JP2017211513A - Speech recognition device, method therefor, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition device, etc., with which it is possible to detect an iteration and avoid the degradation of recognition accuracy due to the iteration.SOLUTION: A speech recognition device includes: a first utterance section detection unit for assuming a non-speech section less than a prescribed threshold θas one portion of a speech section and detecting a first speech section that is a speech section of speech data; a second utterance section detection unit for assuming a non-speech section less than a prescribed threshold θ, where θ<θ, as one portion of a speech section and detecting a second speech section that is a speech section of speech data; a speech recognition unit for performing speech recognition on the speech data that corresponds to the first speech section to obtain a first recognition result and performing speech recognition on the speech data that corresponds to the second speech section to obtain a second recognition result; and a recognition result correction unit for calculating the degree of similarity between the first recognition result of some speech section and the first recognition result of an other speech section and, when the degree of similarity is larger than a prescribed threshold, deletes the first recognition result of the some speech section, and defines the first recognition result that is left as a recognition result.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a speech recognition technique for causing a computer to recognize a human utterance and converting the utterance into a character string.

  In the speech recognition technology, using a speech segment detection technology (see Non-Patent Document 1), only the user's speech is cut out and speech recognition is performed. By cutting out only the user's utterance with high accuracy, it is possible to exclude a non-speaking noise section and to perform voice recognition with high accuracy.

  In speech segment detection, a method is often employed in which speech likelihood and non-speech likelihood are calculated for each frame of input speech to determine speech segments.

  The utterance passed to speech recognition as one utterance does not pass only the section determined as the speech section from the likelihood calculation, but gives a non-speech section as a margin before and after the speech section, or the speech section and the speech section Also included is a short non-speech segment that exists in between (hereinafter also referred to as “gap”).

  In general, how much margin is given and how much gap is used for one utterance is manually adjusted according to the use environment.

Masayoshi Fujimoto, “Fundamentals of Speech Interval Detection and Recent Research Trends”, IEICE Technical Report., SP2010-23 (2010-06), pp.7-12.

  However, there are cases in which the user re-states the desired utterance in a speech dialogue system, etc., and in that case, the re-utterance of the utterance reduces the linguistic connection compared to natural continuous utterance, so the speech recognition accuracy Deteriorates. In addition, it becomes difficult to interpret the semantics of a spoken dialogue system, and the reliability of the spoken dialogue system decreases.

  Since the restated part is also an utterance section, it cannot be excluded by the utterance section detection process.

  By setting the speech segment gap to be combined in the speech segment detection to be short, the speech segment passed to speech recognition is shortened. When rephrasing, a short non-speech segment is inserted, so it can be divided into reutterance and subsequent utterances by adjusting the gap, but it becomes difficult to use the connection as a context if it becomes easier to divide shortly Therefore, recognition accuracy deteriorates.

  It is an object of the present invention to provide a speech recognition device, a method and a program thereof that can avoid deterioration in recognition accuracy due to rephrasing by detecting rephrasing.

In order to solve the above-described problem, according to one aspect of the present invention, the speech recognition apparatus regards a non-speech interval less than a predetermined threshold θ ₁ as a part of a speech segment, and is a speech segment of speech data. A first utterance section detecting unit that detects the first speech sections s ₁ , s ₂ ,..., S _L , and θ ₁ <θ ₂ and a non-speech section less than a predetermined threshold θ ₂ is regarded as a part of the speech section. Te, the second speech section t ₁ is a speech segment of the speech data, t _2, ..., a second utterances section detection unit for detecting a t _J, the first speech section s _1, s _2, ..., corresponding to s _L the first recognition result performs speech recognition sr _1, sr ₂ the audio data to be, ..., give sr _L, the second speech section t _1, t _2, ..., sound the audio data corresponding to t _J A speech recognition unit that performs recognition and obtains second recognition results tr ₁ , tr ₂ ,..., Tr _J , i = 1,..., L-1, m = 1, 2,. and L-1 following any integer, when L> J, first recognition result sr _i and the first recognition result sr _{i + m} Recognition result similarity alpha _{i, m} is calculated and if the similarity alpha _{i, m} is greater than a predetermined threshold value, which remove the first recognition result sr _i, only the remaining first recognition result as recognition result of the And a correction section.

In order to solve the above-described problem, according to another aspect of the present invention, a speech recognition method regards a non-speech interval less than a predetermined threshold θ ₁ as a part of a speech interval, A first speech segment detection step for detecting a certain first speech segment s ₁ , s ₂ ,..., S _L , and θ ₁ <θ _2, and a non-speech segment less than a predetermined threshold θ _{2 as} a part of the speech segment Assuming that the second speech segment detection step for detecting the second speech segment t ₁ , t ₂ ,..., T _J which is the speech segment of the speech data, and the first speech segment s ₁ , s _{2 ,.} the first recognition result performs speech recognition on the corresponding audio data sr _1, sr _2, ..., give sr _L, the second speech section t _1, t _2, ..., the audio data corresponding to t _J A speech recognition step for performing speech recognition and obtaining second recognition results tr ₁ , tr ₂ ,..., Tr _J , i = 1,..., L−1, m = 1, 2,. Any integer less than or equal to L-1 and when L> J, the first recognition result s Calculate the r _i and similarity alpha _{i, m} the first recognition result sr _{i + m,} if the similarity alpha _{i, m} is greater than a predetermined threshold, deletes the first recognition result sr _i, remaining the A recognition result correcting step in which only one recognition result is used as a recognition result.

  According to the present invention, it is possible to avoid deterioration in recognition accuracy due to rephrasing by detecting rephrasing, and there is an effect of improving the reliability of a system using speech recognition.

The functional block diagram of the speech recognition apparatus which concerns on 1st embodiment. The figure which shows the example of the processing flow of the speech recognition apparatus which concerns on 1st embodiment. The figure which shows the example of input audio | voice data, a 1st audio | voice area, and a 2nd audio | voice area. The figure which shows the example of input audio | voice data, a 1st recognition result, and a 2nd recognition result. The figure which shows the example of the processing flow of the recognition result correction part 103. FIG.

  Hereinafter, embodiments of the present invention will be described. In the drawings used for the following description, constituent parts having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted.

<First embodiment>
In the present embodiment, two speech segment detections having different sensitivities (different gap lengths) are operated, and the rephrasing is detected by referring to the speech recognition result for each.

  FIG. 1 is a functional block diagram of the speech recognition apparatus 100 according to the first embodiment, and FIG. 2 shows a processing flow thereof.

  The speech recognition apparatus 100 includes a first utterance section detection unit 101-1, a second utterance section detection unit 101-2, a speech recognition unit 102, and a recognition result correction unit 103.

  The voice recognition apparatus 100 receives voice data collected by a voice device such as a microphone, recognizes the voice data, and outputs a voice recognition result.

<First utterance section detection unit 101-1 and second utterance section detection unit 101-2>
The first utterance section detection unit 101-1 receives voice data, regards a non-speech section less than a predetermined threshold θ ₁ as a part of the voice section, and first voice section s ₁ that is a voice section of the voice data. , s ₂ ,..., s _L are detected (S 101-1) and output to the speech recognition unit 102.

The second utterance section detection unit 101-2 receives voice data as input, regards a non-speech section less than a predetermined threshold θ ₂ as a part of the voice section, and a second voice section t ₁ that is a voice section of the voice data. , t ₂ ,..., t _J are detected (S101-2) and output to the speech recognition unit 102. However, θ ₁ <θ _{2 is} assumed.

In the present embodiment, the first speech section _{s 1, s 2, ...,} s L and the second speech section t _1, t _2, ..., as t _J, and outputs the voice data obtained by cutting out speech interval. However, since it is only necessary to acquire the voice section and the voice data corresponding to the voice section, information indicating the voice data before being cut out and the voice section (for example, (i) start time and end time of the voice section, (ii) voice The start time of the section and the length of the voice section may be output.

For example, the first utterance section detection unit 101-1 and the second utterance section detection unit 101-2 analyze input voice data from the voice device, and detect a voice section. As the speech segment detection method, any existing speech segment detection method may be used, and an optimum method may be selected as appropriate in accordance with the use environment or the like. For example, the speech section detection method of Non-Patent Document 1 may be used. However, in the present embodiment, the voice section detection process is performed with two settings having different detection sensitivities. 3, the input voice data, the first speech section _{s 1, s 2, ...,} s L and the second speech section t _1, t _2, ..., show examples of t _J.

The first utterance section detection unit 101-1 is set so as to divide the utterance sensitively even in a short non-speech section in the utterance that occurs when speaking while rephrasing or thinking (see FIG. 3). That is, the threshold value θ ₁ is set to a small value. On the other hand, the second utterance section detecting unit 101-2 sets such a short non-voice section as a part of the voice section (see FIG. 3). That is, the threshold value θ ₂ is set to a value larger than the threshold value θ ₁ . For example, the first utterance section detection unit 101-1 and the second utterance section detection unit 101-2 internally calculate a ratio between speech likelihood and non-speech likelihood (hereinafter also referred to as likelihood ratio). When a section (non-speech section) where the degree ratio is below a certain threshold β is sandwiched by a section (speech section) above the threshold β, the non-speech section is shorter than a certain time length θ ₁ or θ ₂ The non-voice section is processed as a voice section. The difference in setting here is that this time length θ ₁ is very short, for example, set to 0.1 seconds, etc., while the time length θ ₂ is set to about 0.5 seconds in consideration of general intervals in normal speech. It is to set to. By making θ ₁ very short, the speech section (voice section) can be easily divided finely. In the first utterances section detection unit 101-1 and the second utterances section detection unit 101-2, a plurality of section detection result of moving at different settings (first speech section _{s 1, s 2, ...,} s L and a second audio (Interval t ₁ , t ₂ ,..., T _J , where θ ₁ <θ ₂ , the number L of the first speech segment is always equal to or greater than the number J of the second speech segment, that is, L ≧ J) Is output.

<Voice recognition unit 102>
The voice recognition unit 102 receives voice data, first voice sections s ₁ , s ₂ ,..., S _L and second voice sections t ₁ , t ₂ ,..., T _J (voice data obtained by cutting out voice sections). first speech section s _1, s _2, ..., the first recognition result sr _1, sr ₂ performs speech recognition on the speech data corresponding to s _L, ..., give sr _L, the second speech section t ₁ , t ₂ ,..., t _J are subjected to speech recognition to obtain second recognition results tr ₁ , tr ₂ ,..., tr _J (S102) and output to the recognition result correction unit 103. As the speech recognition method, any existing speech recognition method may be used, and an optimal method may be selected as appropriate in accordance with the usage environment. Figure 4 is a first recognition result _{sr 1, sr 2, ...,} sr L and second recognition result tr _1, tr _2, ..., an example of tr _J.

<Recognition Result Correction Unit 103>
The recognition result correcting unit 103 receives the first recognition result sr ₁ , sr ₂ ,..., Sr _L and the second recognition result tr ₁ , tr ₂ ,..., Tr _J, and when L> J, the first recognition result sr _i a first recognition result to calculate the similarity alpha _{i, m} and sr _{i + m,} if the similarity alpha _{i, m} is greater than a predetermined threshold, deletes the first recognition result sr _i, remaining the Only one recognition result is set as a recognition result (S103) and output as an output value of the speech recognition apparatus 100. However, i = 1, 2,..., L-1 and m = 1. When L> J, J indicates the number of utterances included in the second recognition result, and is at least 1 or more, and L is greater than 1.

FIG. 5 shows an example of the processing flow of the recognition result correction unit 103. The recognition result correcting unit 103 determines that the number of utterance sections L and J input from the first utterance section detection unit 101-1 and the second utterance section detection unit 101-2 are different (L ≠ J, where L ≧ J Therefore, when L> J is satisfied (in the case of yes in S103-A in FIG. 5), rephrasing detection is performed on the recognition result with the larger number of utterance sections (that is, the first recognition result) (S103-B). ~ S103-H). In this example, it is determined that rephrase is not performed when flag is 0, and it is determined that restatement is performed when flag is 1. Therefore, first, flag is set to 0 (S103-B). When the numbers L and J of the utterance sections are the same (L = J, no in S103-A in FIG. 5), no processing is performed, and the recognition result with the longer utterance section length (second recognition result). tr ₁ , tr ₂ ,..., tr _J are output as they are (S103-J). By such processing, the subsequent processing S103-B to S103-l can be omitted. Detection of restatement are those which back and forth on the time axis among the plurality speech period (in this embodiment, sr _i and sr _{i + 1)} to calculate the similarity alpha _{i, m} recognition result string (S103- D), when the when exceeds a certain similarity threshold η of (S103-E yes), sr i + 1 is determined as it rephrasing locations sr _i, sets the flag to 1 (S103 -F). However, rather than using all recognition result string in the calculation of similarity, from the end of the previous recognition results sr _i n characters shall be used from the beginning of sr _{i + 1} n characters only, similarity calculation itself Is calculated using a known technique. Here, n is, for example, 5. The degree of similarity may be calculated as long as two character sequences can be compared and a score can be output. For example, DP matching may be used. When the similarity score is normalized from 0 to 1, η should be about 0.85. From the calculated degree of similarity calculation result, the similarity in speech that proximity when exceeds the η deletes the previous recognition results sr _i close (S103-G). This processing is performed for all the recognition results sr _i (S103-C, The numerical values in the upper end of the S103-C in FIG. 5 represents the initial value of the variable i, closing, the increment value.), Restates may have been performed (in the case of yes in S103-H), utterance similarity (first speech recognition results that have not been deleted sr _i) the recognition result was below the η outputs only adjacent (S103- I). The first speech recognition result sr _L (when i = L) is always output because i = L + 1 is not compared. Incidentally, the first recognition result _{sr 1, sr 2, ...,} sr L , the second recognition result tr _1, tr _2, ..., compared to tr _J, divided shortened. Therefore, first recognition result _{sr 1, sr 2, ...,} sr L , the second recognition result tr _1, tr _2, ..., compared to tr _J, becomes difficult to use the connection as a context, recognition accuracy is degraded I think that. Therefore, if the similarity does not exceed the threshold η in any adjacent recognition result, second recognition results tr ₁ , tr ₂ ,..., Tr _J are output (S103-J).

<Effect>
With the above configuration, it is possible to avoid deterioration in recognition accuracy due to rephrasing by detecting rephrasing, and there is an effect of improving the reliability of a system using speech recognition.

<Second embodiment>
A description will be given centering on differences from the first embodiment.

The recognition result correcting unit 103 receives the first recognition result sr ₁ , sr ₂ ,..., Sr _L and the second recognition result tr ₁ , tr ₂ ,..., Tr _J, and when L> J, the first recognition result sr _i a first recognition result to calculate the similarity alpha _{i, m} and sr _{i + m,} if the similarity alpha _{i, m} is greater than a predetermined threshold, deletes the first recognition result sr _i, remaining the Only one recognition result is set as a recognition result (S103) and output as an output value of the speech recognition apparatus 100. Here, i = 1, 2,..., L−1, m = 1, 2,..., M, and M is an integer from 1 to L−1.

The recognition result correcting unit 103 does not compare the similarity of the recognition result with only the immediately preceding utterance section, but calculates the similarity with all the past utterance sections or the past M utterance sections (recognition results). There is also a form. In this case, since the most recent utterance section is more likely to be rephrased, it is preferable to introduce a weight according to the temporal distance from the comparison target. For example, a score of similarity calculated in DP matching or the like as alpha _{i, m,} shift start time between speech segment (the start utterance of the first recognition results sr _i utterance start time and the first recognition result of sr _{i + m} When Δt _{i + m} is defined as the difference from the time, the similarity considering the distance weight may be defined as α _{i, m} / Δt. That means
α _{i, m} = α _{i, m} / Δt _{i + m}
The similarity α _{i, m} is updated. Also, M should be set to about 3,4. Alternatively, time constraints may be imposed such as utterance intervals within the past T seconds. T may be 5 seconds, for example.

<Effect>
By setting it as such a structure, the effect similar to 1st embodiment can be acquired. Furthermore, by expanding the determination range of whether or not rephrasing is performed, more accurate speech recognition is possible. In addition, since it becomes the same structure as 1st embodiment when M = 1, 1st embodiment can also be said to be an example of 2nd embodiment.

<Third embodiment>
A description will be given centering on differences from the first embodiment.

  The recognition result correcting unit 103 calculates the similarity of the recognition result character string, but instead of the similarity in the notation of the character string, it may be converted into a phoneme once to calculate the similarity of the phoneme sequence. When the speech recognition unit 102 performs speech recognition processing, a phoneme sequence is acquired in the process of converting speech data into a speech recognition result (character string), and the similarity is calculated using the phoneme sequence. May be. Further, in the present embodiment, the similarity of the phoneme series is calculated, but the similarity of the voice data may be calculated, and the feature amount obtained from the voice data (for example, MFCC (Mel Frequency Cepstrum Coefficient), etc.) The similarity may be calculated. However, the accuracy of the similarity is better if the similarity is calculated using a character string or a phoneme sequence.

<Other variations>
The present invention is not limited to the above-described embodiments and modifications. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

<Program and recording medium>
In addition, various processing functions in each device described in the above embodiments and modifications may be realized by a computer. In that case, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its storage unit. When executing the process, this computer reads the program stored in its own storage unit and executes the process according to the read program. As another embodiment of this program, a computer may read a program directly from a portable recording medium and execute processing according to the program. Further, each time a program is transferred from the server computer to the computer, processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program includes information provided for processing by the electronic computer and equivalent to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In addition, although each device is configured by executing a predetermined program on a computer, at least a part of these processing contents may be realized by hardware.

Claims (8)

A first utterance section detector that detects first speech sections s ₁ , s ₂ ,..., S _L that are speech sections of speech data, considering a non-speech section less than a predetermined threshold θ ₁ as a part of the speech section. When,
and θ ₁ <θ _2, the predetermined threshold value theta ₂ less than non-speech section is regarded as part of the speech segment, the second speech section t ₁ is a speech segment of the speech data, t _2, ..., a t _J A second utterance section detection unit to detect;
, S _L to perform speech recognition on the speech data corresponding to the first speech sections s ₁ , s ₂ ,..., S _L to obtain first recognition results sr ₁ , sr ₂ ,. t _1, t _2, ..., the second the recognition result performed speech recognition on the speech data corresponding to _{t J tr 1, tr 2,} ..., a speech recognition unit for obtaining a tr _J,
i = 1, ..., a L-1, m = 1, 2, ..., and M, M is 1 or more L-1 following any integer, when L> J, a first recognition result sr _i similarity alpha _{i, m} the first recognition result sr _{i + m} is calculated, if the similarity alpha _{i, m} is greater than a predetermined threshold, deletes the first recognition result sr _i, the remaining first recognized Including a recognition result correction unit that recognizes only the result as a recognition result,
Voice recognition device. The speech recognition device according to claim 1,
The similarity alpha _{i, m} is said tail from n characters of the first recognition results sr _i, a similarity between the first recognition result sr _{i + m} from the beginning n characters,
Voice recognition device. The speech recognition device according to claim 1 or 2,
M = 1,
Voice recognition device. The voice recognition device according to any one of claims 1 to 3,
The difference between the utterance start time and the utterance start time of the first recognition results sr _{i + m} of the first recognition results sr _i and Δt _{i + m, M} is any integer less than 2 or i, the recognition result in modifying portion, wherein the similarity alpha _{i, m} is updated by the following equation, if the similarity alpha _{i, m} after update is larger than a predetermined threshold, deletes the first recognition result sr _i, the remaining first recognized Only the result is the recognition result.
α _{i, m} = α _{i, m} / Δt _{i + m}
Voice recognition device. The voice recognition device according to any one of claims 1 to 4,
In the recognition result correction unit, calculated from the degree of similarity alpha _i, phoneme series _m for the first recognition result sr _i and phoneme sequences for the first recognition result sr _{i + m,}
Voice recognition device. The voice recognition device according to any one of claims 1 to 5,
The recognition result correction unit, when L = J, omitting the deletion of computation and the first recognition result sr _i of the similarity alpha _{i, m,} and the recognition result of the second recognition result,
Voice recognition device. A first utterance section detection step of detecting first speech sections s ₁ , s ₂ ,..., S _L that are speech sections of speech data, considering a non-speech section less than a predetermined threshold θ ₁ as a part of the speech section. When,
and θ ₁ <θ _2, the predetermined threshold value theta ₂ less than non-speech section is regarded as part of the speech segment, the second speech section t ₁ is a speech segment of the speech data, t _2, ..., a t _J A second utterance section detecting step to detect;
, S _L to perform speech recognition on the speech data corresponding to the first speech sections s ₁ , s ₂ ,..., S _L to obtain first recognition results sr ₁ , sr ₂ ,. t _1, t _2, ..., the second recognition result performs speech recognition on the speech data corresponding to _{t J tr 1, tr 2,} ..., a speech recognition step of obtaining a tr _J,
i = 1, ..., a L-1, m = 1, 2, ..., and M, M is 1 or more L-1 following any integer, when L> J, a first recognition result sr _i similarity alpha _{i, m} the first recognition result sr _{i + m} is calculated, if the similarity alpha _{i, m} is greater than a predetermined threshold, it deletes the first recognition result sr _i, the remaining first recognized A recognition result correction step in which only the result is a recognition result,
Speech recognition method.   A program for causing a computer to function as the voice recognition device according to any one of claims 1 to 6.

Images