JPS59170894A - Voice section starting system - 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 Application of the Invention] The present invention relates to a speech segment segmentation method for a speech recognition device, and particularly relates to a speech recognition device that uses a telephone line [7] or under high noise environments such as inside a factory or outdoors. The present invention relates to a speech segment extraction method suitable for speech recognition devices used in.

[Prior art]

The speech segment extraction method of conventional speech recognition devices is a method that focuses only on the difference in power information between the speech segment and the silent segment, sets an appropriate threshold for the signal level, and extracts the portion above the threshold μ as a speech segment. Tonatsu Shiita. 1 ~ Therefore,
If a telephone line or low-quality microphone is used as the input method, or if it is used in a factory or outdoors where high noise is generated, there will be no noticeable difference in power between the sound I section and the silent section. However, there was a drawback that a voice section was cut out by mistake.

[Purpose of the invention]

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech section extraction method that can accurately detect speech uttered in an environment with a poor signal-to-noise ratio.

[Summary of the invention]

The speech recognition device according to the present invention includes an analysis section that analyzes an input signal, a noise stamp creation section that creates a standard noise stamp, and a speech recognition device that determines whether or not the analyzed input signal is speech using the noise stamp. -In the case of sound or voice, the voice section cutout section (comprised of a part and an identification section that performs a discrimination process on the voice cut out by the voice cutout section) performs cutout processing. .

Here, the noise refers to sounds other than sound waves, such as line noise Ef generated on the telephone line and environmental noise generated around the vocal shore. On the other hand, voiced sounds (vibration waves Cζ of l go voice 弗, voiced sounds that are uttered when the p-way is excited, and unvoiced sounds that are excited by friction and rupture of airflow in the vocal tract and are not accompanied by vibration of the horn band) It is structured and has properties that are essentially different from noise.

In particular, significant differences occur in the frequency range where the properties of waves become apparent. Therefore, in the frequency domain, the sampling (
- By comparing the noise pattern and the input signal,
It is possible to determine whether the input signal is audio. That is, according to the present invention, it is possible to realize a stable speech recognition device that is less likely to erroneously cut out a speech section due to high-power noise.

Since noise is generated by various miscellaneous factors,
There are quite a few areas in which it cannot be said with certainty that the properties have been clearly defined. (~gashi, statistically J(ot
This is called h-noise, and its frequency spectrum tends to have more power concentrated in the low range than in the high range, as shown in the first NN'. On the other hand, unvoiced sounds are excited by air friction on the road, so their frequency spectrum has a structure in which the power is concentrated in the high range rather than the low range, as shown in Figure 2. .゛However, the frequency spectrum of fJ vocal folds is excited by imaging the vocal cords.As shown in Figure 3, Lam expresses the vocal fold movements at the fundamental frequency (fJ).
The harmonic sound structure with pitch frequency) is set as 41. Since the shape of the Sarariko and the torpedo changes slowly, we use the formant that represents its resonance characteristics. As described above, there are significant differences in the frequency spectrum structures of speech and noise. Therefore, it is possible to determine whether or not the input signal is speech by creating a standard noise template in the noise template creation section and comparing this noise template with the analyzed input signal. Therefore, according to the present invention, it is possible to realize a stable sweat voice recognition device that does not erroneously cut out both sound sections due to high-power noise.

[Embodiments of the invention]

An embodiment of the present invention will be described below with reference to FIG. An input signal S input from the microphone 1 is input to an analysis section 2, and an analysis process is performed to extract power information and frequency information. On the other hand, line noise on a telephone line can be expected to remain virtually unchanged during a single call. Utilizing this property, the noise in the silent section after the telephone line is tangential is analyzed and stored in the noise buzz generator 3. The voice section cutout section 4 is configured to compare the analyzed input signal with the noise beat stored in the noise beat generation section 3, and if it exhibits different characteristics, it is determined to be a voice section and is cut out. Similarity calculation unit 6 (C) calculates the similarity between the m-word set to be recognized stored in the standard slam storage unit and the input speech determined to be a speech section. Judgment processing is performed based on the similarity calculated by the similarity calculation unit 6, and the identification result of the input voice is determined.
．． 7.8) can be realized both analogously and digitally. Below, we will mainly explain the case where the circuit is configured with digital circuits.

Persimmons have been proposed as analysis principles for the analysis section 2, but all of them express the resonance characteristics of the vocal tract both implicitly and explicitly. However, regardless of which method is used, the principles of the present invention can be easily applied. - By way of example, the known Channelo-Iruta type spectrado analyzer may be considered. That is, an input signal S input from a microphone is A/I) converted and input to a channel filter type spectrum analyzer. Each channel of the channel filter type spectrum analyzer corresponds to each frequency band, and the spectrum of each frequency band is output as time-series frequency information. The specific configuration of this analysis section 2 is not directly related to the main part of the present invention, and will therefore be omitted.

It's a moat, similarity calculation section 6, judgment section 7, standard slam storage section 5
Various types of identification principles have been proposed for the identification unit including the above, and nr:) can be applied to any of them. - As an example, the known slam matching method may be considered. That is, a set of words to be recognized and recognized is determined, and each word is described using appropriate parameters and stored in the standard button storage section. The input signal is converted into a parameter in the analysis section, and when it is determined to be speech in the speech section extraction section, it is input to the identification section. This manual click is compared with the standard click, ie, click matching is performed, and it is determined that the word is the same as the standard click that gives the maximum match. The specific configuration of this identification section is also omitted since it is not directly related to the main part of the present invention.

Next, a specific example of the configuration of the voice section cutting section 4 and the noise bang creating section 3 will be explained with reference to FIG. The input signal converted into parameters representing power and frequency information is input to the selection section 41. In the selection section 41, a part of the silent section is selected for the noise bang generation section and inputted to the input buffer 31.

In addition, the input signal S including the speech section to be subjected to recognition processing is
is selected to the voice section extraction section, and the input part 7742,
44. Here, the timing for creating the noise bang can be considered, for example, during the question section of the voice response device. That is, speech recognition devices are typically operated as part of a voice response system. Therefore, since the speaker is not expected to make any utterances during the actual time, unvoiced sounds are input into the voice output pause section in the question sentence from the voice response.

Next, the noise parameters in the input buffer are calculated by the arithmetic averaging section 3.
2, the time average over a long period of time (for example, 400018α or more) is taken, resulting in a stable noise slam, which is stored in the noise bang storage unit 33.

On the other hand, the input signal S to be subjected to the recognition process stored in the input buffer 44 is 1
Every frame I/i: Similarity is measured t4. Next, the determination unit 46 determines whether or not the relevant frame is noise using an appropriate threshold value θ as a total criterion. That is, if frames having a similarity of θ or less continue for N frames, the selection unit 43 starts a cutting operation at the starting point of the voice section.

Conversely, if frames having a similarity of 0 or more continue for N or more frames, the selection unit 43 ends the extraction operation.

As described above, according to the present invention, it is possible to realize a stable speech recognition device that does not erroneously cut out speech sections due to high-power noise.

Furthermore, if the speaker-independent recognition device can obtain information on the position of characteristic phonemes such as vowels in the input speech, it can be registered and used as a standard button using the method described in the present invention. It is. That is, it is possible to perform matching using a phonetic standard pattern unique to the speaker, and a high recognition rate can be expected.

〔Effect of the invention〕

According to the present invention, since speech uttered under high noise can be accurately extracted, there is an effect of improving recognition accuracy of a speech recognition device. Furthermore, since the voice recognition device can be introduced into a factory that generates high noise, which was previously unusable due to ambient noise, there is an effect of expanding the market for voice recognition devices. Furthermore, it is no longer necessary to use a high-quality, expensive microphone, and an inexpensive microphone can be used, which has the effect of reducing the cost of the speech recognition device.

[Brief explanation of the drawing]

Figure 1 is a diagram for explaining the spectrum structure of noise, Figure 2 is a diagram for explaining the spectrum composition of unvoiced sounds, Figure 3 is a diagram for explaining the spectrum composition of voiced sounds, and Figure 4 is a diagram for explaining the spectrum composition of voiced sounds. FIG. 5 is a block diagram showing an example of the configuration of a speech recognition device according to the present invention, and FIG. 5 is a block diagram showing an example of the configuration of a speech segment cutting section according to the present invention. In FIG. 4, 1 is a microphone, 2 is an analysis unit, 3 is a noise bang generation unit, 4 is a voice section extraction unit, 5 is a standard bang storage unit, 6 is a similarity calculation unit, 7 is a judgment unit, and 8 is a recognition unit. control section. In FIG. 5, 31Vi buffer 1.32 is an arithmetic averaging section, 33 is a noise button storage section, 41 is a selection section, and 42
is the input buffer 3. 43 is the selection section, 44 is the input buffer 2. 45H matching section, 1st (2) 2nd Fig.

Claims (1)

[Claims] 1. In a speech recognition device comprising a speech signal observation section, an analysis section, a speech section extraction section, and a recognition logic section, a means for detecting a specific section, a means for analyzing the detected signal, and an analysis (~ 1. A voice section extraction method, comprising means for storing a signal, and means for calculating a difference between an analysis result of the particular section and power and frequency information of input voice.