JPS61156295A - 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 Industrial Application] The present invention relates to a speech recognition device that can automatically recognize speech.

[Conventional technology]

One of the conventional speech recognition methods is the pattern recognition method. This method involves creating a standard voice pattern from the power spectra of several standard voices, and comparing this with an input voice pattern created from the power spectrum of the input voice, namely 1. <Tan matching is performed to calculate the value representing the degree of mutual difference (hereinafter referred to as "difference degree"), and the voice standard pattern with the least difference, that is, the voice standard pattern with the minimum degree of dissimilarity, is compared to the same voice. This is a method of determining.

When focusing on monosyllables as speech targets, the monosyllable standard pattern is a time-series feature vector that digitizes the power components of each frequency band of speech from the beginning to the end of the utterance. Further, the monosyllabic human pattern is also a time-series feature vector, and when the input speech pattern is a and the monosyllabic standard pattern is b, the degree of dissimilarity d is given by the following equation.

d-Σlar brl However, a-("+-"z," ・l aR) b-(
b,,b,,...,b*) [Problems to be solved by the invention] DP matching is used as a highly practical speech recognition method using pattern matching, but it only targets monosyllables. In speech recognition, "su" and "ma", "two" and "mi"
, "Yu" and "Ha" were easily misrecognized.

Furthermore, as a method for improving recognition accuracy by capturing the movements of articulators associated with vocalization, there is a method using information such as palate contact, etc., disclosed in Japanese Patent Application Laid-open No. 150997/1983. However, in order to obtain information such as palate contact, a large number of devices are required to be attached to the speaker, and the time and effort required to attach and detach the devices is troublesome.

Considered here, a speaker uses vocal cord vibration, exhaled air turbulence, etc. as a sound source, and uses articulatory organs such as the palate, tongue, and lips to produce speech. During vocalization, articulatory organs such as the palate, tongue, and lips continuously repeat fluctuations such as opening and closing, and these can be considered as an acoustic resonant system.

When a speaker utters a voice, mouth movements such as plosives are an important factor in the utterance, and it is important to capture the opening and closing of the mouth as information. However, in the case of plosive sounds such as M-line plosives, the sound is usually not emitted at the time the lips are closed, or the sound power is weak, making it difficult to capture it as information.

In addition, considering the palate, tongue, lips, etc. as an acoustic resonant system, an external sound wave oscillator is installed to radiate sound waves around the speaker's lips.
One possible method is to capture the echoes from the acoustic resonance system of the palate, tongue, lips, etc., and obtain information on the fluctuations of the articulatory organs such as the opening and closing of the lips. There was a problem that the characteristics changed depending on the positional relationship. Furthermore, when the microphone and the oscillator are integrated, there is a problem in that the mechanical vibrations of the oscillator are transmitted to the microphone.

[Means for solving problems]

In order to solve such problems, the present invention includes a sound wave oscillator that emits sound waves to the lips of a speaker, and a sound wave oscillator that is integrated with the sound wave oscillator and that emits sound waves to the lips of the speaker. an audio input section that detects the speaker's utterances along with the echoes generated by the input to the audio input section; The device includes an audio input switching unit that switches between oscillation of the sonic oscillator and audio signal input by setting a time, and a vibration noise storage unit that is connected to the audio pattern creation unit and stores vibration noise when no audio is input. This is what I did.

[Effect]

In the present invention, when a voice is input, the voice signal is corrected using numerical values stored in the vibration noise storage section to create an input voice pattern and a voice standard pattern.

〔Example〕

An embodiment of a speech recognition device according to the present invention is shown in FIG. In FIG. 1, 1 is a sound wave oscillator that emits sound waves to the speaker's lips, etc., 2 is an acoustic resonance system such as the roof of the mouth, and 5 is a sound wave that is integrated with the sound wave oscillator 1 and is emitted from the sound wave oscillator 1. Reverberant sound that occurs when it enters the speaker's lips etc. 3
6 is an analog-to-digital converter that converts the echo sound 3 and the voice 4 into digital audio signals; 7 is a microphone 5;
The time is set so that the input time interval of the audio signal input to the sound wave oscillator 1 does not overlap with the oscillation time interval of the sound wave oscillator 1, and the oscillation of the sound wave oscillator 1 and the input of the audio 4 are switched, and the audio signal a is output. An input switching unit 8 is an audio pattern creation unit that converts the audio signal a into an input audio pattern as a time series of feature vectors and an audio standard pattern, and 9 is connected to the audio pattern creation unit 8 and vibrates when no audio is input. a vibration noise storage unit that stores noise; 10 a standard pattern registration unit that stores a plurality of standard audio patterns; 11 an input pattern registration unit that registers an input audio pattern created from the digitized audio signal a; 12 13 is a pattern transfer unit that transfers the monosyllabic standard pattern registered in the standard pattern registration unit 10 and the input voice pattern registered in the input pattern registration unit 11; A DP matching section 14 performs recognition matching with a syllable standard pattern, and 14 is an output determination section that selects candidates for input speech patterns.

Next, the operation of the apparatus configured as described above will be explained. First, the audio input switching unit 7 sends the sound wave oscillator l to the sound wave oscillator l as shown in FIG.
The oscillation instruction signal S1 shown at l is sent out.

In addition, the voice manual switching section 7 turns off the gate signal S2 shown in FIG. 2(b) by 2 at a time τ1+, which is 1 plus the time τ2 during which the oscillation instruction signal S1 is on. Prohibit input of audio signals.

A sound wave oscillator 1 placed near the speaker's lips emits sound waves toward the acoustic resonance system 2 for a period of time τ1 during which the oscillation instruction signal S1 is on.

τ2 can be determined as follows. That is, if the speed of sound is V, the distance between the sound wave oscillator 1 and the microphone 5 is l, and the sound wave oscillator 1 starts oscillating a sound wave at time TO shown in FIG. 2(b) and ends at time T1, then The time when the sound wave oscillated just before reaches the microphone 5 is (T
I'+ j! /v). therefore,
When the relationship T2-Tl-τ2≧l/V is satisfied, direct human input of sound waves from the sound wave oscillator 1 to the microphone 5 can be prevented.

For example, v-340m/see, l1=1.5c
m, the slot time of one frame is 9 m5ec,
Assuming that the sampling frequency of the audio signal is 8 kHz and the oscillation frequency of the sound wave oscillator 1 is 5 to 15 kHz, τ1, τ2
When calculating, one period of a sound wave with a frequency of 5kHz is 0.
At 2m5ec, l/v = 0.0441 again. Therefore, the audio input switching unit 7 turns on the transmission instruction signal S1 for 0.2IIlsec and turns off the gate signal S2 for 0.2441m5ec at regular intervals.

In this way, the echo sound 3 from the acoustic resonant system 2 such as the palate and the speaker's voice 4 are input to the microphone 5 and converted into digital signals by the analog-to-digital converter 6.

The signal is converted into an audio signal of the same number, and is inputted to the audio pattern creation unit 8 as an intermittent audio signal a via the audio input switching unit 7. The audio pattern creation unit 8 uses this audio signal a to
Create the audio pattern shown in Figure 3. This audio pattern is an example of a time-series arrangement when lip opening/closing information is added, and frequency band-based audio patterns 15, 16, 17, and 18 are registered in chronological order in the order of frames 19, 20, and 21.

The acoustic wave oscillator 1 and the microphone 5 are coupled to the support rod 22 by struts 23, 24, respectively, as shown in FIG. Since the sound wave oscillator 1 and the microphone 5 have such a configuration, if mechanical vibration is generated at the same time as the sound wave is generated by the sound wave oscillator 1, this mechanical vibration will cause the sound wave oscillator 1 to operate regardless of whether or not the speaker is speaking. During this time, vibration noise is input to the microphone 5. When there is no speaker, only vibration noise due to attenuating mechanical vibrations is input to the voice pattern creation unit 8. At this time, the vibration noise storage unit 9 connected to the voice pattern creation unit 8 stores The vibration noise components of are stored for each channel.

When the speaker speaks, the voice pattern creation section 8 shown in FIG.
In the voice pattern, a value obtained by subtracting the value stored in the vibration noise storage unit 9 from the frequency band component ratio of the voice power including the voice 4 produced by the speaker and the echo sound 3 from the sound wave oscillator 1 is registered. be done.

The speech pattern creation section 8 operates in two modes: a monosyllabic standard pattern registration mode and an input pattern registration mode. In the case of a monosyllabic standard pattern register, a monosyllabic standard pattern is created from the digitized audio signal a and the vibration noise stored in the vibration noise storage section 9, and is registered in the standard pattern registration section 10. When the audio pattern creation section 8 is in the input pattern registration mode, the audio signal a and the vibration noise storage section 9
An input voice pattern is created from the vibration noise stored in and registered in the input pattern registration section 11.

The monosyllabic standard pattern registered in the standard pattern registration section 10 and the input speech pattern registered in the input pattern registration section 11 are transferred to the DP matching section 13 by the pattern transfer section 12. The DP matching unit 13 performs recognition comparison between all the monosyllabic standard patterns transferred from the standard pattern registration unit IO via the pattern transfer unit 12 and the input speech pattern sent from the input pattern registration unit 11, and identifies the differences. The monosyllabic standard patterns are rearranged in descending order of degree of dissimilarity and transferred to the output determination section 14 with the degree of dissimilarity added thereto. The output determination unit 14 selects m pieces from the array of monosyllabic standard patterns inputted from the DP matching unit 13 in order of decreasing degree of difference,
This is used as a candidate for the input voice pattern. However, among the m monosyllabic standard patterns that are candidates for input speech patterns, those whose degree of dissimilarity is greater than a certain value Re are excluded from the candidates.

In this way, when the sound source is the sound wave oscillator 1 provided near the lips and is emitted toward the acoustic resonance system 2 such as the cavity of the lips 9, it is possible to capture the echo sound 3, and the speaker It becomes possible to capture information on fluctuations in the articulatory organs that are closely related to vocalization, including at times when vocalization is not occurring.

Furthermore, by integrating the sonic wave oscillator 1 and the microphone 5, stable characteristics can be reproduced.

Furthermore, when creating a monosyllabic standard pattern and when creating an input speech pattern, it is possible to create a speech pattern with vibration noise components removed.

Note that the frequency of the sound waves emitted by the sound wave oscillator 1 may be in a high frequency range, for example, avoiding the audio frequency range.

〔Effect of the invention〕

As explained above, the present invention stores in advance the vibration noise generated by mechanical vibrations accompanying the generation of sound waves in the sound wave oscillator fixed to the sound input section, and the sound waves radiated from the sound wave oscillator are directly input to the sound input section. By blocking the voice signals generated by the two speakers and correcting the voice signals generated by the voices of the two speakers using stored vibration noise, we were able to reliably capture the echoes from the articulatory organs such as the palate. It is possible to add information on the opening and closing of lips, etc. to monosyllabic standard patterns and monosyllabic manual patterns, thereby improving recognition accuracy.

Furthermore, there is an effect that information on the opening and closing of lips, etc., can be obtained without using any equipment worn by the speaker. Furthermore, even when performing voice recognition repeatedly, since the mutual positions of the sonic wave oscillator and the microphone are fixed, there is an effect that stable and highly accurate recognition can be performed every time.

[Brief explanation of drawings]

FIG. 1 is a block system diagram showing an embodiment of the speech recognition device according to the present invention, FIG. 2 is a time chart for explaining its operation, FIG. 3 is a speech pattern diagram showing the state of speech registration, and FIG. FIG. 4 is a configuration diagram showing a state of coupling between a sound wave oscillator and a microphone. 1... Sonic oscillator, 2... Acoustic resonance system, 5...
...Microphone, 6...Analog-digital converter, 7
...Audio input switching unit, 8...Audio pattern creation unit, 9...Vibration noise storage unit, 10...Standard pattern registration unit, 11...Input pattern registration unit, 12・・・・・・
Pattern transfer section, 13...DP matching section, 14.
... Output determination section, 22... Support rod, 23.24.
...Strut.

Claims (1)

[Claims] a voice pattern creation section that converts a voice signal into an input voice pattern and a voice standard pattern as a time series of feature vectors; and a standard pattern registration section that stores a plurality of the voice standard patterns; A speech recognition device that determines input speech by performing pattern matching with a standard speech pattern includes a sonic oscillator that emits sound waves to the lips of a speaker, and a system that is integrated with the sonic oscillator and that emits sound waves from the sonic oscillator. an audio input unit that detects the utterance of the speaker along with a reverberant sound generated by the sound waves incident on the lips or the like of the speaker; and an input time interval of the audio signal input from the audio input unit and oscillation of the sound wave oscillator. an audio input switching section that switches between the oscillation of the sound wave oscillator and the input of the audio signal by setting a time so that the time intervals do not overlap; and an audio input switching section that is connected to the audio pattern creation section and stores vibration noise when no audio is input. and a vibration noise storage unit, and when a voice is input, the voice signal is corrected using the numerical value stored in the vibration noise storage unit to create an input voice pattern and a voice standard pattern. Speech recognition device.