JPS6355599A - Voice recognition equipment - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] Industrial applications The present invention relates to a speech recognition device that improves recognition rate.

Conventional technology Phonological recognition means is a field that is expected to be put to practical use as a man-machine interface such as input to word processors and computers.

Speech recognition devices include those that use monosyllables (CV, C: consonant, V: vowel) as units for recognizing input speech, those that use Cv and VCV, and those that use phonemes (C and V).
) may be considered.

In addition, the user speaks a standard voice in advance.

There is a registered type that starts recognition processing after registration, and an unspecified type that performs statistical processing based on a large amount of vocalization data, prepares universal standard patterns, and does not require user registration. . In addition, the mainstream methods of feature extraction are those using linear predictive analysis and filter punching. In this specification, an unspecified speech recognition device using CV and VCV as the recognition unit and linear predictive analysis as the feature extraction method will be described in both the conventional example and the embodiment, but the present invention is not limited to these. An example of a conventional speech recognition device will be described below with reference to FIG.

FIG. 2 is a block diagram showing the configuration of an unspecified speech recognition device. The audio input from the audio input terminal 1 is processed by the linear predictive analysis unit 2 with a window length of 20 m (with a frame shift) of 5 m.
5ec, analyzed using the autocorrelation method of order 15,
It is output as a sequence of 16 parameters in total, including 15 cepstral coefficients and residual power (0th order cepstral coefficient). (See Corona).

Next, the silence detection section 3 uses the residual power to determine the beginning, end, and middle of the word. The vowel recognition unit 4 uses a vowel discrimination function (Yasuda Mibe, Social Statistics, Chapter 2, Section 7, 1) obtained by processing a large amount of vocalization data in advance.
Discriminant function storage unit 5 that stores coefficients of 969 Maruzen)
The coefficients are read in, and vowel recognition is performed for each frame in a portion other than the silent portion detected by the silence detector 3. Reference numeral 6 denotes a stationary point detection unit which extracts stable vowel recognition results for each frame obtained by the vowel recognition unit 4 and outputs them as a stationary point sequence. The number of stationary points indicates the number of syllables of the input voice. 8 is a phoneme recognition unit and a standard pattern storage unit 9
DP matching is performed between the standard pattern read from the input speech and the parameter string obtained from the input speech, and the standard pattern with the minimum distance is output as the recognized phoneme string.

The standard pattern storage section 9 stores universal standard patterns created in advance from a large number of utterance data using statistical processing. 10 is a language processing unit and a phonological recognition unit 8
Linguistic processing is performed on the phoneme recognition results obtained in , and the final recognition results are obtained at the recognition result output terminal 12 . Reference numeral 11 is a language dictionary in which dictionaries used for processing 10 languages are stored (for example, Mifune et al.: Institute of Electronics and Communication Engineers PRI, 83-40.
;This paper uses a filter bank rather than linear predictive analysis, and uses inter-frame variance rather than vowel recognition results to detect stationary points; however, phonological recognition is performed after vowel stationary points are detected. This can be cited as a conventional example. ).

Problems to be Solved by the Invention In such conventional speech recognition devices, phoneme identification is controlled using the detection results of stationary points. is having a big impact on In order to improve the vowel recognition rate at stationary points, there are countermeasures such as using vowel recognition candidates up to the second rank, but there is a problem that processing time increases.

The present invention has been made in view of this point, and after detecting a stationary point, vowel recognition is performed again at the detected stationary point to determine the certainty of the existence of the stationary point and the reliability of the first vowel candidate, The purpose is to measure the improvement of stationary point detection rate and vowel recognition rate.

Means for Solving the Problems The present invention is a speech recognition device that is provided with means for performing vowel recognition again at the detected stationary point after detecting a stationary point.

According to the above-described configuration, the present invention performs vowel recognition again at the detected stationary point after detecting a stationary point, determines the certainty of the existence of the stationary point and the reliability of the first vowel candidate, and performs necessary recognition. In some cases, the recognition rate can be improved by performing recognition processing without determining vowel candidates.

Embodiment 1 FIG. 1 is a block diagram showing the configuration of an unspecified speech recognition device according to an embodiment of the present invention. In the figure, solid lines indicate the flow of processing, and dotted lines indicate data references. The audio input from the audio input terminal 1 is processed by the linear predictive analysis unit (feature extraction unit) 2 with a window length of 20 m5ec and a frame shift).
It is analyzed using the autocorrelation method of order 1 and order s, and 1
A total of 16 parameters, including 5 cepstrum coefficients and residual power (0th order cepstrum coefficient), are output as a string.

Next, the silence detection section 3 uses the residual power to determine the beginning, end, and middle of the word. The first vowel recognition unit 4 reads the coefficients from the discriminant function storage unit 6 that stores the coefficients of the vowel discriminant function obtained by processing a large amount of utterance data in advance, and reads the coefficients from the discriminant function storage unit 6 that stores the coefficients of the vowel discriminant function obtained by processing a large amount of utterance data in advance. For the other parts, vowel recognition is performed for each frame.

6 is a stationary point detection unit which extracts stable vowel recognition results for each frame obtained by the vowel recognition unit 4 and outputs them as a stationary point sequence. 7 performs more detailed vowel recognition on the stationary points detected by the second vowel recognition section. The first vowel recognition means and the second vowel recognition means will be explained in detail later. A phoneme recognition unit 8 performs DP matching between the standard pattern read from the standard pattern storage unit 9 and the parameter string obtained from the input speech, and outputs the standard pattern with the minimum distance as a recognized phoneme string. The standard pattern storage unit 9 stores universal standard patterns created in advance from a large number of voice data using statistical processing. 1o is a language processing unit that performs language processing on the phoneme recognition result obtained by the phoneme recognition unit 8, and outputs the final recognition result to the recognition result output terminal 12. Reference numeral 11 denotes a language dictionary in which dictionaries used for processing 10 languages are stored.

Next, the first vowel recognition section and the second vowel recognition section will be explained in detail. A method using Mahalanobis general distance is already known as a vowel recognition method.

Now, the covariance matrix of the sample set used to create the discriminant function is V,
The average value for each vowel is xv (v=a,...N),
Letting the incoming covert turn be X, the Mahalanobis general distance between each vowel and the incoming covert turn can be found using equation (1).

d(x, v)=x V-x-2"vtv-+engine+xvt
y-1xv (1) Valley V is a 16-by-15-column matrix, where X indicates a 15th-order vector, and 16 corresponds to the analysis order in the feature extraction section.

In equation (1), the second term xv"V', the third term xv
Since t■-1xv is a constant, equation (1) can be replaced as equation (2).

d(x,su=x V-x-271,x+Da ・-
--=-=(2)lv=xOtv-1Da=xvv-x
The first term of v(2) acid requires a considerable amount of calculation as it involves matrix operations with 16 rows and 16 columns, but this part does not depend on the vowel and is determined only by the input cover pattern, so it is difficult to determine which vowel is closer to the first term. If you only want to make a judgment, the first term can be ignored. Therefore, in the first vowel recognition means, vowel recognition is performed using equation (3), which has a small number of calculation formulas, as a distance measure, and the result is used to perform stationary point detection.

d' (x, v)=-21Vx+Da -・
Yama・-Bun- (3) Then, in the second vowel recognition means, x V−x is determined only for the detected stationary points (4
) to obtain the Mahalanobis general distance.

d(x, v)=x V"'-x+d'(x, v)
-----・-・-・(4) The Mahalanobis general distance is a statistical distance and follows a -distribution with 16 degrees of freedom, so it can be highly reliable.

For example, when d(x, a)=11.04, the probability (=reliability) that X is a is determined to be 76%. Therefore, vowels are redetermined according to the following rules.

Rule 1) The reliability of the first candidate vowel is 10 chi or less (
d(x, v)≧22.3), all vowels are candidates.

Rule 2) When the reliability of the first candidate vowel is 50% or less, the second candidate vowel is considered.

Rule 3) If the reliability of the first candidate vowel is 60% or more and the reliability of the second candidate is not 20% or more, both the first and second vowel candidates are considered.

Rule 4) Except for the above, only the first candidate is used.

As described above, according to this embodiment, the vowel recognition rate can be improved by providing means for performing vowel recognition again on the detected stationary points and determining vowel candidates according to reliability.

As described in detail, according to the present invention, the vowel recognition rate can be improved, and its practical value is significant.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention, and FIG. 2 is a block diagram of a conventional speech recognition device. 2... Feature extractor, 3... Silence detector, 4... First vowel recognition unit, 6... Discrimination function storage unit, 6... ...Steady point detection section, 7...--
・Second vowel recognition unit, 8... Phoneme recognition unit, 9.
...Standard button storage section.

Claims (1)

[Claims] a voice input means; a feature extraction means for extracting features at regular intervals from the voice data inputted from the voice input means and outputting a feature parameter string; and a vowel recognition means for performing vowel recognition on the feature parameter string. , a stationary point detection means for detecting a stable part from the result of the vowel recognition and outputting it as a vowel stationary point sequence; a phoneme recognition means for recognizing the phoneme of the characteristic parameter sequence; and a vowel obtained by the stationary point detection means. A speech recognition device characterized by comprising means for performing vowel recognition again at a stationary point.