JP4297349B2 - Speech recognition system - Google Patents

Description

  The present invention relates to a speech recognition system, and more particularly to a speech recognition system that can sequentially output speech recognition result candidates during speech production.

  An example of a conventional voice recognition system (conventional method 1) will be described with reference to FIG. The voice input unit 1 inputs a voice signal which is a user's voice and sends it to the voice detection unit 2. Next, the voice detection unit 2 cuts out the voice signal obtained from the voice input unit 1 from the beginning of the voice section sequentially by shifting by ΔT for each frame length T as shown in FIG. send. In addition, an end detection signal is sent to the recognition result candidate determination unit 6 at the time when the end of the speech section is detected. The acoustic analysis unit 3 extracts an acoustic parameter representing the characteristics of the audio signal from the audio signal cut out for each frame, and sends it to the matching unit 4. The collation unit 4 collates the recognized vocabulary standard pattern 5 obtained by concatenating models representing the features of phonemes and the acoustic parameters from the first frame to the current time, and determines the likelihood of the phoneme string constituting the recognized vocabulary. calculate. Note that the likelihood at each time can be calculated using the Viterbi algorithm based on the result of matching up to one hour before. This is described in, for example, “Recognition of Speech by Stochastic Model” by Seiichi Nakagawa. The recognition result candidate determination unit 6 outputs a phoneme string hypothesis having a high likelihood as a recognition result candidate at the time when the termination detection signal is received from the voice detection unit 2.

  Next, another example of the conventional system (conventional method 2) will be described with reference to FIG. The same reference numerals are given to the same or equivalent parts in FIG. In this conventional example, the functions of the voice input unit 1, the voice detection unit 2, the acoustic analysis unit 3, and the collation unit 4 are the same as those in FIG. 3 is different from the conventional example in FIG. 3 in that a phoneme string having a high likelihood is output from the matching unit 4 as a recognition result candidate at each time.

As documents describing the technology according to the prior art systems, for example, Patent Documents 1 and 2 below.
JP 2003-255972 A JP 2003-345386 A July 1, 1988, Corona Inc. Luminescence, Seiichi Nakagawa "Speech Recognition by Stochastic Model"

  The conventional methods 1 and 2 have the following problems. That is, in the conventional method 1, a recognition result candidate cannot be obtained until the utterance is completed, and it takes time to present information based on the recognition result.

  In the conventional method 2, since a phoneme string having a high likelihood is used as a recognition result candidate at each time during collation, a phoneme shortly after the start of collation may be included in the recognition result candidate. There is a possibility that many phoneme sequences that have been processed are output as recognition result candidates.

  For example, FIG. 6 shows an example of a phoneme string having a high likelihood obtained as a result of collating acoustic parameters from the first frame to a certain time. Even when the phoneme sequence actually spoken is “o / NN / s / e / i”, “o / NN / s / e”, which is a part of the phoneme sequence spoken as shown in FIG. "O / NN / s / e / i / n", "o / NN / s / e / i / g", "o / NN / s / e / i" with phonemes following the phoneme sequence In many cases, a phoneme sequence similar to the uttered phoneme sequence such as “/ s” is listed as a highly likely phoneme sequence.

  The present invention has been made in view of the above-described problems of the prior art, and an object of the present invention is to provide a speech recognition system capable of obtaining a recognition result candidate with high accuracy during utterance.

In order to achieve the above-described object, the present invention provides a speech recognition system capable of sequentially outputting speech recognition result candidates during speech production, a speech detection unit that extracts a speech signal by shifting it by ΔT for each frame length T, and the speech detection An acoustic analysis unit that extracts an acoustic parameter representing the characteristics of the speech signal from the extracted speech signal obtained from the speech unit, a matching unit that calculates the likelihood of a phoneme sequence from acoustic parameters from the start of speech to the middle of speech, A recognition result candidate determining means for determining a recognition result candidate in the middle of utterance based on a phoneme string collation obtained from the speech analysis unit, and The acoustic parameters up to the time are collated with a recognition vocabulary standard pattern that is composed by a model representing the phoneme characteristics according to the phoneme sequence of the recognition vocabulary, and the recognition result phoneme string candidate The first matching unit that outputs the first likelihood, and the acoustic parameters from the first frame to the current time are compared with a background pattern that models all Japanese, and the second likelihood is output. Ri Do and a second verification unit that, the recognition result candidate determining means,
Taking a difference between the first and second likelihoods of the phoneme sequence obtained from the collating unit, and calculating a normalized score that is a value obtained by dividing the difference by the execution interval of the speech recognition process; normalized score calculator for the normalization score to output the first likelihood m _i of advance greater than the determined threshold phoneme string and the sound Motoretsu, and indexed the likelihood m _i higher than the threshold phoneme sequence The first posterior probability is obtained, the second posterior probability of the phoneme sequence included in the input speech from the start of utterance to the middle of utterance is obtained based on the first posterior probability, and the second posterior probability is determined in advance. It is characterized in that it comprises a recognition result candidate determination unit that determines a phoneme string that is equal to or greater than the reference value as a recognition result candidate .

According to the first and second aspects of the present invention, it is possible to obtain an accurate recognition result candidate in the middle of utterance, and the time until presentation of information can be shortened.

  Further, since the recognition result candidate can be selected in detail using the normalized score and the posterior probability, it is possible to reduce the output of unnecessary recognition result candidates.

  Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention. In FIG. 1, the same reference numerals are assigned to the same or equivalent parts as in FIG.

In the figure, a voice input unit 1 inputs a voice signal from a user and sends it to a voice detection unit 2. The voice detection unit 2 cuts out the voice signal obtained from the voice input unit 1 sequentially by ΔT (time) every frame length T (time) as shown in FIG. 4 and sends each to the acoustic analysis unit 3. The acoustic analysis unit 3 extracts an acoustic parameter representing the characteristics of the audio signal from the audio signal cut out for each frame, and sends the acoustic parameter to the first verification unit 4 and the second verification unit 7 .

The first collation unit 4 collates the recognition vocabulary standard pattern 5 obtained by concatenating the models representing the features of phonemes with the acoustic parameters from the first frame to the current time, and recognizes the recognition result phoneme string candidate and its likelihood m _K. Is sent to the recognition result candidate determination means 10. The second collation unit 7 collates a background pattern 8 that models all Japanese language and an acoustic parameter from the first frame to the current time, and sends the likelihood m _B to the recognition result candidate determination means 10.

  The recognition result candidate determination means 10 selects a phoneme string having a high likelihood from the phoneme strings obtained from the first collating unit 4, and based on the likelihood of each phoneme string, the input speech from the start of utterance to the middle of utterance Only a phoneme string that is likely to be a phoneme string is determined as a recognition result candidate, and as one specific example, a normalization score calculation unit 11 and a recognition result candidate determination unit 12 may be configured. it can.

The normalization score calculation unit 11 calculates a normalization score S from the likelihood m _K obtained by the first collation unit 4 and the likelihood m _B obtained by the second collation unit 7 by the following formula. Then, a phoneme string hypothesis whose score is higher than a predetermined threshold (hereinafter referred to as an N-best phoneme string) and a likelihood m _i of each phoneme string (= a likelihood m _k higher than the threshold ) are determined as a recognition result candidate. Send to part 12.

S = (m _K −m _B ) / (t _P −t _S )

Here, t _S indicates the start time of the first frame, and t _P indicates the current time. _That, t P -t _S is the length of a section used for verification. The normalized score S is described in detail in Japanese Patent Application No. 2003-048608, which is a patent application filed by the present applicant.

  Next, the recognition result candidate determination unit 12 calculates the posterior probability for each phoneme sequence by the following method using the N-best phoneme sequence and the likelihood of each phoneme.

(1) N-best phoneme sequence likelihood m _i a (1 ≦ i ≦ N) and indexing to determine the i-th likelihood m posteriori probability (first posterior probability) of _i p _i from the following equation.

(2) Next, if a certain phoneme sequence x is included in the i-th candidate and δ _{x, i} = 1, and if it is not included in δ _{x, i} = 0, the phoneme sequence x is included in the input speech. The posterior probability (second posterior probability) p _x is calculated by the following equation.

  Thereafter, the recognition result candidate determination unit 12 determines a phoneme string included in a certain range θ from the maximum value of the posterior probability in the phoneme string as a recognition result candidate.

In FIG. 2, recognition is performed using a recognition dictionary including words such as “speech”, “speech recognition”, “speech synthesis”, and “speech operation”, and the input speech “onsei” and recognition vocabulary standard up to time t are recognized. An example of a phoneme sequence obtained when matching patterns is shown. Here, it is assumed that five phoneme sequences having high normalization scores are listed, and the posterior probabilities p _i (i = 1, 2,..., 5) are calculated from the likelihoods of the respective phoneme sequences.

Originally, the phoneme string corresponding to the voice “onsei” until time t is “o / NN / s / e / i” (p ₁ = 0.22), but in addition, the phoneme string “o / NN / “o / NN / s / e” (p ₅ = 0.18) that forms part of “s / e / i” or “speech recognition” that follows “o / NN / s / e / i” on the recognition dictionary ”,“ Speech synthesis ”, phoneme sequence“ o / NN / s / e / i / n ”(p ₂ = 0.21) with phonemes in the words“ speech operation ”,“ o / NN / s / “e / i / g” (p ₃ = 0.20) and “o / NN / s / e / i / s” (p ₄ = 0.19) are also listed as phoneme strings with high normalized scores.

From the definition of this time, the posterior probability p _x, the posterior probability _{p o / NN / s / e} = 1.0 of phoneme string "o / NN / s / e". Similarly, the posterior probability p _{o / NN / s / e / i} of the phoneme string “o / NN / s / e / i” is calculated as 0.82. Also, similar phoneme strings such as “o / NN / s / e / i / n”, “o / NN / s / e / i / g”, “o / NN / s / e / i / s” are candidates. The posterior probabilities are in the following cases: p _{o / NN / s / e / i / n} = 0.21, p _{o / NN / s / e / i / g} = 0.20, p _{o / NN /} It is expected to take a low value of _{s / e / i / s} = 0.19.

  Here, when θ = 0.50 is used as the criterion for selecting the candidate, for example, phoneme sequences “o / NN / s / e” and “o / NN / s / e / i” having high posterior probabilities are recognized as recognition results. It is output as a candidate, and "o / NN / s / e / i / n", "o / NN / s / e / i / g", "o / NN / s / e / i / s", etc. are recognized results Excluded from the candidate.

  As described above, according to the present embodiment, it is possible to output a recognition result candidate with high accuracy during the utterance.

It is a block diagram which shows the structure of one Embodiment of this invention. It is explanatory drawing which shows the example of a phoneme sequence with a high normalization score when the phoneme sequence which uttered is "o / NN / s / e / i". It is a block diagram which shows the structure of an example of the conventional speech recognition system. It is explanatory drawing of audio | voice area detection and feature parameter extraction. It is a block diagram which shows the structure of the other example of the conventional speech recognition system. It is explanatory drawing of the recognition result candidate output by the speech recognition system of FIG.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Voice input part, 2 ... Voice detection part, 3 ... Acoustic analysis part, 4 ... 1st collation part, 5 ... Recognition vocabulary standard pattern, 7 ... 2nd collation part , 8 ... Background pattern, 10 ... Recognition result candidate determination means, 11 ... Normalization score calculation unit, 12 ... Recognition result candidate determination unit.

Claims (2)

In a speech recognition system that can sequentially output speech recognition result candidates during utterance,
An audio detection unit that extracts audio signals by shifting them by ΔT for each frame length T;
An acoustic analysis unit that extracts acoustic parameters representing the characteristics of the audio signal from the cut-out audio signal obtained from the audio detection unit;
From the acoustic parameters from the start of utterance to the middle of utterance, a matching unit that calculates the likelihood of the phoneme sequence,
A recognition result candidate determining means for determining a recognition result candidate in the middle of utterance based on a matching result of the phoneme string obtained from the matching unit;
The collation unit collates the acoustic parameters from the first frame extracted from the acoustic analysis unit to the current time with a recognition vocabulary standard pattern configured by a model representing a phoneme characteristic according to the phoneme sequence of the recognition vocabulary. The first collation unit that outputs the recognition result phoneme string candidate and the first likelihood, and the acoustic parameters from the first frame to the current time are collated with a background pattern that models all Japanese. , Ri Do and a second matching unit for outputting a second likelihood,
The recognition result candidate determination means includes
Taking a difference between the first and second likelihoods of the phoneme sequence obtained from the collating unit, and calculating a normalized score that is a value obtained by dividing the difference by the execution interval of the speech recognition process; normalized score calculator for the normalization score to output the first likelihood m _i of advance greater than the determined threshold phoneme string and the sound Motoretsu,
Obtains a first posterior probability exponentially the likelihood m _i higher than the threshold phoneme string, the second phoneme strings contained in the input speech halfway utterance from the utterance start based on the posterior probability of the first A speech recognition system including a recognition result candidate determination unit that obtains a posterior probability and determines a phoneme string having a second posterior probability equal to or higher than a predetermined reference value as a recognition result candidate . The speech recognition system according to claim 1 ,
The recognition result candidate determination unit indexes the likelihood m _i (1 ≦ i ≦ N) of the phoneme string higher than the threshold, and sets the first posterior probability p _i of the i-th likelihood m _i as (1 ) calculated from the equation, the second posterior probability p _x of the phoneme sequence x contained in the input speech halfway utterance from the utterance start, when the sound Motoretsu x is included in the i-th candidate [delta] _{x, i =} 1 When not included, δ _{x, i} = 0 and calculated by the following equation (2).

Images