JPS61273599A - 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 Application of the Invention] The present invention targets monosyllables constituting a speech signal. 0 Regarding a speech recognition device configured with a processing circuit [Background of the invention] Conventionally, for a speech signal that is uttered by separating each - word,
In the so-called monosyllabic recognition method, which recognizes each monosyllable individually, an envelope signal is formed from the speech signal of each word, and the rising region of the envelope signal is used as the region from which consonant information is obtained, and the maximum of the envelope is A method has been used to determine the spectral shape or change in the spectral shape of the original speech signal in each region from the point in time when the amplitude value is present to the region from which vowel information is obtained. Then, the degree of similarity with standard characteristic values or patterns set in advance is calculated, the closest monosyllable is selected, and this is identified as the monosyllable of the input speech.

There are five types of vowels in Japanese: ``aiueo.'' When a vowel itself without a consonant is uttered, the spectral shape of each vowel differs greatly, so it is easy to recognize the vowel based on the spectral shape. However, in monosyllables with consonants, especially in voiced sounds such as ``zajizuzezo'', the spectral shape of the vowel region is significantly deformed compared to the shape of the vowel itself, even though they are the same vowel. This is one of the factors that worsens the vowel recognition of the entire monosyllable.

In order to obtain as much of the same information as possible for the same vowel regarding the above-mentioned transformation of the vowel region spectrum, various calculation processing methods have been conventionally applied to the spectrum.

One of the representative methods is the method by Kim, Makino, and Kido in IEICE Technical Report, EA-5i-6B, P57-64 (1985). As described in the literature, a method has been proposed in which a straight line determined by the least squares method is drawn on the spectrum, this straight line is used as a threshold, and the position of the local peak of the spectrum exceeding this value is determined. has been done. As shown in Figure 2, this method allows the detection of peak positions in spectrum 1 to be easily performed using straight line 2, and even when the slope of the entire spectrum changes variously, the peak position This has the advantage that it is possible to easily obtain frequency information where .

Another method is the Electronic Technology Research Institute report, /11
1841. As described in the document "Basic research on automatic recognition of speech based on phonemic units" by Nichichi in P1-122 (1984), a peak location with a predetermined frequency width is determined from the spectral shape, and this peak location is A method has been proposed that emphasizes the pseudo formant frequency. If the emphasis at the peak point is appropriately increased, there is an advantage that frequency information at which pseudo formants appear can be easily obtained even when the slope of the entire spectrum varies in the same way as described above.

However, since all of the above methods require complex calculations based on spectral amplitude data, the calculation time is not negligible from the viewpoint of shortening the recognition calculation time.

Further, within the range of so-called 50 sounds, the change in spectral shape between vowels is small, and the spectral shape itself is useful information for identifying monosyllables. Therefore, the method of eliminating the spectral shape as described above causes a problem in that useful information is lost.

[Purpose of the invention]

In order to eliminate the above-mentioned conventional problems, an object of the present invention is to provide a speech recognition system that makes it possible to obtain information about vowels in a timely manner by providing a signal processing circuit after an analysis circuit that analyzes the frequency of a speech signal. We are in the process of providing equipment.

[Summary of the invention]

As mentioned above, in order to deal with the situation where the slope of the entire vowel spectrum changes in various ways, and to remove this slope, the spectrum shape is regarded as one waveform, and this signal waveform is processed by a high-pass filter.ノ(
After being analyzed by the NONDONOKUSUFIK1 group, the individual filter output signals are converted to a filter with a predetermined time constant: c 7
-) A forming circuit generates an envelope signal, and a multiplexer sweeps the envelope signal at predetermined time intervals to obtain a time-series signal ranging from a low frequency to a high frequency. When the time axis of this signal is replaced with the frequency axis, it becomes a frequency characteristic. If this time-series signal is processed by a high-pass filter, large changes over the entire time-series signal can be removed and only local changes can be obtained.

FIG. 5 shows an example of signal processing using a high-pass filter. Figure 3 (2) shows the circle and the vocalized signal.
This is a time-series signal obtained in the vowel region by a bandpass filter followed by a detection circuit and a multiplexer. The signal obtained by passing this signal through a high-pass filter having a predetermined cutoff frequency and attenuation characteristic is the signal shown in FIG. This eliminates the above-mentioned slope across the entire frequency range.

In order to perform monosyllable recognition calculations based on this high-pass filter passed signal, it is necessary to normalize the amplitude immediately beforehand. This is because the phonation intensity of the voice changes with each utterance, so this change is removed. After that, it is necessary to perform recognition calculations. Comparing this normalized nine-level data with the data uttered as "a", the shape matches well with the shape obtained from the vowel region data of "za".

[Embodiments of the invention]

The overall structure and operation of the apparatus of the present invention will be explained below with reference to FIG. The uttered audio signal is sent to microphone 3 in the figure. The signal is input to an envelope forming circuit 5 and a group of bandpass filters 10 via a microphone amplifier 4.

When the amplitude of the audio envelope formed in the envelope forming circuit 5 exceeds a predetermined reference value, that information is transmitted to the controller 9 by a trigger pulse, and the controller starts operating.

At the same time, the audio envelope signal is sent to the A/D converter 6.
is recorded in the memory 7 as a digital signal.

The time constant in the above envelope forming circuit 5 is 15 m.
The sampling period in the A/D converter 6 is also set to 15 msec.

Since the rise time of a monosyllable is usually about aom seconds,
Here, the rising region is divided into approximately five parts, and the average spectrum at each time interval is measured. In rare cases, it may rise within 15 msec, or it may take 150 msec. Therefore, a time constant of 15 msec is chosen in order to smooth out spectral fluctuations occurring within 15 msec. The memory 7 is a backup memory capable of storing approximately 8 seconds of audio envelope signals. This memory 7 is the first α36
Once the signal is filled with a second signal, the envelope parameter calculation section 8 starts calculation. Q. In 36 seconds, 24 average spectra obtained in 15 m seconds are obtained, and from this number, one or two monosyllables are finely divided to obtain a time series spectrum.

Among the syllables existing in the above 0.56 seconds, the temporal position where the first single syllable is extracted and shows the maximum value of the envelope amplitude, and the maximum amplitude value is determined as 1.0.9, O at the rising edge of the envelope, The temporal position indicating the amplitude of S is determined, and these values are transmitted to the spectrum calculation section 16 via the controller 9.

On the other hand, in the signal analysis section 20, each output signal of the audio signal transmitted to the bandpass filter group 10 is converted into an envelope signal by the envelope forming circuit 11.

The contents of the band-pass filter group 10 described above are such that one band-pass filter has a frequency band of 173 octaves, ranging from 200...2 to 5K11z, and consists of a filter group of 15 channels. Further, the time constant of the envelope forming circuit 11 is the same as the time constant of the envelope forming circuit 5, and corresponds to the envelope parameter calculated by the envelope forming circuit 5.

The output signal of the envelope forming circuit 11 is scanned by a multiplexer 12 at a frame interval of 15 msec.
is processed by the memory 1 through the A/D converter 14.
Recorded in 5. The function of this high-pass filter 13 will be explained in detail later.Recording in the memory 15 is started when a signal having an amplitude equal to or higher than a preset reference amplitude value is input, similarly to the memory 7. The memory 15, like the memory 7, is a buffer memory capable of storing approximately 8 seconds of voice analysis data. When the data for α36 seconds is recorded in the memory 15 and the calculation by the envelope parameter calculation unit 8 has been completed, the spectrum calculation unit 1
Start the calculation in step 6. If the envelope calculation has not yet been completed, wait for it to be completed, and after the calculation is completed, the spectrum calculation unit 16 calculates envelope data for 0.36 seconds according to instructions from the controller 9.

The calculation in the spectrum calculation section is to obtain a spectrum for 30 m seconds from the envelope maximum value obtained in envelope parameter calculation W58, and to obtain a difference spectrum by calculating the difference between this spectrum and the spectrum at the envelope amplitude value I15 described above. Here, the spectrum for 30 m seconds is obtained by averaging two lines of data obtained by scanning for 15 m seconds in the multiplexer 12. Next, the envelope amplitude value is 0.5
Find the spectrum in the final region of a single syllable. In the end, the following three types of spectra will be recorded in the pattern meso 17.

(1) Spectrum at the maximum amplitude value of the audio envelope.

(2) Difference spectrum (5) Ending region spectrum Here, (1) is the spectrum of the vowel region, and (2)
represents the change in the spectrum in the rising region of the audio signal, and consonant information can be obtained from this. (3) is used to determine whether the vowel in the final region is the same as (1) or different. By calculating the similarity, if the vowel is the same, it is a monosyllable with one vowel, and if it is different, it is a monosyllable with one vowel. "Kiya" included
It can be determined that it is a monosyllable such as "kiyu". Here, it is necessary to normalize the amplitudes of the three types of spectra calculated. The method is the sum of squares of the individual amplitude values,
That is, individual amplitude values are determined so that the power over the entire frequency range is a constant value. As a result, the next calculation in the pattern matching section 18 will be performed with high accuracy.

Each time the calculation is completed in the spectrum calculation unit 1, it is recorded in the pattern memory 17. When the recording is completed, calculation in the pattern matching section 18 is started.

The pattern matching section calculates the degree of similarity with the pattern in the standard pattern memory 25 stored in advance, first identifies the vowel of the input monosyllable, then identifies the consonant, and finally identifies the vowel in the input monosyllable. It determines whether there are one or two syllables, and identifies the input monosyllable by combining the results. The six results are sent to the output terminal 21 via the output buffer 19. Output from

Next, the function of the high-pass filter 1S will be explained using FIG. 4.5. 22 in FIG. 4 is a multiplexer 22
represents the output signal of Since the scanning time of the multiplexer is 15 msec, the bandpass filter output signal is sequentially output from the low frequency to the high frequency every 1 msec. 15 msec corresponds to 66.7Hz. Therefore, the cutoff frequency of the high pass filter is set to 250.
Hz and have an attenuation characteristic of 12 dB10at, the attenuation will be approximately 22 dB at 66.7111z.

That is, the large variation across the frequency range is 20 dB.
It becomes possible to attenuate the amount by more than 100%.

FIG. 5 shows a 12 dB 10 ct high pass filter. In the figure, 23 is an operational amplifier that constitutes a buffer amplifier, and 24 is an operational amplifier that constitutes a high-pass filter. Examples include, but are not limited to, a method of sweeping the center frequency using a bandpass filter with a variable center frequency, and a Fourier spectrum using discrete Fourier transform.

Spectral envelopes based on the linear prediction method can be handled in the same way by converting the spectral values into time series data.

Please note that in this example, an analog filter is used.
The output signal of multiplexer 12 is transferred to memory 1
5 and then remove the low frequency components using a digital filter, the same effect will be obtained.

Furthermore, it is also possible to perform calculations equivalent to filtering in the spectrum calculation section 16 without using a filter. However, in this case, it does not contribute to shortening the calculation time, and it is only possible to take advantage of the recognition calculation based on pattern matching, which maintains the spectral shape and plans the shape of the entire spectrum.

[Effects of the Invention] Conventionally, in response to the phenomenon that the spectrum in the vowel region of a single syllable changes greatly depending on the type of consonant, the characteristics of the vowel are extracted through relatively complicated calculations, which requires an unreasonable amount of calculation time. However, with the present invention, it has become possible to extract vowel features in real time with a simple circuit configuration.In contrast, in the past, the spectral shape was discarded and only the features were extracted, but the present invention According to this method, since the spectral shape of the vowel can be maintained, recognition calculations that take advantage of the pattern matching method's ability to evaluate the overall spectral shape were made possible.

[Brief explanation of the drawing]

Fig. 1 is a circuit block diagram of the present invention, Fig. 2 is a waveform diagram showing a conventional feature extraction method in a vowel spectrum, Fig. 5 is a waveform diagram showing the functions of the present invention, and Fig. 4 is a device of the present invention. FIG. 5 is a waveform diagram of a multiplexer output signal in FIG. 5, and FIG. 5 is a high-pass filter circuit diagram in the apparatus of the present invention.

Claims (1)

[Claims] A speech recognition system that recognizes the input speech by frequency-analyzing the input speech, generating multiple spectral patterns ranging from consonants to vowels, and calculating the degree of similarity with pre-stored standard spectral patterns. The apparatus includes means for frequency-analyzing the audio signal, means for converting the frequency-analyzed amplitude or power spectrum value into time-series data, and a high-pass filter for removing low frequency components of the time-series data. , a speech recognition device characterized in that it has means for storing the time-series data or the high-pass filter passed signal, and stably extracts spectral features.