JPS6225796A - Voice recognition equipment - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to an improvement in a speech recognition device, and more particularly to an improvement in recognition rate.

BACKGROUND OF THE INVENTION Speech recognition technology is a field that is expected to be put to practical use as a man-machine interface for input into word processors and computers.

Speech recognition devices include those that use monosyllables (CV% C represents consonants, ■ represents vowels), those that use cv and vcv, and those that use phonemes (C and V
). In addition, there is a registration type in which the user utters and registers a standard voice in advance and then starts the recognition process, and a registration type in which a ti-like pattern is prepared by performing statistical processing based on a large amount of vocalization data and the user registers it. There are unspecified types that do not require . In addition, the mainstream methods of feature extraction are those using linear predictive analysis and filter banks. Here, as well as the conventional example and the embodiment, Cv and vcv are used as recognition units, and unspecified speech L2. The identification device will be explained below. An example of a conventional speech recognition device will be described below with reference to the drawings.

FIG. 3 is a block diagram showing the configuration of an unspecified speech recognition device. The audio input from the audio input terminal 21 is processed by the feature extraction unit 22 with a window length of 20 m5ec and a frame shift of 5
A linear prediction analysis using the m5ec15th order autocorrelation method is performed and output as a set of 16 parameters including 15 cepstral coefficients and residual power. (For information on linear predictive analysis, see Markel Gray, Translated by Hisaki Suzuki, Linear Prediction of Speech, 1980, Corona Publishing). Next, the silence detection section 23 uses the residual power to determine the beginning, end, and middle of a word. In the vowel recognition unit 24, a discrimination function that stores coefficients of a vowel discrimination function (Social Statistics by Sanbe Yasuda, see Chapter 2, Section 7, Maruzen, 1969) obtained by statistical processing of the vocalization data of Mr. The coefficients are read from the function storage unit 25, and vowel recognition is performed for each frame for parts other than the silent portion detected by the silence detection unit 23. 26 is a stationary point detection unit that recognizes the vowel for each frame obtained by the vowel recognition unit 24. The fulcrum is extracted from the vowel recognition results and output as a vowel stationary point string.27 reads the standard pattern from the standard pattern storage section 28 created in advance by the phoneme recognition section, performs input cover pattern and leaf matching, and outputs the vowel stationary point string. The standard pattern with the minimum result distance is output as a recognized phoneme string. 29 is a word recognition unit that compares the contents of the phoneme recognition result obtained in 9 with the word dictionary 30 stored in the symbol string notation and finalizes the result. A recognition result as a typical word is obtained at the recognition result output terminal 12.

(For example, Mifune et al.: Institute of Electronics and Communication Engineers, PRL83-40,
This paper uses a filter bank instead of linear predictive analysis as a feature extraction means, and uses the variance between each frame instead of using the vowel recognition results of each frame as a vowel stationary point detection means, but this is an example of a conventional example. It can be given as follows. ) Problems to be Solved by the Invention In such conventional speech recognition devices, the stability in detecting vowel stationary points (which can be expressed as the number of frames in which the same vowel recognition result continues) is determined by a threshold value for detecting stability. Since only one threshold value is used as one threshold value, there is a problem that when the speech rate becomes faster, the vowel stationary point is dropped, and when it becomes slower, the vowel stationary point is inserted, and the recognition rate decreases as the speech rate changes. Furthermore, it is unavoidable that a person's speaking speed fluctuates, and it is difficult to always maintain the same speaking speed, which places a heavy burden on the speaker.

The present invention has been made in view of this point, and it provides several thresholds for detecting vowel stationary points, and uses which threshold is used to determine the first stationary point to easily estimate the utterance rate. The purpose of this method is to reduce the insertion and omission of vowel steady points caused by fluctuations in the speaking speed by detecting steady points according to the speaking speed, and to further reduce the burden on the speaker.

Means for Solving the Problems In order to solve the above-mentioned problems, the present invention provides several threshold values in stages for vowel steady point detection, estimates the utterance rate, and performs regular inspection taking the utterance rate into consideration. The purpose of this is to deal with fluctuations in the rate of speech by performing the following steps.

Function The function of this means is as follows. In other words, several thresholds are set regarding the stability (number of mm frames) when detecting a vowel steady point, and the speaking rate is first estimated, and then the vowel steady point is detected according to the speaking speed. It is possible to improve the recognition rate by reducing the insertion and omission of vowel stationary points due to fluctuations in the number of vowels. Furthermore, since it is not necessary to keep the speaking speed constant, it is possible to reduce the burden on the speaker.

EXAMPLE An example of the present invention will be described below. FIG. 1 is a block diagram showing one embodiment of the present invention. The audio input from the audio input terminal l is subjected to linear predictive analysis in the feature extraction unit 2 using a window length of 2 Q msec, a frame shift of 5 m5 ec, and a 15th order autocorrelation method, and 15 cepstral coefficients and residual power are extracted. are output as a set of 16 parameters in total. Next, the silence detection section 3 uses the residual power to determine the beginning, end, and middle of a word. The vowel recognition unit 4 reads the coefficients from the discriminant function storage unit 5 which stores the coefficients of the discriminant function obtained by statistically processing the utterance data of Mr. 6 is a vowel stationary point detection unit.First, the utterance rate estimation unit 6a estimates the utterance rate, and based on the result, the vowel recognition unit 6b performs the vowel recognition. From the vowel recognition result sequence for each frame obtained in step 4, stable ones are extracted and output as a stationary point sequence.

The steady point detection section 6 will be explained in detail later. 7 is a standard pattern storage unit 8 that has been revised and created by the phoneme recognition unit.
The standard pattern is read from the input cover pattern and DP matching is performed, and the phoneme of the standard pattern with the minimum distance is output as a recognition result phoneme string. 9 is a word recognition unit that compares the phoneme recognition result obtained in step 7 with the contents of the word dictionary 10 stored as symbol strings, and outputs the final recognition result as a word to the recognition result output terminal 11. .

Next, the algorithm in the stationary point detection section 6 will be explained in detail. FIG. 2 is a flowchart showing an outline of processing in the stationary point detection section 6 (both 6a and 6b). Using the vowel recognition results for each frame obtained by the vowel recognition unit 4 as input, calculate the number of "!a" continuation frames of the same vowel. For example, if the input vowel string is aaaaaaaaaaaa
eeeeee 1eee i i i i then a12
Find the number of continuous frames of the same vowel as e6i4 (however, cases where the number of consecutive frames is 3 or less are excluded), and this 12.6.4 is the number of continuous frames for each (=nf
i). (In this example, there are three values, nfl, nf2, and nf3.) Therefore, the maximum value of the number of continuous frames in the input cover turn is determined by setting nf = max (nfl, nf2, nf3). This vowel has the highest stability. When this nf is nf>T)It (threshold 1), stationary point detection is performed assuming that the speaking speed is relatively slow, and when nf>Tl12 (threshold 2, THI>TH2), the speaking speed is normal. Otherwise, steady point detection is performed assuming that (nf < TH2) and the speaking speed is fast. Thereafter, using timing processing, processing for adding and erasing stationary points is performed and output as vowel stationary points. (Timing processing is 2
Stationary points are added and deleted on the assumption that no two stationary points are too close or too far apart. It takes advantage of the fact that in Japanese, each vowel has approximately equal timing. In this example, τ old = 25, TH2・15
is adopted. ) As described above, the speaking rate is estimated by setting the threshold value in stages for detecting the steady point, and the steady point is detected in accordance with the speaking speed. As a result, it is possible to reduce the insertion and omission of vowel stationary points due to variations in speech rate. Furthermore, since there is no need to forcibly keep the speaking speed constant, the burden on the speaker can be reduced.

In the embodiment, two threshold values, THI and TlI2, are used, but the present invention is not subject to any restrictions on the number of threshold values. Further, although a CV and VC cine specific type recognition device has been described as a recognition unit, this does not limit the present invention in any way.

Effects of the Invention As described above, according to the present invention, the speaking rate can be estimated using a simple method, and by detecting vowel steady points based on the speaking rate, it is possible to reduce the omission and insertion of vowel steady points due to fluctuations in the speaking rate. Therefore, it is possible to measure the improvement in the recognition rate. Furthermore, since there is no need to forcibly keep the speaking speed constant, it is possible to reduce the burden on the speaker.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the phoneme recognition results of an embodiment of the present invention, FIG. 2 is a flowchart showing an outline of the algorithm of the present invention, and FIG. 3 is a block diagram of a conventional example. l...Audio input end, 2...Feature extraction unit, 3...Silence detection unit, 4...Vowel recognition unit, 5... - Discrimination function storage unit, 6... Steady point detection unit, 6a... Speech rate estimation unit, 6b.
...Steady point detection unit, 7...Phonological recognition unit,
8...Standard pattern storage unit, 9...Word recognition unit, 10...Word dictionary storage unit, 11...
...Audio output terminal.

Claims (1)

[Claims] a voice input means, a feature extraction means for performing feature extraction on the voice inputted from the voice input means at regular intervals to extract a feature parameter string, a vowel recognition means for performing vowel recognition on the feature parameter string, and the vowel recognition. A stationary point detection means for detecting a stable part from the result and outputting it as a vowel stationary point sequence; a standard pattern storage means for storing a standard pattern prepared in advance for each phoneme to be recognized; It has a phonological recognition means that compares each of the standard patterns and converts the characteristic parameter sequence into a phonological sequence, and the stationary point detection means sets a threshold in stages to detect a stationary point to estimate the speech rate. A voice recognition device characterized by performing stationary point detection on the top of the screen.