been performed yet. , The word spotting process of the N highest priority words is performed, the process proceeds to step S23, and the word spotting process result (score, start point, and end point) in step S22 is output to the syntax analysis unit 5, and step S24. Proceed to.

In step S24, the variable N is subtracted from the variable W, the subtracted value is set in the variable W, and the process proceeds to step S25. That is, in step S24, the word spotting process is performed in step S22 from the number of words that have not been subjected to the word spotting process among the words for which the syntax analysis unit 5 has requested the word spotting process. The number of words
It proceeds to step S25.

In step S25, if there is a new word spotting processing request (search request) from the syntax analysis unit 5, the number of requested words is added to the variable W, and the process proceeds to step S26, where the syntax analysis unit 5 It is determined whether or not the word spotting processing request (search request) has been completed, and whether or not the variable W is 0 is determined.

If it is determined in step S26 that the word spotting processing request (search request) from the syntax analysis unit 5 has not ended, or if it is determined that the variable W is not 0, the process returns to step S21, and again. The processing from step S21 is repeated.

When it is determined in step S26 that the word spotting processing request (search request) from the syntax analysis unit 5 has been completed and the variable W is 0, the processing is completed.

On the other hand, in the syntactic analysis unit 5, when the word spotting processing unit 4 supplies the score, start point, and end point of each word for which word spotting processing is requested, the score or the section length from the start point to the end point is supplied. ,
Also, based on the syntax information registered in advance in the built-in syntax dictionary, the input voice is analyzed, based on the analysis result, a new sentence hypothesis is created,
Priority is added to the words needed to analyze the new sentence hypothesis as described above (for words for which a word spotting process was previously requested, the priority is changed), and the word Spotting processing unit 4
Is output to.

In the syntactic analysis unit 5, after the speech section detection unit 3 detects the frame in which the utterance is completed,
Based on the result of its own parsing, the word candidate string as a word spotting result is excluded from the wrong (probably wrong) word candidate, and a correct (presumably correct) sentence (sentence) is obtained. , That is the external device operation part 6
Is output to.

In the external device operating unit 6, the external device connected to the external device operating unit 6 is operated according to the content of the sentence (sentence) output from the syntax analyzing unit 5. That is, the external device connected to the external device operation unit 6 is, for example, an AV device,
The sentence (sentence) output from the syntax analysis unit 5 is, for example, “C
In the case of “D playback”, in the external device operation unit 6, C
The AV device is operated so that the reproduction of D is started.

As described above, during voice input (steps S1 to S13), data (voice feature parameter) remains in the input buffer built in the word spotting processing unit 4, and the word spotting processing unit 4 operates. When the processing is delayed, the number N of words to be subjected to the word spotting processing is reduced (step S7), and the input buffer incorporated in the word spotting processing unit 4 is empty, so that the word spotting processing unit 4 has a sufficient processing capacity. The number N of words to be word-spotted is increased (step S10), and in the word-spotting processing unit 4, among the words for which word-spotting processing has been requested by the syntax analysis unit 5, N words are ordered in descending order of priority. Word spotting processing is performed (step S8).

After the voice input is completed (steps S21 to S26), the word spotting processing unit 4 has the highest priority word out of the words requested by the syntax analyzing unit 5 for the word spotting processing. Word spotting processing is performed sequentially from the group.

Therefore, the word spotting processing unit 4 sequentially performs word spotting processing from a word having a high priority, that is, a word required by the syntax analysis unit 5 to perform the syntax analysis. The response can be improved.

Furthermore, since the number of words to be word-spotted is changed in accordance with the processing capability of the word spotting processor 4, the number of words that can be processed in real time by the word spotting processor 4 is exceeded. However, even if given by the syntactic analysis unit 5, the processing can be performed quickly.

The case where the voice recognition device of the present invention is applied to an external device controller has been described above, but the present invention can be applied to any device that recognizes voice in addition to the external device controller.

Although the word spotting method in the word spotting processing unit 4 is not mentioned in the present embodiment, the word spotting processing unit 4 is not mentioned.
For example, for example, DP matching method or HMM
Method, JP-A-60-249198, JP-A-60-24991
The word spotting process can be performed based on any voice recognition algorithm, such as the voice recognition algorithm of the voice recognition device disclosed in JP-A-60-252396.

Further, in the present embodiment, the voice analysis unit 2 extracts the linear prediction coefficient from the voice as the feature parameter of the voice, but the present invention is not limited to this. That is, the voice analysis unit 2 can extract all characteristic parameters such as power, cepstrum coefficient, Percoll coefficient, formant, and zero-cross number for each predetermined frequency bandwidth from the voice.

[0060]

According to the speech recognition apparatus of the first aspect, the recognition means determines the number of words to be subjected to the speech recognition processing by the priority of the words added by the analysis means or the own speech.
The speech recognition processing is performed based on the recognition processing capability, and the speech recognition processing of the word in the speech is performed based on the priority of the word.
Therefore, it is possible to speed up the response process to the input voice.

According to the voice recognition device of the second aspect,
The recognition unit changes the word to be recognized in the voice from the voice based on the priority of the word during the voice section and after the end of the voice section. Therefore, for example, after the end of the voice section, it is possible to perform the voice recognition processing only on the word having the highest priority, so that the real-time property of the device can be improved.

According to the voice recognition device of the third aspect,
Since the analysis unit changes the priority of the word based on the analysis result of the recognition result of the recognition unit, it is possible to further speed up the response process to the input voice.

[0063]

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of an external device controller to which a voice recognition device of the present invention is applied.

FIG. 2 is a word spotting processing unit 4 of the embodiment shown in FIG.
3 is a flowchart illustrating the operation of the above.

FIG. 3 is a flowchart following the flowchart of FIG.

[Explanation of symbols]

1 Voice input section 2 Speech analysis section 3 Voice section detector 4 Word spotting processing section 5 Parsing part 6 External device operation section

─────────────────────────────────────────────────── ─── Continuation of the front page (56) Reference JP-A-1-255925 (JP, A) JP-A-6-161488 (JP, A) JP-A-3-177899 (JP, A) JP-A-63- 165900 (JP, A) (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 15/00-15/28 JISST file (JOIS)

Claims (3)

(57) [Claims] 1. A recognition unit for recognizing a voice, and an analysis unit for requesting the recognition unit to perform a voice recognition process of a word in the voice and analyzing a recognition result of the recognition unit to understand the voice. And the analysis means adds a priority for speech recognition processing to the word and supplies the word to the recognition means, and the recognition means calculates the number of words of the speech recognition processing as
Based on written word priority or own speech recognition processing ability
The speech recognition apparatus is characterized in that the speech recognition processing of the word in the speech is performed based on the priority of the word. 2. A detection unit for detecting a voice section of the voice is further provided, wherein the recognizing unit selects from among the voice based on the priority of the word during the voice section and after the end of the voice section. The voice recognition apparatus according to claim 1, wherein the voice recognition word is changed. 3. The voice recognition device according to claim 1 , wherein the analysis unit changes the priority of the word based on an analysis result of a recognition result of the recognition unit.