JP5427140B2 - Speech recognition method, speech recognition apparatus, and speech recognition program - Google Patents

Description

  The present invention relates to a speech recognition method, a speech recognition apparatus, and a speech recognition program for efficiently recognizing speech data of various sound qualities.

  In recent years, it has become possible to easily obtain a large amount of audio data as a memory element for recording audio data becomes cheaper. When recognizing such audio data, the recognition accuracy and processing time greatly vary depending on the quality of the audio data.

  In view of this, conventionally, a method for suppressing a defect caused by a speech recognition error by giving a reliability to the speech recognition result has been studied. For example, Patent Document 1 is known as a prior art that gives reliability to a speech recognition result. FIG. 1 shows a functional configuration of a speech recognition apparatus 900 disclosed in Patent Document 1. The speech recognition apparatus 900 includes an acoustic analysis unit 120, an acoustic model storage unit 140, a dictionary / language model storage unit 150, a search unit 160, and a reliability calculation unit 190.

  The acoustic analysis unit 120 generates an acoustic feature parameter series 130 by analyzing, for example, a mel frequency cepstrum coefficient (MFCC) of the input speech signal 110 in units called frames of several tens of ms. The searching unit 160 searches the acoustic feature parameter series 130 for a speech recognition result candidate using the acoustic model storage unit 140 and the dictionary / language model storage unit 150. As a result of the search, the speech recognition results 170 from the top to the Nth and the score 180 for each speech recognition result are output.

  The reliability calculation unit 190 calculates and outputs a reliability score 195 corresponding to each of the plurality of speech recognition results 170 based on the speech recognition result 170 and the score 180. The reliability score 195 is obtained from, for example, N best candidates obtained as a speech recognition result, a simple score difference between those scores, and an addition average.

  By referring to the reliability score 195, the speech recognition result 170 corresponding to the reliability score 195 is discarded or the speech recognition result is confirmed with respect to the speaker. By performing such processing, occurrence of problems due to erroneous recognition has been suppressed.

JP 2005-148342 A

  However, in the conventional speech recognition apparatus 900, the reliability score is calculated from the speech recognition result after the speech recognition processing and the score accompanying the speech recognition result. Therefore, it takes time for speech recognition processing to obtain a reliability score. Among those with low recognition accuracy due to reasons such as poor S / N ratio, even if the beam width at the time of search is widened or unsupervised adaptation is performed, it is possible to improve recognition accuracy only by misrecognition. There are some voice data that cannot be used. Therefore, when the reliability score is calculated from the score after the voice recognition process is performed, there is a problem that extra processing time is required for the unusable voice data. Also, when performing speech recognition processing on a large number of audio files, it takes time to process an audio file with low speech recognition accuracy, and processing of other audio files with high speech recognition accuracy does not proceed. There is a problem of lowering the processing efficiency. In addition, since the processing is based on the speech recognition result using the language model, there is a problem that the reliability score value depends on the language model.

  The present invention has been made in view of such a problem, and is capable of calculating a reliability score in a short processing time without performing a voice recognition process, and outputting a reliability score independent of a language model. An object is to provide a recognition device, a speech recognition method, and a speech recognition program.

In order to solve the above problems, a speech recognition method according to the present invention obtains a speech feature amount sequence by analyzing speech feature amounts of a speech digital signal in units of frames, and uses a speech feature amount sequence for each frame, The highest product of the output probability b _s (o _t ) obtained from the GMM belonging to each state of the monophone HMM for the speech feature quantity sequence and the appearance probability P (s) of each state s is obtained, and the highest From the logarithm of the product P (s ^) b _{s ^} (o _t ) or the logarithm of the output probability b _{s ^} (o _t ) and the GMM belonging to each state of the speech model or pose model HMM belonging to the state of the speech model for that input the difference between the logarithm of the highest output probability obtained b _{g ^} (o _t) and pre reliability of the frame, obtains a confidence score of the audio file units by averaging the pre-reliability, audio feature Using columns, performs speech recognition processing based on the confidence scores.

The speech recognition apparatus according to the present invention also includes a feature amount analysis unit that analyzes a speech feature amount of an input speech digital signal in units of frames and outputs a speech feature amount sequence, and inputs a speech feature amount sequence for each frame. And find the highest product of the output probability b _s (o _t ) obtained from the GMM belonging to each state of the monophone HMM for that input and the appearance probability P (s) of each state s. Obtained from the logarithm of P (s ^) b _{s ^} (o _t ) or the logarithm of output probability b _{s ^} (o _t ) and the GMM belonging to each state of the speech model or the pause model HMM for the input. the highest difference between the logarithm of the output probability b _{g ^} (o _t) and pre reliability of the frame to be pre reliability of outputting the confidence score of the audio file units by averaging the pre-reliability It includes a score calculation unit, an input voice feature amount sequence, and a voice recognition processing unit that performs speech recognition processing based on the confidence scores.

  In the present invention, before performing the speech recognition processing, the reliability of the speech recognition result obtained as a result of speech recognition is estimated in advance, and the speech recognition processing is performed based on the obtained reliability. As a result, it is possible to reduce the processing time for unavailable audio data. Further, it is possible to preferentially process voice data with high reliability, that is, voice data that can be expected to have high voice recognition accuracy, thereby improving the processing efficiency of the entire voice recognition process. Furthermore, since the language model is not used when obtaining the reliability, there is an effect that (prior) reliability that does not depend on the language model can be obtained.

The figure which shows the function structure of the conventional speech recognition apparatus 900 disclosed by patent document 1. FIG. The figure which shows an example of a phoneme model. The figure which shows typically 1 state which comprises a phoneme model. The figure which shows the function structural example of the speech recognition apparatus 100,200. The figure which shows the operation | movement flow of the speech recognition apparatus 100. The figure which shows the function structural example of the prior reliability score calculation parts 30 and 30 '. The figure which shows typically the time passage of a monophone maximum likelihood state and a speech / pause maximum likelihood state. The figure which shows the case where FIG. 7 is made into two types of acoustic models. The figure which shows an experimental result. The figure which shows the function structural example of the prior reliability score calculation part 230. FIG. The figure which shows typically the relationship between an audio | voice feature-value and likelihood (or output probability) in order to demonstrate the fundamental view of Example 2. FIG. The figure which shows the function structural example of the speech recognition apparatus 300. FIG. The figure which shows the example of the relationship between the reliability score C and beam search width | variety N (C). The figure which shows the function structural example of the speech recognition apparatus 400. FIG. The figure which shows the operation | movement flow of the speech recognition apparatus 400. The figure which shows the function structural example of the speech recognition apparatus 500. The figure which shows the operation | movement flow of the speech recognition apparatus 500.

Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated. Prior to the description of the embodiments, the basic concept of the present invention will be described.
[Basic concept of this invention]
A general reliability measure is expressed by the following word posterior probability P (W ^ | O).

O is an acoustic feature quantity sequence (O = (o ₁ , o ₂ ,..., O _T )), W is a speech recognition result word sequence, and P (W) is a speech obtained based on the speech recognition result. It is the appearance probability of the recognition result word sequence, and “^” indicates the word or state with the highest likelihood, and the word sequence or state sequence with the highest likelihood.

  Here, in order to obtain a speech recognition result word sequence W using a language model including a dictionary of large vocabularies, enormous calculations related to speech recognition processing are required. In order to reduce the amount of calculation, in the present invention, the language model is not used, and the state series S is used instead of the speech recognition result word series W. Therefore, the word posterior probability P (W ^ | O) is approximated by the following expression.

The type of state series S is preferably all state series that can be generated from all possible states s _j (where j = 1, 2,..., J), but in order to reduce the amount of calculation. Alternatively, a high-speed method used in speech recognition may be introduced, and a low possibility state may be excluded from the calculation target in advance.

Furthermore, in order to increase the speed, each state s _j in the state sequence S is limited to only states included in the monophone. Here, the monophone is an environment-independent phoneme model, which is a phoneme model that has no restriction on the preceding and following phonemes, compared to an environment-dependent phoneme model (for example, a triphone) that has restrictions on the preceding and following phoneme environments. There are few types. For example, when the 30 number of phonemes, the number of phoneme models in monophone acoustic model is 30, the number of cases of triphone is 30 ³ (27000 pieces). The monophone includes a pose model that is a model of a part other than the voice, that is, a non-voice part. A phone model of a monophone is constructed by a probability chain of one or more (usually about three) states, and is represented as a monophone HMM (Hidden Markov Model). The monophone HMM is represented by, for example, a left-to-right type HMM as shown in FIG. FIG. 2 shows three states s ₁ (first state), s ₂ (second state), and s ₃ (third state) arranged side by side, and the state probability chain (state transition) is self-transition. a ₁₁ , a ₂₂ , a ₃₃ and a ₁₂ , a ₂₃ , a ₃₄ to the next state. Each state s is composed of a mixed distribution composed of one or more basis distributions (hereinafter, mixed normal distribution GMM: mixed distribution including Gaussian Mixture Model). For example, as shown in FIG. Is expressed as The mixed normal distribution M is, for example, three (basic) normal distributions, N (μ _{s, 1} , Σ _{s, 1} ), N (μ _{s, 2} , Σ _{s, 2} ), N (μ _{s, 3} , Σ _{s). , 3} ). Here, μ _{s, m} represents an average vector of the normal distribution m belonging to the state s, and Σ _sm represents a covariance matrix of the normal distribution m belonging to the state s.

Furthermore, in order to reduce the amount of calculation in equation (2), in the same way as many speech recognition decoders ignore the transition probability (see Reference 1), the present invention also ignores the transition probability, and each monophone HMM The reliability for each frame is estimated using only the output probability obtained from the GMM belonging to the state (hereinafter simply referred to as “monophone GMM”).
[Reference 1] JR Glass, "A probabilistic framework for segmentbased speech recognition", Computer Speech and Language, Elsevier, 2003, Vol.17, No.2-3, pp.137-152
Therefore, the state posterior probability P of formula (2) (S ^ | O ) , the state posterior probability P of each frame with respect to the acoustic feature quantity _{o t} at time _t (s ^ | _o t) approximately as follows from Calculated.

T represents the total number of frames. Further, the state posterior probability P (s ^ | o _t ) for each frame is calculated for each frame from the output probability b _s (o _t ) of the state s as follows.

S ^ is a state when the value of P (s) · b _s (o _t ) is the highest at time _t (hereinafter referred to as “maximum likelihood state s ^”), and M _s is a mixed distribution belonging to the state s. W _{s, m} is a mixture weight coefficient of the normal distribution m, N _{s, m} (·) means a Gaussian distribution function of the normal distribution m, and N _{s, m} (o _t | μ _{s, m} sigma _{s, m)} denotes the output probability of the normal distribution m belonging to the state s for the acoustic feature quantity o _t at time t. Note that w _{s, m} is determined by the result of acoustic model learning, and takes a range of 0 ≦ w _{s, m} ≦ 1. For example, _assuming that the mixed distribution number _Ms is 16, the average value is 1/16.

In Reference Document 2, the speed is increased by reducing the amount of calculation of acoustic likelihood using a monophone based on the assumption that the monophone is an approximate model of a phoneme environment dependent model (triphone). Similarly, in the present invention, in the calculation of Expression (4), the speed is increased by using only a monophone.
[Reference 2] A. Lee, T. Kawahara, K. Shikano, "Gaussian mixture selection using context-independent HMM", in Proceedings of ICASSP, 2001, vol.1, pp.69-72
The denominator Σ _s P (s) b _s (o _t ) of Equation (4) is approximated as follows using a speech model composed of speech GMM learned from all phoneme features other than pauses.

g is a state belonging to the speech model, and is learned from the acoustic features of all phonemes, in other words, from all states. Here, if this speech model is constructed to have only one state g, the appearance probability P (g) of g is 1 in the speech frame. Therefore,

Therefore, from Equation (4) and Equation (6) ′, the state posterior probability P (s ^ | o _t ) for each frame is approximately calculated by the following equation.

Here, normally, since speech recognition uses a probability value converted to a logarithmic score region for calculation, the prior reliability c (o _t ) for each frame is approximately calculated for each frame obtained by Equation (7). The state posterior probability P (s ^ | o _t ) is assumed to be a logarithmic score area as in the following equation.

Pre-reliability c (o _t ) for each frame is the speaker verification as seen in Reference 3 when the speech model is considered as UBM (Universal Background Model) and the state appearance probability P (s ^) is ignored. This is equivalent to the logarithm of the likelihood ratio often used in. In the present invention, by introducing the state appearance probability P (s ^), the appearance frequency of the state and thus the appearance frequency for each phoneme is taken into consideration in the estimation of the maximum likelihood state s ^.
[Reference 3]
DA Reynolds, TF Quatieri, and RB Dunn, “Speaker verification using adapted gaussian mixture models,” Digital Signal Processing, 2000, vol.10, pp.19-41
The reliability score C is calculated from the prior reliability c (o _t ) for each frame. At that time, in order to enable comparison of audio data of different lengths, normalization is performed as follows according to the total number of frames T.

Based on such a concept, the present invention obtains a reliability score using a monophone and voice data without using a voice recognition result.
Hereinafter, embodiments of the present invention will be described in detail.

<Voice recognition apparatus 100>
The speech recognition apparatus 100 according to the first embodiment will be described with reference to FIGS. 4 and 5. The speech recognition apparatus 100 includes an A / D conversion unit 10, a feature amount analysis unit 20, a prior reliability score calculation unit 30, a speech recognition processing unit 40, an acoustic model parameter memory 50, a language model parameter memory 60, It comprises. The speech recognition apparatus 100 is realized by reading a predetermined program into a computer configured with, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

  The A / D conversion unit 10 converts the audio signal x (u) into a discrete value, for example, at a sampling frequency of 16 kHz and converts it into the audio digital signal x (v) (step S10). However, u represents continuous time and v represents discrete time. When the audio digital signal x (v) is directly input, the A / D converter 10 is not necessary.

The feature quantity analysis unit 20 receives the voice digital signal x (v) as an input, for example, sets 320 voice digital signals x (v) as one frame (for example, 20 ms), and for each frame, the voice feature quantity o _t. And an audio feature amount series O is output (step S20). As the audio feature amount, for example, dynamic parameters such as MFCC (Mel-Frequenct Cepstrum Coefficient) 1 to 12 elements, ΔMFCC which is the change amount, power, Δ power, and the like are used. Also, processing such as cepstrum average normalization (CMN) may be performed.

Pre confidence score calculation unit 30 is input with audio feature series O, probabilities of occurrence of output probability b _{s (o t)} and state s Field of the GMM obtained from monophones GMM for the speech feature quantity o _t for each frame The one having the highest product of P (s) (hereinafter referred to as “monophone maximum likelihood value P (s ^) b _{s ^} (o _t )") is obtained. Furthermore, pre confidence score calculator 30, the highest output probability obtained from GMM belonging to each state of the GMM or poses Model HMM belonging to the state of the speech model (hereinafter referred to as "speech / pause GMM") for the input o _t A thing (hereinafter referred to as “speech / pause maximum likelihood b _{g ^} (o _t )”) is obtained. As described above, this speech model is learned from the feature quantities of all phonemes other than the pose. Furthermore, the difference between the logarithm of the obtained monophone maximum likelihood value P (s ^) b _{s ^} (o _t ) and the logarithm of the speech / pause maximum likelihood value b _{g ^} (o _t ) is determined as the prior reliability c ( o _t ) (see equation (8)), the prior reliability c (o _t ) is averaged to obtain a reliability score C for each audio file and output (step S30).

  The speech recognition processing unit 40 receives the speech feature amount series O and the reliability score C and performs speech recognition processing based on the reliability score. For example, it is determined whether or not speech recognition processing is performed according to the reliability score C (step S40a), and when it is determined that speech recognition processing is performed, the acoustic model recorded in the acoustic model parameter memory 50, With reference to the language model recorded in the language model parameter memory 60, speech recognition processing is performed on the speech feature amount series O, and the speech recognition result W and the reliability score C are output (step S40b).

  Note that the voice recognition process in steps S40a and s40b is repeated until the process is completed for all frames of the voice file.

According to the speech recognition apparatus 100, the prior reliability score calculation unit 30 assigns the prior reliability c (o _t ) to each frame, averages the prior reliability (that is, calculates the average prior reliability per frame). And a reliability score C for each audio file is calculated. The reliability score C based on the speech feature amount series O requires a smaller amount of calculation than the conventional method of obtaining the reliability score from the speech recognition result. In addition, when processing a plurality of audio files, it is determined whether or not the audio recognition processing is performed according to the value of the reliability score C, so that the audio file with low reliability C, that is, with low audio recognition accuracy This also solves the problem that the voice recognition processing takes time. Next, a more specific configuration example of the prior reliability score calculation unit 30, which is a main part of the first embodiment, will be described in detail.

<Pre-reliability score calculation unit 30>
The prior reliability score calculation unit 30 will be described with reference to FIG. The prior reliability score calculation unit 30 includes a monophone maximum likelihood detection unit 32, a speech / pause maximum likelihood detection unit 33, a prior reliability calculation unit 34, and a reliability score calculation unit 35.

FIG. 7 schematically shows a time course of the output probability of the monophone, the output probability of the pause model, and the speech model. In the horizontal direction, the passage of time is represented by a frame t. The vertical direction represents the state of each of a plurality of monophones and voice models for each frame t. For example, each monophone consists of three states, and the monophone “* -a + *” consists of states a ₁ , a ₂ , and a ₃ . The thick circle state represents the monophone maximum likelihood state s ^ corresponding to the monophone maximum likelihood value P (s ^) b _{s ^} (o _t ). The state of the circle with diagonal lines represents the speech / pause maximum likelihood state 対 応 corresponding to the speech / pause maximum likelihood value b _＾ (o _t ). When the monophone maximum likelihood state ＾ and the speech / pause maximum likelihood state ＾ coincide (s ＾ = g ＾), the state is indicated by a thick circle with diagonal lines.

At times t _{1 to} t ₃ , the monophone maximum likelihood state s ^ is the first state p ₁ to the third state p ₃ of the pose model, respectively. Similarly, the speech / pause maximum likelihood state ＾ is the first state p ₁ to the third state p ₃ of the pose model, respectively. Therefore, the times t ₁ to t ₃ are in a non-voice state. For example, at the time t ₁ , using the expression (8), the appearance probability P (p ₁ ) of the first state p ₁ of the monophone “* -pause + *” and the output probability b _{p1 of the} GMM belonging to the state p ₁ ( The difference between the logarithm of the product of o _t1 ) and the logarithm of the output probability b _p1 (o _t1 ) of the GMM belonging to the state p _{1 of the} pose model is defined as a prior reliability c (o _t1 ). In other words, it is obtained as follows.
c (o _t1 ) = log (P (p ₁ ) b _p1 (o _t1 )) − logb _p1 (o _t1 )

At time t _4, monophones maximum likelihood state s ^ is the third state a ₃ of monophones "* -a + *", in the voice state from that voice / pause maximum likelihood state g ^ is the state g of voice model It is believed that there is. Using equation (8), the logarithm of the product of the appearance probability P (a ₃ ) of the third state a ₃ of the monophone “* -a + *” and the output probability b _a3 (o _t4 ) of the GMM belonging to the state a ₃ And the logarithm of the output probability b _g (o _t4 ) of the GMM belonging to the state g of the speech model is defined as a prior reliability c (o _t4 ). In other words, it is obtained as follows.
c (o _t4 ) = log (P (a ₃ ) b _a3 (o _t4 )) − logb _g (o _t4 )

In addition, at time _{t 19,} monophones maximum likelihood state s ^ is, monophones "* -i + *" is the second state _{i 2} of, is the third state _{p 3} of the voice / pause maximum likelihood state g ^ pose model . At this time, using the expression (8), the appearance probability P (i ₂ ) of the second state i ₂ of the monophone “* −i + *” and the output probability b _i2 (o _t19 ) of the GMM belonging to the state i ₂ The difference between the logarithm of the product and the logarithm of the output probability b _p3 (o _t19 ) of the GMM belonging to the third state p ₃ of the pose model is defined as a prior reliability c (o _t19 ). In other words, it is obtained as follows.
c (o _t19 ) = log (P (i ₂ ) b _i2 (o _t19 )) − logb _p3 (o _t19 )
FIG. 7 shows only a part of the time. The length of the audio file is, for example, about several minutes (for example, 30,000 frames). Hereinafter, the processing of each means will be specifically described.

(Monophone maximum likelihood detection means 32)
Monophones maximum likelihood detecting means 32, output probability obtained from each monophones GMM for the speech feature quantity _{o t} for each frame t _b s _{(o t)} and the product P of the probability P (s) of the GMM belongs state s ( _s) b s (from _{o t),} monophones maximum likelihood value _{P (s ^) b s ^} (o t) the request, the log-log (P (s ^) bs ^ (ot)) a pre-reliability calculation means 34. Note that the monophone maximum likelihood detection unit 32 can obtain the appearance probability P (s) of each monophone GMM and each state s with reference to the acoustic model parameter memory 50. In addition, the monophone maximum likelihood detection unit 32 may acquire and store each monophone GMM and the appearance probability P (s) of each state s from the acoustic model parameter memory 50 in advance.

  It should be noted that the appearance probability P (s ^) of the monophone maximum likelihood state s ^ is assumed that there is no difference between the appearance probability of each state in the learning data of the acoustic model and the evaluation speech data that is the target speech recognition target. Approximately, the following equation (10) may be used.

The denominator of Equation (10) represents the sum of the appearance frequencies of each state s in the acoustic model learning data, and the numerator represents the appearance frequency of the maximum likelihood state s ^ in the acoustic model learning data. If the expected value Γ (s) of the appearance frequency of each state s obtained at the time of learning the acoustic model is stored in the acoustic model parameter memory 50, it can be easily realized by using it.

(Voice / pause maximum likelihood detection means 33)
Voice / Pause maximum likelihood detecting means 33 obtains the output probability obtained from the speech / pause GMM, audio / pause maximum likelihood value _b g ^ a _{(o t)} for the audio feature amount _{o t} for each frame t, the logarithm logb _{g ^} (o _t ) is output to the prior reliability calculation means 34. The voice / pause maximum likelihood detection means 33 can acquire the voice / pause GMM with reference to the acoustic model parameter memory 50. The voice / pause maximum likelihood detection means 33 may acquire the voice / pause GMM from the acoustic model parameter memory 50 in advance and store it.

(Advance reliability calculation means 34)
Prior reliability calculation means 34 is input with logarithm log (P (s ^) b _{s ^} (o _t )) of monophone maximum likelihood value and logarithmic logb _{g ^} (o _t ) of speech / pause maximum likelihood value, The difference is obtained as the prior reliability c (o _t ) of the frame by the following equation (11), and is output to the reliability score calculation means 35.

(Reliability score calculation means 35)
The reliability score calculation means 35 receives the prior reliability c (o _t ) for each frame, averages the prior reliability c (o _t ) for each frame according to the equation (9), and is based on the audio file unit ( In other words, the prior reliability c (o _t ) obtained by accumulating and averaging the duration T (total number of frames) of the audio file is obtained as the reliability score C and output.

Thus, pre confidence score calculator 30, a confidence score C representing the reliability of the audio file units by averaging the frame advance confidence c a (o _t) by the total number of frames T audio file calculate. In addition, since the reliability score C for each audio file is obtained, no elaborate processing is required. Further, with such a configuration, the prior reliability calculation can be performed at a stable processing speed without changing the processing speed due to the quality of the input voice signal, consistency with the acoustic model, and the like. Next, details of the voice recognition processing unit 40 will be described.

<Voice recognition processing unit 40>
The speech recognition processing unit 40 receives the speech feature amount series O (= o ₁ , o ₂ ,..., O _T ) output from the feature amount analysis unit 20 and the reliability score C, and receives the acoustic model parameter memory 50 and the language model. With reference to the parameter 60, voice recognition processing is performed, and a voice recognition result W is output. At this time, the reliability score C may be output simultaneously. In this speech recognition process, a recognition process using all acoustic models recorded in the acoustic model parameter memory 50 is performed. The voice recognition processing unit 40 switches the execution of the voice recognition process according to the value of the reliability score C.

For example, the speech recognition processing unit 40 stops the speech recognition process when the reliability score C is equal to or less than a certain value _Cth . Since the reliability score C is a value calculated for each voice file, the voice recognition processing unit 40 switches whether or not the voice recognition process is executed for each voice file. For example, a method of calculating the constant value C _th from the reliability score distribution for the learning data of the acoustic model is conceivable. When the average value μ and the standard deviation σ of the reliability score distribution are used, for example, C _th = μ−2σ. For example, the voice recognition processing unit 40 obtains and accumulates reliability scores C of a plurality of voice files, and stores the top N voice files (for example, the number corresponding to 20% of all voice recognition target voice files). Only a voice recognition process may be performed.

<Effect>
As described above, according to the speech recognition apparatus of the present invention, the prior reliability based on the speech feature amount is obtained, the prior reliability for each frame is averaged, and the reliability score for each audio file is calculated. Therefore, the reliability score is obtained by processing that is lighter than that of the conventional speech recognition apparatus. Further, since the processing is based on the speech feature amount, a reliability score that does not depend on the language model can be obtained. Further, by determining whether or not to perform speech recognition processing according to the value of the obtained reliability score, for example, for speech recognition processing of a speech file with low speech recognition accuracy due to a poor S / N ratio or the like. It can solve time-consuming problems. In contrast to the conventional calculation of reliability in units of words or keywords, or in units of utterances (sentences), the speech recognition apparatus 100 of the present embodiment has a reliability score in units of audio files composed of a plurality of utterances. Can be calculated.

<Modification 1>
Only parts different from the first embodiment will be described with reference to FIGS. 4 and 6. The processing content of the prior reliability score calculation unit 30 ′ is different from that of the first embodiment.

<Pre-reliability score calculation unit 30 '>
The prior reliability score calculation unit 30 ′ includes a monophone maximum likelihood detection unit 32 ′, a speech / pause maximum likelihood detection unit 33 ′, a prior reliability calculation unit 34, and a reliability score calculation unit 35. The processing of the monophone maximum likelihood detection means 32 ′ and the voice / pause maximum likelihood detection means 33 ′ is different from that of the first embodiment. The prior reliability score calculation unit 30 ′ calculates the reliability score C for each audio file by averaging the prior reliability for each frame calculated based on the monophone and the audio model included in the two or more types of acoustic models. . FIG. 8 shows an example of the time course of output probability when two or more kinds of acoustic models are a male acoustic model and a female acoustic model.

(Monophone maximum likelihood detection means 32 ')
Monophones maximum likelihood detecting means 32 'is first output probability obtained from GMM belonging to the state _{s m} Male monophone HMM for speech features _{o t} for each frame t (hereinafter referred to as "male monophones GMM") _b sm _{(o t} ) And the appearance probability P (s _m ) of the state s _m to which the GMM belongs, P (s _m ) b _sm (o _t ), the highest value (hereinafter referred to as “male monophone maximum likelihood value P (s ^ _m )” b _{s ^ m} (o _t ) ”). Then, the probability of occurrence of the output probability _b sf _{(o t)} and its GMM belongs state _{s f} obtained from GMM belonging to the state _{s f} of women monophones HMM for the speech feature quantity _{o t} (hereinafter referred to as "women monophones GMM") from P product _{P (s f) b sf (} o t) of _{(s f),} the highest value (hereinafter referred to as "women monophones maximum likelihood value _{P (s ^ f) b s} ^ f (o t) " hereinafter) a Ask. Men monophones maximum likelihood value _{P (s ^ m) b s} ^ m (o t) and women monophones maximum likelihood value _{P (s ^ f) b s} ^ f (o t) of the, the larger the monophones maximum likelihood value P (S ^) b _{s ^} (o _t ) and the logarithm thereof is output to the prior reliability calculation means 34.

(Speech / pause maximum likelihood detection means 33 ')
Voice / Pause maximum likelihood detection unit 33 ', the output probability is first obtained from a male voice / pause GMM for the speech feature quantity o _t for each frame t, male voice / pause maximum likelihood value b _{g ^ m} a _{(o t)} Ask. Then, first, the output probability obtained from female voice / pause GMM for the speech feature quantity _{o t} for each frame t, obtaining the female voice / pause maximum likelihood value _{b g ^ f (o t)} . Of the male voice / pause maximum likelihood value b _{g ^ m} (o _t ) and the female voice / pause maximum likelihood value b _{g ^ f} (o _t ), the larger one is the voice / pause maximum likelihood value b _{g ^} (o _t ). And the logarithm is output to the prior reliability calculation means 34.

The prior reliability calculation means 34 calculates the logarithm log (P (s ^) b _{s ^} (o _t )) of the monophon maximum likelihood value and the logarithmic logb _{g ^} (o _t ) of the speech / pause maximum likelihood value to obtain the equation (11). The difference is obtained as the prior reliability c (o _t ) of the frame. The reliability score calculation means 35 receives the prior reliability c (o _t ) for each frame, averages the prior reliability c (o _t ) for each frame according to the equation (9), and obtains the one for each audio file. Obtained as a reliability score C.

  With such a configuration, even when the subsequent speech recognition process uses a plurality of acoustic models, the reliability score C is recognized by using a plurality of types of acoustic models in the prior reliability score calculation. It can be accurately determined according to the processing. Note that there may be three or more types of acoustic models used in the prior reliability score calculation unit 30 ′.

The reliability score C is a value calculated by comparing the output probabilities of the maximum likelihood states of two or more types of speech models or pause models with respect to the speech feature amount series, and limited to monophones of a type having a large output probability. May be. In other words, all of the frame maximum likelihood value _P of monophones of men and women _{(s ^ m) b s ^} m and _{(o t) P} the _{(s ^ f) b s ^} f (o t) as of the above-mentioned example The frame in which the output probability of the speech model or the pose model is higher for the male (female) than the female (male) may be limited to the male (female) monophone.

That is, the speech / pause maximum likelihood detection means 33 ″ uses the larger one of the speech / pause maximum likelihood values b _{g ^ m} (o _t ) and b _{g ^ f} (o _t ) of the male and female voice / pause maximum likelihood. The value b _{g ^} (o _t ) is set, and the monophone maximum likelihood detection unit 32 ″ receives the determination result as an input, and either of the monophone maximum likelihood values P (s ^) b _{s ^} (o _t ) In this example, since the product P (s) b _s (o _t ) of the output probability b _s (o _t ) of all types of monophones and the appearance probability P (s) of the state is not calculated, the amount of calculation is reduced. Can be expected.

<Other variations>
A speech section determination unit (not shown) may be provided in the preceding stage of the feature amount analysis unit 20. For example, the voice section determination unit determines that the voice section is not a voice section when a frame whose power is equal to or lower than a predetermined value continues for a predetermined time or longer. And when it determines with a non-voice area, an instruction | indication signal is output so that the process after that with respect to the area may be stopped. With such a configuration, it is possible to omit the speech recognition process in the non-speech section. Note that large noise or the like cannot be omitted by the speech segment determination unit, but the monophone maximum likelihood detection unit 32 and the speech / pause maximum likelihood detection unit 33 determine whether the speech is speech or non-speech (pause). Can be prevented.

Since the appearance frequency or the appearance probability of each state s used in the monophone maximum likelihood detection means 32 is not used in the actual speech recognition processing, there is also an acoustic model parameter memory 50 that does not hold this information. In that case, all the appearance frequencies may be set to 1 (P (s) = 1), and the prior reliability c (o _t ) for each frame may be obtained by Expression (8). There is also an acoustic model parameter memory 50 in which appearance frequencies or appearance probabilities are stored only for some states. In that case, the average value of the appearance frequency or appearance probability of some stored states is obtained, and the obtained average value is used as the appearance frequency of other states (states where the appearance frequency or appearance probability is not stored) or You may substitute as an appearance probability.

Moreover, when using a some acoustic model in a prior reliability score calculation part, it is good also as a structure which estimates an utterance area and estimates the optimal acoustic model for every utterance area. For example, as in Reference Document 4, a gender is estimated in advance using a voice / pause GMM, and an acoustic model (male acoustic model or female acoustic model) that matches the estimated gender is used.
[Reference 4] S. Kobashikawa, A. Ogawa, Y. Yamaguchi, and S. Takahashi, “Rapid unsupervised adaptation using frame independent output probabilities of gender and context independent phoneme models”, INTERSPEECH, 2009, pp.1615-1618.
The monophone maximum likelihood detection means 32 and the speech / pause maximum likelihood detection means 33 replace the logarithm log (P (s ^) b _{s ^} (o _t )) and logarithm logb _{g ^} (o _t ), respectively. ^) B _{s ^} (o _t ) and b _{g ^} (o _t ) are output, and the prior reliability calculation means 34 outputs the logarithm log (P (s ^) b _{s ^} (o _t )) and logb _{g ^} ( o _t ) may be determined.

  Note that the processes described in the method and apparatus are not only executed in time series according to the order of description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Good.

<Experimental result>
The acoustic analysis conditions of this experiment are a sampling frequency of 16 kHz, a Hamming window with a window width of 20 msec, a window shift of 10 msec, a feature amount of 25 dimensions (MFCC12, ΔMFCC12, ΔPOWER), and an evaluation task of 48 speakers (male) A total of 240 calls (total of 19.81 hours, 17,672 utterances) by 17 people and 31 women), the utterance contents are free utterances in a one-to-one dialogue. The acoustic model is an unspecified speaker model by gender, the total number of states is 1,958, the total number of distributions is 26,567 men and 29,836 women. Gender selection was performed in advance using voice / pause GMM as in Reference 4. The language model is a word trigram constructed based on the transcription of dialogue voice, and the vocabulary size is 59,676 words. A voice recognition engine VoiceRex (see Reference 5) was used as the decoder.
[Reference 5] H. Masataki, D. Shibata, Y. Nakazawa, S. Kobashikawa, A. Ogawa, and K. Ohtsuki, “VoiceRex-Spontaneous speech recognition technology for contact-center conversations,” NTT Tech. Rev., 2007, vol. 5, no. 1, pp. 22-27

In order to show the effectiveness of the proposed recognition target data selection based on the pre-reliability estimation, the average recognition rate (character unit) of the selected call voice against the data selection rate per call unit is evaluated, and the ideal condition: recognition accuracy is Ideal conditions selected in descending order, average recognition rate: average recognition rate of all call voices used in the experiment, conventional technology: selection in descending order of reliability score using speech recognition results after speech recognition processing, proposed technology : Comparison was made under the four conditions of selection in the order of proposed prior reliability (voice recognition apparatus 100 of Example 1). Moreover, the prior art employ | adopted the method which estimates the reliability based on N best of the speech recognition result like the reference literature 6. FIG. Furthermore, the speed of prior reliability estimation was evaluated by comparison with the prior art including speech recognition processing.
[Reference 6] B. Rueber, “Obtaining confidence measures from sentence probabilities”, In EUROSPEECH-1997, pp.739-742

  The effect of selecting recognition target data by the proposed method is shown in FIG. Although the speech recognition apparatus 100 according to the first embodiment does not satisfy the ideal condition, the recognition rate is higher than the average recognition rate at all selection rates, and the selection has an effect on the improvement of the recognition rate. Furthermore, it showed the same performance as the method based on the ex-post reliability after speech recognition processing (prior art). In addition, the processing time of the prior reliability estimation is only 0.0184 in comparison with the prior art, realizing a speed improvement of 50 times or more. When all call voices cannot be recognized and processed under limited computing resources, the selection based on the a posteriori reliability as shown in FIG. 9 cannot be realized. Therefore, the selection based on the proposed prior reliability is effective. It can be said.

  In this paper, we proposed a method for selecting recognition target speech data before speech recognition based on fast prior reliability estimation using environment-independent phoneme model and speech model. As a result of the experiment, the same selection performance was realized at a speed about 54 times faster than the ex-post reliability estimation after the speech recognition processing.

<Voice recognition apparatus 200>
Only the parts different from the first embodiment of the speech recognition apparatus 200 according to the second embodiment will be described with reference to FIG. The speech recognition apparatus 200 is different from the first embodiment in the processing content of the prior reliability score calculation unit 230.

<Pre-reliability score calculation unit 230>
The prior reliability score calculation unit 230 will be described with reference to FIG. The prior reliability score calculation unit 230 includes a monophone maximum likelihood detection unit 232, a speech / pause maximum likelihood detection unit 33, a prior reliability calculation unit 234, and a reliability score calculation unit 35, and the monophone maximum likelihood detection unit. Processing contents of the H.232 and the prior reliability calculation means 234 are different from those in the first embodiment.

(Monophone maximum likelihood detection means 232)
Monophones maximum likelihood detection unit 232, the output probability obtained from each monophones GMM for the speech feature quantity _{o t} for each frame t _b s _{(o t)} and the product P of the probability P (s) of the GMM belongs state s ( _s) b s (from _{o t),} monophones determine the maximum likelihood value _{P (s ^) b s ^} (o t), the output probability obtained from GMM belonging to monophones maximum likelihood state _{s ^ b s ^ (o t} ) Loglog _{s ^} (o _t ) is output to the prior reliability calculation means 34.

(Advance reliability calculation means 234)
The prior reliability calculation means 234 calculates the logarithm logb _{s ^} (o _t ) of the output probability b _{s ^} (o _t ) obtained from the GMM belonging to the monophone maximum likelihood state s ^ and the log logb _{g ^ of the} speech / pause maximum likelihood value. (O _t ) is input, and the difference is obtained as the prior reliability c (o _t ) of the frame by Expression (12), and is output to the reliability score calculation means 35.

Even if the equation (12) is used instead of the equation (11), the prior reliability c (o _t ) can be obtained as in the first embodiment.

  In addition, it turns out that Formula (12) is effective as prior reliability also from the following viewpoints. FIG. 11 shows the relationship between the voice feature quantity and the likelihood. Likelihood is a value representing likelihood, and an output probability value may be substituted. The horizontal axis is the voice feature amount, and the vertical axis is the likelihood. In the figure, the respective distributions of the speech model (broken line) and monophone phoneme models “* -a + *”, “* -i + *”, “* -u + *” included in the acoustic model are shown. Note that-represents left side dependence, + represents right side dependence, and * represents any phoneme. In FIG. 11, the number of states of the phoneme model is represented as 1 and the number of mixture distributions is represented as 1 for simplification.

  The GMM used for the speech model is a model learned based on learning data of all speech, that is, all phonemes. Therefore, the distribution is a distribution in which the likelihood value with respect to the voice feature amount is relatively gentle. On the other hand, the monophone is a model learned from learning data of each phoneme. Therefore, the likelihood value for the speech feature amount corresponding to the phoneme has a steep distribution.

Therefore, it is possible to determine the reliability of the audio file by comparing the likelihood of the audio model for a certain audio feature amount and the likelihood of the monophone for the same audio feature amount. In other words, the likelihood of monophones "* -a + *" for the audio feature of the phoneme a which was recorded without being affected by the noise ^{_{o t clean (a) b s}} (o t clean (a)) shows a large value . On the other hand, the likelihood _b g of speech model for the same speech features ^{_{o t clean (a) (o}} t clean (a)) shows a relatively small value. As a result, there is a large difference between these values.

On the other hand, the audio feature amount of the phoneme a which was recorded strongly affected by the noise o _t ^noisy (a) is, the likelihood b _{s (o t} ^noisy _(a in monophones is different from the original feature value )) and the difference between the likelihood ^{_{b g (o t noisy (a}} )) in the speech model is reduced.

Thus the speech features for monophone likelihood b _{s (o t),} by looking at the difference between the speech model likelihood b _{g (o t),} it is possible to evaluate the quality of the recorded voice. Therefore, it can be seen that the prior reliability c (o _t ) can be obtained from the equation (12).

By adopting such a configuration, the same effect as in the first embodiment can be obtained. Further, in the expression (11) used in the first embodiment, since the first term includes the appearance probability P (s ^) (<1) of the maximum likelihood state s ^, the value of the prior reliability c (o _t ) is small. This is likely to be a negative region. In the expression (12) used in the second embodiment, the first term and the second term are similar log probabilities of output probabilities, and the distribution of the speech model is gentle compared to the distribution of the monophone as described above. The value of the second term is smaller than that of the first term and is likely to be a positive region. In other words, the range of possible values of the value of the prior reliability c (o _t ) and the reliability score C is limited. Therefore, when performing voice recognition processing control at a later stage, setting of the threshold value C _th for controlling the voice recognition processing becomes easy.

<Voice recognition apparatus 300>
A speech recognition apparatus 300 according to the third embodiment will be described with reference to FIG. The speech recognition apparatus 300 is different from the speech recognition apparatuses 100 and 200 in that it includes a recognition processing control unit 380 and the processing content of the speech recognition apparatus 340.
<Recognition processing control unit 380>
The recognition processing control unit 380 outputs the beam search width N (C) as a control signal. An example is shown in equation (13).

FIG. 13 illustrates the relationship between the reliability score C and the beam search width N (C). The horizontal axis is the reliability score C, and the vertical axis is the beam search width N (C).
As shown in FIG. 13, the equation (13) expresses the beam search width N (C) (N _{min to} N _max ) corresponding to the reliability score C (C _{min to} C _max ) in a predetermined range, and the reliability score C It is an idea of proportionally distributing with the value of. Here, since the proportionality coefficient is a negative value, the reliability score C is small and the beam search width N (C) is large, and C is large and N (C) is small. Of course, the relationship between the reliability score C and the beam search width N (C) may be a relationship that can be expressed by a non-linear function. Further, when the beam search width N (C) is used as the control signal, the beam search width is not limited to the number beam width. For example, the score search width, the word end score beam width, the word end number beam width, etc. There may be.

Here, for example, C _max = μ + σ, C _min = μ−σ, N _max may be 1.5 times the beam width normally used, N _min may be half the beam width normally used, and the like. In addition, when the average sound quality is extremely low (for example, C <C _min ), even if the beam search width is increased, the accuracy cannot be improved and it takes much processing time. Therefore, the beam search width is reduced, for example, N _min. good. Further, the speech recognition process may not be performed by including the non-recognition instruction signal in the control signal. Further, a signal for stopping the speech recognition process and a control signal for the beam search width may coexist.

<Voice recognition processing unit 340>
The speech recognition processing unit 340 performs speech recognition processing using the speech feature amount series as an input based on at least one of a signal for stopping the speech recognition processing or a control signal for beam search width. For example, when a signal for stopping the voice recognition process is received from the recognition process control unit 380, the voice recognition process is stopped for the corresponding voice file. When a control signal for the beam search width N (C) is received, speech recognition processing is performed based on the beam search width N (C).

<Effect>
As described above, the speech recognition apparatus 300 including the recognition processing control unit 380 can improve the efficiency of the speech recognition processing of a plurality of audio files and improve the recognition accuracy. Note that the speech recognition processing unit 40 may have the function of the recognition processing control unit 380.

<Voice recognition apparatus 400>
A speech recognition apparatus 400 according to the fourth embodiment will be described with reference to FIGS. 14 and 15.
The voice recognition apparatus 400 is different from the voice recognition apparatuses 100 and 200 in that the voice recognition apparatus 400 includes a voice file processing unit 401 and a sort voice recognition processing unit 440.

<Audio file processing unit 401>
The audio file processing unit 401 rearranges the plurality of audio files in descending order of the reliability score C of the plurality of audio files (step S401).
<Sort voice recognition processing unit 440>
The sorted speech recognition processing unit 440 performs speech recognition processing in descending order of the reliability score C (step S440).

<Effect>
By adopting such a configuration, the same effect as in the first embodiment can be obtained. Furthermore, by executing the speech recognition processing in the order of the reliability score C in this way, it is possible to improve the processing efficiency when performing speech recognition processing of a plurality of speech files. For example, when it is difficult to perform speech recognition processing on all speech files due to the relationship between computer resources and processing time, speech recognition processing is not performed on speech files with a low reliability score C, and speech recognition accuracy is high. The voice recognition process is performed only on a voice file having a high reliability score C that is expected to be high, and it is possible to collect a highly accurate voice recognition result. Note that the function of the voice file processing unit 401 may be included in the function of the sort voice recognition processing unit 440. Note that the same effect can be obtained by combining the speech recognition apparatus 300 according to the third embodiment with the speech file processing unit 401 and the sorted speech recognition processing unit 440.

<Voice recognition apparatus 500>
A speech recognition apparatus 500 according to the fifth embodiment will be described with reference to FIGS. 16 and 17.
The speech recognition apparatus 500 differs from the speech recognition apparatuses 100 and 200 in that it includes an unsupervised adaptive control unit 501, an unsupervised adaptation unit 502, a post-adaptation acoustic model parameter memory 503, and a second recognition processing unit 504.

<Unsupervised adaptive control unit 501>
The unsupervised adaptive control unit 501 receives the prior reliability C, and the value of the prior reliability C is within a certain range (for example, C> C _th2 and C _th2 > C _th , where C _th2 is Then, using the above-described average value μ and standard deviation σ of the reliability score distribution, it is determined whether or not C _th2 = μ−σ, for example, and the unsupervised adaptive control signal p is output (step S501). ). If the value of the prior reliability C is not within a certain range, the processing of the audio file is terminated (NO in step S501). The unsupervised adaptive control signal is a signal for controlling whether or not the speech recognition result output from the speech recognition processing unit 40 is used as an adaptation label.

<Unsupervised adaptation unit 502>
The unsupervised adaptation unit 502 uses the speech recognition result W as an adaptation label when the unsupervised adaptation control signal p instructs to use the speech recognition result W output from the speech recognition processing unit 40 as an adaptation label. The acoustic model recorded in the acoustic model parameter memory 50 is learned, and an adaptive acoustic model is generated (step S502). The post-adaptation acoustic model is recorded in the post-adaptation acoustic model parameter memory 503.

<Second recognition processing unit 504>
The second recognition processing unit 504 performs speech recognition processing of the speech feature amount series O using the post-adaptation acoustic model recorded in the post-adaptation acoustic model parameter memory 503, and outputs a speech recognition result W ′ (step S504). . At this time, the reliability score C obtained by the prior reliability score calculation unit 30 may be output together.

<Effect>
By adopting such a configuration, the same effect as in the first embodiment can be obtained. Furthermore, the speech recognition apparatus 500 learns an acoustic model using the speech recognition result W as an adaptation label and performs speech recognition processing only when the value of the prior reliability C is within a certain range. When the prior reliability score C is low and the recognition accuracy of the voice file is low, the voice recognition processing result W at that time is not suitable as an adaptation label in unsupervised adaptation, and an improvement in the accuracy of the acoustic model by unsupervised adaptation is expected. Can not. In such a case, the calculation time can be reduced by omitting the unsupervised adaptation and the second speech recognition process. In addition, since the acoustic model is learned using the speech recognition result W having the high reliability score C and the recognition accuracy of the speech file as the adaptive label, the accuracy of the acoustic model can be automatically improved. The same effect can be obtained by combining the speech recognition apparatuses 300 and 400 according to the third and fourth embodiments, the unsupervised adaptive control unit 501, the unsupervised adaptation unit 502, the post-adaptation acoustic model parameter memory 503, and the second recognition processing unit 504. You can

<Program>
Further, when the processing means in the device is realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is 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. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can 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 recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

100, 200, 300, 400, 500, 900 Speech recognition device 20 Feature amount analysis unit 30, 30 ′, 230 Prior reliability score calculation unit 40, 340 Speech recognition processing unit 50 Acoustic model parameter memory 60 Language model parameter memory 380 Recognition Processing control unit 401 Audio file processing unit 440 Sorted speech recognition processing unit 501 Unsupervised adaptive control unit 502 Unsupervised adaptation unit 503 Adaptive acoustic model parameter memory 504 Second recognition processing unit

Claims (11)

A feature amount analysis process for obtaining a speech feature amount sequence by analyzing a speech feature amount of a speech digital signal in units of frames;
Using the speech feature amount sequence for each frame, the output probability b _s (o _t ) obtained from the GMM belonging to each state of the monophone HMM for the speech feature amount sequence, and the appearance probability P (s) of each state s Is obtained from the logarithm of the highest product P (s ^) b _{s ^} (o _t ) and the GMM belonging to each state of the speech model or pose model HMM for the input speech model. A prior reliability score calculation process in which a difference from the logarithm of the highest obtained output probability b _{g ^} (o _t ) is used as the prior reliability of the frame, and the prior reliability is averaged to obtain a reliability score for each audio file. When,
Using the speech feature amount sequence, a speech recognition processing step of performing speech recognition processing based on the reliability score,
A speech recognition method characterized by the above. A feature amount analysis process for obtaining a speech feature amount sequence by analyzing a speech feature amount of a speech digital signal in units of frames;
The product of the output probability b _s (o _t ) obtained from the GMM belonging to each state of the monophone HMM for the input and the appearance probability P (s) of each state s using the speech feature amount sequence for each frame. Is the logarithm of the output probability b _{s ^} (o _t ) when is the highest and the highest output probability b _{g ^} (o obtained from the GMM belonging to each state of the speech model HMM or the pose model HMM for the input. a prior reliability score calculation process in which a difference between the logarithm of _t ) is defined as a prior reliability of the frame, and the prior reliability is averaged to obtain a reliability score for each audio file;
Using the speech feature amount sequence, a speech recognition processing step of performing speech recognition processing based on the reliability score,
A speech recognition method characterized by the above. In the voice recognition method according to claim 1 or 2,
A recognition processing control step for obtaining at least one of a signal for stopping speech recognition processing or a control signal for beam search width using the reliability score ;
The speech recognition processing step performs speech recognition processing of the speech feature amount series based on at least one of a signal for stopping speech recognition processing or a control signal for beam search width.
A speech recognition method characterized by the above. The speech recognition method according to any one of claims 1 to 3,
From the confidence scores of the plurality of audio files, and the audio file processing step of rearranging a plurality of audio files with high confidence score order,
Sort voice recognition processing process that performs voice recognition processing in order of high reliability score ;
A speech recognition method, further comprising: In the voice recognition method according to any one of claims 1 to 4,
Using said confidence score, and unsupervised adaptive control process of obtaining the no adaptive control signals teacher determines whether the value of the confidence scores or within a predetermined range,
Using the speech recognition result and the unsupervised adaptive control signal, learning an acoustic model using the speech recognition result as an adaptation label and generating an adapted acoustic model;
A second recognition process for performing speech recognition processing of the speech feature amount sequence using the post-adaptation acoustic model when the post-adaptation acoustic model is generated;
A speech recognition method, further comprising: A feature amount analysis unit that analyzes a speech feature amount of an input speech digital signal in units of frames and outputs a speech feature amount sequence;
The product of the output probability b _s (o _t ) obtained from the GMM belonging to each state of the monophone HMM with respect to the input speech feature quantity sequence for each frame and the appearance probability P (s) of each state s Which is obtained from the logarithm of the highest product P (s ^) b _{s ^} (o _t ) and the GMM belonging to each state of the speech model or pose model HMM for the input. A prior reliability score calculator that outputs a reliability score for each audio file by averaging the prior reliability as a difference between the logarithm of the high output probability b _{g ^} (o _t ) and the prior reliability of the frame;
A speech recognition processing unit that performs speech recognition processing based on the reliability score using the speech feature amount series as an input;
A speech recognition apparatus characterized by that. A feature amount analysis unit that analyzes a speech feature amount of an input speech digital signal in units of frames and outputs a speech feature amount sequence;
The product of the output probability b _s (o _t ) obtained from the GMM belonging to each state of the monophone HMM with respect to the input speech feature quantity sequence for each frame and the appearance probability P (s) of each state s Is the logarithm of the output probability b _{s ^} (o _t ) when is the highest and the highest output probability b _{g ^} (o obtained from the GMM belonging to each state of the speech model HMM or the pose model HMM for the input. a prior reliability score calculation unit that sets a difference between the logarithm of _t ) as a prior reliability of the frame, averages the prior reliability, and outputs a reliability score for each audio file;
A speech recognition processing unit that performs speech recognition processing based on the reliability score using the speech feature amount series as an input;
A speech recognition apparatus characterized by that. In the voice recognition device according to claim 6 or 7,
A recognition processing control unit for obtaining at least one of a signal for stopping speech recognition processing or a control signal for beam search width, using the reliability score as an input;
The speech recognition processing unit performs speech recognition processing using the speech feature amount series as an input based on at least one of a signal for stopping speech recognition processing or a control signal for beam search width,
A speech recognition apparatus characterized by that. The speech recognition apparatus according to any one of claims 6 to 8,
From the confidence scores of the plurality of audio files, and the audio file processing unit that rearranges the plurality of audio files with high confidence score order,
Sort voice recognition processing unit that performs voice recognition processing in order of high reliability score ;
A speech recognition apparatus further comprising: The speech recognition apparatus according to any one of claims 6 to 9,
Wherein the confidence scores as inputs, and unsupervised adaptive controller that outputs without adaptive control signals teacher determines whether the value of the confidence scores or within a predetermined range,
An unsupervised adaptation unit that receives the speech recognition result and the unsupervised adaptive control signal as input, learns an acoustic model using the speech recognition result as an adaptation label, and generates a post-adaptation acoustic model;
A second recognition processing unit for performing speech recognition processing of the speech feature amount sequence using the post-adaptation acoustic model when the post-adaptation acoustic model is generated;
A speech recognition apparatus further comprising:   A program for causing a computer to execute the speech recognition method according to any one of claims 1 to 5.

Images