JPH07230293A - Voice recognition device - Google Patents

Abstract

PURPOSE:To improve the degree of freedom of ultering of a speaker and the recognition rate of voice. CONSTITUTION:In a control section 5, the silence interval of the voice detected by a voice interval detection section 5 and prescribed set times T1 and T2 are respectively compared (where T1<T2). In accordance with the results of the comparison, the voice is divided into prescribed segment units and is divided into sentence units constituted by segments. Then, in a voice recognition section 3, the sentence is voice recognized in the unit of the segments which constitute the sentence. In this case, corresponding to the recognition results of segments in the sentence proceeding in terms of time, the handling of succeeding segments is varied.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device suitable for use in recognizing voice, for example.

[0002]

2. Description of the Related Art In a conventional voice recognition apparatus, when a voice is input, for example, a voiceless portion (voiceless section) for a predetermined time or more, which is presumed to be a sentence segment in the voice, is detected. Detected and input speech is a unit separated by its silent part, that is, a sentence (an ordinary sentence, a word string having a dependency relation such as a compound word, a word string connected in some sense (for example, It is designed to perform continuous word speech recognition in units of (including addresses).

[0003]

By the way, the speaker is
In some cases, words in one sentence may be uttered while being separated from each other, or may be uttered without being separated from each other. Therefore, words separated by silent parts are not always sentence units. Therefore, there is a method of restricting the utterance of the utterer such that the utterance is divided into sentence units, but this method has a problem that the degree of freedom of the utterance is reduced and thus the operability of the apparatus is deteriorated.

Therefore, there is a method of increasing the degree of freedom of the pause time (pause) inserted between words by the speaker by setting the above-mentioned predetermined time to a relatively large value.

However, in this case, the speaker
There is a problem in that the utterance feels annoying because the sentence-to-sentence segment is forced to be silent for a predetermined time or longer.

[0006] Further, the speaker often coughs or utters an unnecessary word that is not a recognition target word, such as "that?", After the voice to be originally recognized. Therefore, as described above, when the predetermined set time is set to a large value, the cough and unnecessary words are also recognized as voice recognition targets, which reduces the recognition rate. there were.

The present invention has been made in view of such a situation, and is to improve the operability and recognition rate of the apparatus.

[0008]

A voice recognition apparatus according to claim 1 comprises a detection means for detecting a voice section of an input voice (for example, a voice section detection section 4 shown in FIG. 1) and a detection means. A non-voice section which is a section between the detected voice section and the voice section subsequently detected by the detection means,
A dividing unit that compares the first and second set times, divides the voice into predetermined segment units according to the comparison result, and divides the voice into sentence units composed of segments (for example, in FIG. The present invention is characterized by including a control unit 5) shown and a recognition unit (for example, the voice recognition unit 3 shown in FIG. 1) for recognizing a sentence divided by the dividing unit in units of segments constituting the sentence.

According to another aspect of the speech recognition device of the present invention, the first set time is shorter than the second set time, and the dividing means has a non-voice section shorter than the first set time. ,
The voice sections before and after the voiceless section are combined into one segment, and when the voiceless section is a value between the first and second set times, the voice sections before and after the voiceless section are respectively set. One segment, and the two segments are combined into one sentence, and when the non-voice section is longer than the second set time, the voice sections before and after the non-voice section are each one sentence. The feature is that

According to another aspect of the speech recognition apparatus of the present invention, if the recognition unit recognizes at least one word from the preceding segment when the sentence is composed of a plurality of segments, the succeeding segment in the sentence is recognized. Is characterized by ignoring.

According to another aspect of the speech recognition apparatus of the present invention, when the recognizing means recognizes at least one word from the preceding segment when the sentence is composed of a plurality of segments, the following segment in the sentence is recognized. All of them are treated as one new sentence, and voice recognition is further performed.

According to another aspect of the speech recognition apparatus of the present invention, when the recognition means recognizes a word from a preceding segment when the sentence is composed of a plurality of segments, the recognition means is restricted under the constraint corresponding to the word. , A segment following a preceding segment is speech-recognized.

In the voice recognition apparatus according to the sixth aspect, when the recognition means is composed of a plurality of segments, the recognition means performs voice recognition on all the segments, and among the recognition results obtained as a result, the likelihood The feature is that the segment with the highest frequency is used as the final speech recognition result.

[0014]

In the voice recognition apparatus according to the present invention, the voiceless section of the input voice is compared with the first and second set times, and the voice is given a predetermined value according to the comparison result. It is divided into segment units, and also divided into sentence units composed of segments. Then, the sentence is voice-recognized in units of the constituent segments. Therefore, by setting each of the first and second set times to an appropriate value, it becomes unnecessary to limit the utterance of the speaker, and the operability of the device can be improved.

In the voice recognition apparatus according to the second aspect, the first set time is shorter than the second set time,
When the voiceless section is shorter than the first set time, the dividing means includes the voice sections before and after the voiceless section to be 1
When the non-voice section has a value between the first and second set times, the voice sections before and after the non-voice section are each set as one segment, and
The two segments are combined into one sentence, and when the non-voice section is longer than the second set time, the voice sections before and after the non-voice section are each set to one sentence.
Therefore, by setting each of the first and second set times to an appropriate value, it becomes unnecessary to limit the utterance of the speaker, and the operability of the device can be improved.

In the speech recognition apparatus according to the third aspect, when a sentence is composed of a plurality of segments, when at least one word is recognized from the preceding segment, the subsequent segment in the sentence is ignored. To be done. Therefore, a part which is not a recognition target word, such as a cough and an unnecessary word, which follows the voice to be originally recognized is ignored, and the voice recognition rate can be improved.

In the speech recognition apparatus according to the fourth aspect, when the sentence is composed of a plurality of segments, when at least one word is recognized from the preceding segment, all the subsequent segments in the sentence are recognized. Speech recognition is further performed as one new sentence. Therefore, when a voice to be further recognized is followed by a voice to be originally recognized, the voice is recognized, so that the recognition rate can be improved.

In the speech recognition apparatus according to the fifth aspect, when a sentence is composed of a plurality of segments, when a word is recognized from the preceding segment, it is preceded by a constraint corresponding to the word. The segment following the segment to be recognized is recognized by voice. Therefore, the voice recognition rate can be further improved.

In the speech recognition apparatus according to the sixth aspect, when a sentence is composed of a plurality of segments, all of the segments are speech-recognized, and among the recognition results obtained as a result, the likelihood is the highest. The one with the higher segment is the final recognition result. Therefore, for example, it is possible to improve the recognition rate of the voice including many unnecessary words.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram showing the configuration of an embodiment of a voice recognition device of the present invention. The microphone 1 converts the input voice into a voice signal as an electric signal and outputs the voice signal to the A / D converter 2. A / D converter 2
Is configured to sample (A / D convert) the voice signal from the microphone 1 at a predetermined sampling interval and output the resulting digital voice signal to the voice recognition unit 3 and the voice section detection unit 4. There is.

The voice section detector 4 calculates, for example, the short-time power of the voice signal from the A / D converter 2, and detects a section whose power is equal to or more than a predetermined value as a voice section. Then, the remaining section is detected as a non-voice section and is output to the control unit 5.

The control section 5 compares the non-voice section from the voice section detecting section 4 with a preset predetermined time T ₁ or T _2, respectively, and determines the section of the input voice based on the comparison result. , Is divided into sections of a predetermined segment and sections of a sentence composed of segments.
Then, a control signal corresponding to the division result is output to the voice recognition unit 3. Furthermore, the control unit 5
Is also configured to control the voice recognition unit 3 in accordance with the recognition result (recognition result candidate) output from the voice recognition unit 3. Further, the control unit 5 is configured to output the voice recognition result from the voice recognition unit 3 to a device (not shown) in the subsequent stage.

The voice recognition unit 3 recognizes the voice signal from the A / D converter 2 based on a control signal supplied from the control unit 5, for example, continuous word voice.
That is, the voice recognition unit 3 is configured to recognize the continuous word voice of the portion of the voice signal corresponding to the section of the sentence in the section unit corresponding to the segment. The voice recognition in the voice recognition unit 3 is performed by, for example, HMM (Hidden Marko).
v Models) method (this method is effective for continuous word speech recognition) etc., and the recognition result candidate with the highest probability (likelihood) is output as the recognition result. There is. In the voice recognition unit 3, the HMM
As a symbol (code) observed from, the input voice signal is acoustically analyzed and is a voice feature amount, for example, LPC.
A cepstrum coefficient or the like is calculated, and a vector quantization of this is used.

Next, the operation will be described. First, when a voice is input to the microphone 1, the voice is converted into a voice signal which is an electric signal and is output to the A / D converter 2. In the A / D converter 2, the audio signal from the microphone 1 is A
It is D / D converted and output to the voice recognition unit 3 and the voice section detection unit 4.

The voice section detection unit 4 detects the short-time power of the voice signal from the A / D converter 2, and successively detects the sections whose values are equal to or more than a predetermined value Th as the voice section.
That is, for a voice signal having a short-time power as shown in FIG. 2A, a voice section as shown in FIG. 2B is detected. Further, in the voice section detection unit 4, among the non-voice sections other than the voice section, the sections (time) t ₀ , t ₁ , t ₂ , t ₃ , t ₄ sandwiched between the voice sections are controlled by the control section. 5 are sequentially output. The control unit 5, and a voiceless interval t ₀ to t _4, respectively from the voice section detection unit 4, and a preset predetermined time T _1, T ₂ respectively are sequentially compared.

Here, the predetermined time T ₁ is the upper limit value of the non-speech portion in the word, specifically, in Japanese, for example, the length of the consonant (for example, the statistical maximum value of the consonant length, etc.). ) Is set to, so to speak, a reference value for determining a break between words in the input voice. Further, the predetermined time T ₂ is larger than the time allowed as a pause between words (for example, pause time due to breathing) when the words are uttered continuously, and is inserted by the user of the device to separate sentences. That is, it is, so to speak, a value equal to or less than the minimum length of the spontaneous silence, and is a reference value for determining the separation between sentences in the input voice, so to speak. Therefore, the predetermined time T ₁ or T ₂ is a relatively small value or a large value, respectively, and the relationship between them is T ₁ <T ₂ .

When the control section 5 determines that the non-voice section is shorter than the predetermined time T ₁ , the voice sections before and after the non-voice section are considered to constitute one segment. It In addition, the non-voice section is a predetermined time T
_When it is determined that the value is ₁ or more and T ₂ or less, the speech sections before and after the non-speech section are considered to form one segment, respectively, and the two segments form one sentence. It is said that Further, when it is determined that the non-voice section is longer than the predetermined time T ₂ , the voice sections before and after the non-voice section constitute one sentence (independent sentence).

Therefore, in the case shown in FIG. 2B, T ₁ ≤t ₀ ≤T ₂ , T ₂ <t ₁ , T ₂ <t ₂ , T ₁ ≤t ₃ ≤
Assuming that T ₂ and T ₁ ≤t ₄ ≤T ₂ , the segment of the input speech is a sentence A composed of two segments (divided) and one segment, as shown in FIG. Sentence B consisting of and sentence C composed (divided) of three segments
Is divided into three sentence sections.

That is, in FIG. 2B, T ₁ ≤t ₀ ≤T
The speech sections before and after the non-speech section t ₀ of ₂ are set as independent segments and form the sentence A. In addition, a voiceless section t with T ₂ <t ₁
The speech sections before and after ₁ are one sentence A and B, respectively. Furthermore, T ₂ <before and after the speech section of the voiceless section t ₂ which is t ₂ are respectively one sentence B, and C. Also, T
The voice section before and after the non-voice section t ₃ where ₁ ≤ t ₃ ≤ T ₂ is
The sentence C is composed of one segment each. Further, the speech sections before and after the non-speech section t ₄ with T ₁ ≤t ₄ ≤T ₂ are also defined as one segment and form the sentence C.

Further, for example, when a voice signal of short time power as shown in FIG. 3A is inputted, the voice section detecting section 4 sequentially detects the voice sections as shown in FIG. 3B. Of the non-voice section that is the other section,
Sections (time) t ₅ , t ₆ , t ₇ , t sandwiched between voice sections
₈ is sequentially output to the control unit 5.

Now, t ₅ <T ₁ , T ₁ ≤t ₆ ≤T ₂ , t ₇ <T
_Assuming that ₁ and T ₂ <t ₈ , t ₅ <
The voice sections before and after the non-voice section t ₅ which is T ₁ are combined into one segment (FIG. 3 (c)). Furthermore, the voice section before and after the non-voice section t ₆ where T ₁ ≤t ₆ ≤T ₂ (however, the voice section preceding the non-voice section t ₆ in time is
As mentioned above, the total of 1
Since it is regarded as one segment, the voice section before the non-voice section t _{6 in} this case can be considered as that segment) is regarded as one independent segment (of the non-voice section t ₆ ). The preceding and following voice sections are configured to form independent segments),
It is assumed that sentence D is formed. Further, the voice sections before and after the non-voice section t ₇ where t ₇ <T ₁ are combined into one segment. Further, before and after the speech section of the T ₂ <is t ₈ voiceless interval t ₈ are each single statement D, E
Is configured.

As a result, the section of the input voice is shown in FIG.
As shown in (c), consists of two segments (division)
The sentence D is divided into two sentences, that is, the sentence D and the sentence E which is one segment.

The control unit 5 outputs a control signal to the voice recognition unit 3 in accordance with the result of the division, whereby the voice recognition unit 3 is operated.
To control. In the voice recognition unit 3, the voice signal from the A / D converter 2 is sequentially divided for each sentence in response to the control signal from the control unit 5, and each sentence is segmented in units of, for example, H.
Continuous word speech recognition is performed based on the MM method or the like.

That is, for example, the ward name of Tokyo's 23 wards and its town name are the target vocabulary for voice recognition, and the utterance pattern is "ward name" + "town name"("townname" is generated following "ward name". Pattern), and when only “ward name” or “town name” is allowed, when a voice is uttered, the process as described in FIG. 2 or FIG. 3 is first performed. Then, for example, as shown in the flow chart of FIG. 4, when the control section 5 detects a non-voice section longer than a predetermined time T ₂ based on the output of the voice section detection section 4 in step S1, the control section 5 detects the non-voice section. The voice section immediately before the voice section (the section where the short-time power is equal to or greater than the predetermined value Th) is detected as the section corresponding to the sentence.

Here, as described above, the predetermined time T ₂
Is a relatively large value (compared to the above-mentioned conventional predetermined time), so if a speaker inserts a slightly longer pause between words that should be included in one sentence, However, the sentence is not divided into different sentences, so that the degree of freedom of the utterance of the speaker can be improved.

After the processing of step S1, the process proceeds to step S2, and it is determined how many segments the voice section corresponding to the sentence detected in step S1 includes.

When it is determined in step S2 that the voice section corresponding to the sentence is composed of one segment, that is, only "ward name" or "town name" is uttered, or "ward name" and "town name" are spoken. If both of them are spoken continuously with little pause time, step S
Go to 3 and its segment (in this case equal to sentence)
A control signal is output from the control unit 5 to the voice recognition unit 3 so as to perform continuous word voice recognition on the voice signal corresponding to. As a result, in the voice recognition unit 3, the voice signal corresponding to the segment is subjected to continuous word voice recognition, the process proceeds to step S4, the recognition result is held in the variable R, and the process proceeds to step S12.

On the other hand, when it is determined in step S2 that the voice section corresponding to the sentence includes two or more segments, the process proceeds to step S5, and the leading segment (first segment (FIGS. 2 and 3)) A control signal is output from the control unit 5 to the voice recognition unit 3 so as to perform continuous word voice recognition on the voice signal corresponding to. As a result, in the voice recognition unit 3, the voice signal corresponding to the segment is subjected to continuous word voice recognition, the process proceeds to step S6, and the recognition result is held in the variable R ₁ .

Then, the process proceeds to step S7, and step S
At 6, it is determined whether the recognition result held in the variable R ₁ represents a “ward name”. In step S7, the recognition result is held in the variable R _1, when it is determined not to represent a "ward name", i.e., the variable R recognition result stored in _1, for example, "street" only, or "ward Name "+
If it represents a "town name", proceed to step S8,
The recognition result (only “town name” or “ward name” + “town name”) held in the variable R ₁ is held in the variable R, and the process proceeds to step S12.

If it is determined in step S7 that the recognition result held in the variable R ₁ indicates a "ward name", the variable R ₁ is changed from the voice recognition unit 3 to the control unit 5.
Is output to. Then, in the control unit 5, the recognition target is limited to the “town name” of the town actually existing in the ward indicated by the “ward name” which is the recognition result held in the variable R ₁ , and the next segment (in this case, A control signal is output to the voice recognition unit 3 so that the voice recognition of the second segment (FIGS. 2 and 3) is performed.

As a result, in the voice recognition section 3, in step S9, the voice signal corresponding to the next segment is
Continuous word speech recognition (in this case, word recognition may be used) by limiting the recognition target vocabulary.

Therefore, in this case, the speech recognition target vocabulary is narrowed down, so that the speech recognition rate can be improved.

After that, the process proceeds to step S10, and the "town name" which is the recognition result of the voice signal corresponding to the next segment.
Is held in the variable R ₂ and the process proceeds to step S11. In step S11, the recognition result stored in the variable R ₁ (in this case, “ward name” (eg, “Shinagawa Ward”)
The recognition result held in the variable R ₂ (in this case, the character string “ward name” + “town name” (eg, “Shinagawa-ku Kitashinagawa”) that is followed by “town name” (eg, “Kita-Shinagawa”) But,
The recognition result is held in the variable R, and the process proceeds to step S12.

In step S12, the variable R is supplied from the voice recognition unit 3 to the control unit 5, and the control unit 5 outputs the recognition result held in the variable R to the device at the subsequent stage, thus ending the processing. To do.

For the next sentence, step S is again performed.
The process from 1 is performed.

If there are three or more segments that make up the sentence, a segment after the second segment,
That is, the third and subsequent segments are ignored because they correspond to coughs and unnecessary words. Therefore, in this case, immediately after the second segment, if there is at least a non-voiced section for a predetermined time T ₁ , the subsequent portion is ignored, so that the speaker should be careful of the beginning of the utterance. Just by getting it, you can improve the voice recognition rate. That is, both the operability of the device and the recognition rate of voice can be improved.

Further, the voice recognition unit 3 outputs the recognition result candidate having the highest probability (likelihood) (hereinafter referred to as a score) among the recognition result candidates obtained as a result of the voice recognition processing as described above. It is done like this, but for example 2
It is possible to output the recognition result candidates with the highest scores after the th.

That is, for example, when the speech section corresponding to the sentence detected in step S1 includes two or more segments and a recognition result candidate of "ku name" is obtained from the first segment, the highest score among them is obtained. Of the recognition result candidates obtained by speech recognition of the second segment, some recognition result candidates with higher scores are recognized as the highest score of the first segment under the restriction of “ward name” which is a recognition result candidate having a high score. It is possible to output together with a high recognition result candidate (recognition result).

Further, the speech recognition of the second segment is performed under the restriction of each of the "ward names" which are the recognition result candidates having the higher score of the first segment, and the recognition of the higher score obtained as a result is performed. It is possible to output the result candidates together with each of the recognition result candidates having a higher score in the first segment.

Further, for example, when the speech section corresponding to the sentence detected in step S1 is composed of one segment, the one with the highest score among the recognition result candidates of that segment is output. You can

Next, FIG. 5 is a flow chart for explaining another operation example of the voice recognition device of FIG. In this case, in steps S21 to S25, the same processes as steps S1 to S5 described in FIG. 4 are performed.

Then, in step S26, the continuous word voice recognition result for the voice signal corresponding to the first segment in step S25 similar to step S5 of FIG. 4 is held in the variable R ₁ , and the score thereof is held in the variable S ₁ Held in.

Then, the process proceeds to step S27, and it is determined whether or not the recognition result held in the variable R ₁ in step S26 represents a "ward name". When it is determined in step S27 that the recognition result held in the variable R ₁ represents the “ward name”, the process sequentially proceeds to steps S29 to S32, and steps S9 to S described in FIG.
The same processing as that of 12 is performed, and the processing ends.

On the other hand, in step S27, the variable R ₁
When it is determined that the recognition result held in _No. does not represent the “ward name”, that is, the recognition result held in the variable R ₁ is, for example, only “town name” or “ward name” + “town name”. If so, the process proceeds to step S33, the speech signal corresponding to the next segment, that is, the second segment is recognized as a continuous word, and the process proceeds to step S34.

In step S34, step S3
The recognition result for the voice signal corresponding to the second segment in 3 is held in the variable R ₂ and the score thereof is held in the variable S ₂ and the process proceeds to step S35 where the score held in the variable S ₁ is , It is determined whether the score is smaller than the score held in the variable S ₂ . Step S
In 35, when it is determined that the score held in the variable S ₁ is smaller than the score held in the variable S ₂ , the process proceeds to step S 36, and the recognition result of the second segment held in the variable R ₂ Is held in the variable R, and the process proceeds to step S32.

Therefore, when the score for the recognition result of the second segment is higher than the score for the recognition result of the first segment, the first segment and the third and subsequent segments are ignored, and the recognition result of the second segment is ,
The final recognition result is output from the voice recognition unit 3 to the control unit 5.

On the other hand, in step S35, the variable S ₁
If it is determined that the score held in the is not smaller than the score held in the variable S ₂ , step S3
7, the recognition result of the first segment held in the variable R ₁ is held in the variable R, and the process proceeds to step S32.

Therefore, if the score for the recognition result of the first segment is equal to or higher than the score for the recognition result of the second segment, the second and subsequent segments are ignored and the first segment is ignored.
The recognition result of the segment is output from the voice recognition unit 3 to the control unit 5 as the final recognition result.

Usually, it is considered that the score obtained as a result of the speech recognition processing is small for the words that are not the recognition target words, that is, the unnecessary words. Therefore, as described above, the recognition result of the first segment is "ku name". If it is not, by deciding the recognition result of the first and second segments, which has the highest score, as the final recognition result, unnecessary words and the like can be ignored, and thus the recognition rate can be improved. Can be improved.

In the case shown in FIG. 5, as in the case of FIG. 4, not only the recognition result candidate with the highest score (recognition result) but also the recognition result candidates with the second and higher scores are output. can do.

Furthermore, in this case, the speech recognition up to the second segment is performed, but the speech recognition is performed for all the segments included in one sentence, and the recognition result of the segment with the highest score among them is obtained. The final recognition result may be used.

For the next sentence, step S is again performed.
The processing from 21 is performed.

Next, the voice recognition apparatus shown in FIG. 1 can be operated according to the flow chart shown in FIG. That is, first, in steps S41 to S46, the same processing as in steps S21 to S26 of FIG. 5 is performed.

Then, in step S47, it is determined whether or not the score for the recognition result of the first segment is less than or equal to a predetermined value C (for example, the minimum value that can be regarded as the score of the recognition result of the speech recognition target vocabulary). Is determined.
When it is determined in step S47 that the score for the recognition result of the first segment is not equal to or smaller than the predetermined value C, the process proceeds to step S48, and the recognition result of the first segment held in the variable R ₁ is set in the variable R. It is held, and the process proceeds to step S49.

If it is determined in step S47 that the score for the recognition result of the first segment is equal to or less than the predetermined value C, the process proceeds to step S50, and the voice signal corresponding to the next segment, that is, the second segment. But,
The continuous words are recognized, and the process proceeds to step S51. In step S51, the recognition result for the audio signal corresponding to the second segment in step S50 is held in the variable R, and the process proceeds to step S49.

In step S49, the variable R is supplied from the voice recognition unit 3 to the control unit 5, and the control unit 5 outputs the recognition result held in the variable R to the subsequent apparatus, and the process is terminated. To do.

Therefore, in FIG. 6, when the score for the recognition result of the first segment is larger than the predetermined value C, the second and subsequent segments are ignored and the recognition result of the first segment is the final recognition result. If the score for the recognition result of the first segment is less than or equal to a predetermined value C, the first segment and the third and subsequent segments are ignored, and the recognition result of the second segment is the final recognition result. Will be said.

As described above, it is conceivable that an unnecessary word or the like has a small score obtained as a result of the voice recognition process. Therefore, when the score of the recognition result of the first segment is large, the recognition result is large. , And when the recognition result of the first segment has a small score, by ignoring the recognition result and adopting the recognition result of the second segment, unnecessary words are ignored. This can improve the recognition rate.

In this case as well, as in the case of FIG. 4, not only the recognition result candidate with the highest score (recognition result) but also the recognition result candidates with the second and higher scores can be output. it can.

For the next sentence, step S is again performed.
The processing from 41 is performed.

Next, in the above description, since the vocalization pattern is limited to only "ward name" + "town name", and only "ward name" or "town name", the segments forming the sentence are quite special. As long as no utterance is made, the number of segments is at most two. Therefore, in FIGS. 4 to 6, the third and subsequent segments are ignored, but the utterer may say, for example, “ward name” + “town name”, When the word “ward name” or “town name” is reworded, or when an unnecessary word is uttered between the recognition target words, the recognition target word may be included in the third segment and thereafter.

Therefore, the voice recognition apparatus shown in FIG. 1 can be operated according to the flowchart shown in FIG. 7, for example. In FIG. 7, step S6
7 is the same as that of steps S41 to S51 of FIG. 6 except that the processing of step S72 is performed between steps S7 and S68.

That is, in FIG. 7, when the score for the recognition result of the first segment is larger than the predetermined value C, in step S72, the process of setting the second and subsequent segments as the next sentence is performed. Then, in step S68,
After performing the process of S69, the process of step S62 and the subsequent steps is performed on the portion newly made as a sentence in step S72. Therefore, in this case, even if the recognition target vocabulary is included in the third segment and thereafter, it can be recognized. However, since it is considered rare that there is a recognition target vocabulary in the third segment and thereafter, it is actually appropriate to ignore it as an unnecessary word in the third segment and thereafter.

Also in this case, similarly to the case in FIG. 4, not only the recognition result candidate with the highest score (recognition result) but also the recognition result candidates with the second and higher scores can be output. it can.

The processing from step S61 is performed again for the original next sentence.

As described above, the predetermined time T ₁ for cutting out the word-like portion of the voice and the predetermined time T ₂ for cutting out the sentence related to the user interface are used to set the predetermined time. Of two voice intervals with an interval shorter than the time T ₁ is one segment, and two voice intervals (segments) with an interval of a value within a predetermined time T _{1 to} T ₂ are one sentence. , Two voice intervals (segments) that are separated by an interval longer than a predetermined time T ₂ are included in
Since each sentence is included in a separate sentence, and when the sentence includes a plurality of segments, the handling of the following segments is determined according to the recognition result of the first segment. The recognition rate can be improved.

In the present embodiment, the voice recognition unit 3
Then, the LPC cepstrum coefficient is used as the feature amount of the voice, but the present invention is not limited to this. Also,
The voice recognition in the voice recognition unit 3 can also be performed according to an algorithm other than the HMM method.

Further, in the present embodiment, the voice section is detected based on the short time power of the voice signal.
Besides this, for example, the voice section can be detected based on the number of zero-crossings of the voice signal per unit time. Alternatively, the voice section may be detected based on both the short-time power of the voice signal and the number of zero crossings per unit time.

Further, in this embodiment, the predetermined time T
_Although the actual values of ₁ and T ₂ were not mentioned, it was found from the experiment that the predetermined times T ₁ and T ₂ were approximately 0.3, respectively.
The result has been obtained that ms and 1.0 s are preferable.

Further, in the present specification, "or more"
Alternatively, "less than or equal to" may be read as "greater than" or "less than", respectively, and "greater than" or "less than" may be read as "greater than or equal to" or "less than or equal to".

[0081]

As described above, according to the present invention, the operability of the apparatus can be improved and the recognition rate of voice can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a voice recognition device of the present invention.

FIG. 2 is a diagram illustrating an operation of a control unit 5 in the embodiment of FIG.

FIG. 3 is a diagram illustrating an operation of a control unit 5 in the embodiment of FIG.

FIG. 4 is a flowchart illustrating a first operation example of the embodiment of FIG.

FIG. 5 is a flowchart illustrating a second operation example of the embodiment of FIG.

FIG. 6 is a flowchart illustrating a third operation example of the embodiment of FIG.

7 is a flowchart illustrating a fourth operation example of the embodiment of FIG.

[Explanation of symbols]

 1 Microphone 2 A / D converter 3 Voice recognition unit 4 Voice section detection unit 5 Control unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hiroaki Ogawa 6-735 Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Inventor table Masanori 6-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Inventor Kazuo Watanabe 6-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Inventor Katsuki Minamino 6-7-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Soni -Inside the corporation

Claims (6)

[Claims] 1. A detection unit for detecting a voice section of an input voice, a section between the voice section detected by the detection unit and a voice section subsequently detected by the detection unit. A dividing unit that compares the voice section with the first and second set times, divides the voice into predetermined segment units according to the comparison result, and divides the voice unit into sentence units composed of the segments. And a recognition unit that recognizes the sentence divided by the dividing unit in units of segments that compose the sentence. 2. The first set time period is shorter than the second set time period, and the dividing means sets the voiceless voice when the voiceless period is shorter than the first set time period. The voice sections before and after the section are combined into one segment, and the non-voice section is
When the value is between the first and second set times, the voice sections before and after the non-voice section are set as one segment, respectively, and the two segments are combined to form one sentence. The speech recognition apparatus according to claim 1, wherein when the section is longer than the second set time, the speech sections before and after the non-speech section are each one sentence. 3. The recognition means, when the sentence is composed of a plurality of segments, from the preceding segment,
2. When recognizing at least one word, ignore subsequent segments in the sentence.
Alternatively, the voice recognition device according to item 2. 4. The recognizing means, when the sentence is composed of a plurality of segments, from the preceding segment,
The speech recognition apparatus according to claim 1 or 2, wherein when at least one word is recognized, all subsequent segments in the sentence are treated as one new sentence for further speech recognition. 5. The recognition means, when the sentence is composed of a plurality of segments, from the preceding segment,
The speech recognition apparatus according to claim 1 or 2, wherein when a word is recognized, a segment following the preceding segment is speech-recognized under a constraint corresponding to the word. 6. The recognizing means, when the sentence is composed of a plurality of segments, performs speech recognition on all the segments, and of the recognition results obtained as a result, the one having the highest likelihood. The voice recognition device according to claim 1 or 2, wherein is a final recognition result.