JPH0462598B2 - - Google Patents

Description

[Detailed Description of the Invention] [Field of Application of the Invention] The present invention targets monosyllables of speech and accurately quantifies the amount of change in the spectrum in the so-called "crossing" region, from the consonant region to the vowel region. , relates to a speech recognition device that recognizes speech by calculating the degree of similarity with standard characteristics prepared in advance.

[Background of the invention]

Conventionally, to recognize monosyllabic speech, the input signal is first divided into a consonant region and a vowel region, and the spectra of each consonant and vowel, or multiple time-series spectra spanning from the consonant to the vowel region, are obtained.
A method is being used to calculate the similarity between these and a standard spectrum group.

One of the typical methods for dividing an input signal into consonant and vowel regions is the one described in the Journal of the Acoustical Society of Japan, Vol. 40, No. 2.
As described in P63-70 (1984), the envelope of the input signal is determined, and shift matching is performed between this and a standard envelope prepared in advance in the rising region of the envelope.

Here, the principle will be explained using FIG. 2.
The solid line 1 in FIG. 2A represents the rising edge of the envelope in a monosyllabic signal. For this solid line 1, the standard envelope 2 is set in advance.
By setting and sequentially shifting this on the time axis, the distance between the solid line 1 and the dotted line 2 is sequentially calculated. The calculated distance becomes a solid line 3 as shown in FIG. 2B, and this minimum point 4 becomes the reference point for consonant region detection. The standard envelope corresponding to point 4 is 2' in FIG. 2A. The range of consonants and the range of vowels can be determined from the standard envelope 2'.

By the method described above, it is possible to roughly set the consonant region of a monosyllable. Furthermore, individual differences in speaking speed and voice waveform can be dealt with by preparing a plurality of standard envelopes.

On the other hand, as an improvement plan of the above method,
As described in Publication No. 26797, division of consonant and vowel regions is achieved by setting reference values that define vowel and consonant regions near the maximum and minimum values of the envelope signal for each input voice. I am doing it. Figure 3 shows the relationship between the minimum value Ec and maximum value Ev in the standard envelope.

Once a consonant is detected, the spectrum of the consonant region is determined. Using the above method, it is in principle easy to obtain the spectrum of consonant and vowel regions, and it is an excellent method.

However, the maximum amplitude value of the speech envelope varies greatly depending on the strength of the speaker's vocalization and the type of monosyllable, so if the amplitude of the speech envelope is small, consonants cannot be detected using simple methods. A problem arises in that the voice is buried in ambient noise other than the voice. This problem will be explained in detail with reference to FIG.

FIG. 4A shows an example of analyzing a monosyllabic envelope 1 having a relatively large amplitude. The period from the time when the envelope intersects with a preset amplitude reference value (trigger level) to the time indicated by position 5 indicating the maximum value of the envelope is defined as the consonant region and the transition region from the consonant to the vowel.

For a speech signal such as that shown in FIG. 4A, which has an amplitude value sufficiently larger than the reference amplitude value, the consonant region and the following transition region can be determined by a simple method of starting the analysis from the signal that has reached the reference amplitude value. The vowel region can be input to the signal analyzer. If the analysis interval in which the frequency is analyzed in a predetermined time width from the point where the reference amplitude value and the envelope intersect is defined as a, then the consonant region is analyzed. Next, if the analysis interval to be analyzed at a predetermined time interval from time 5, which is the maximum value of the envelope, is c, then the vowel region will be analyzed.

On the other hand, when the amplitude of the audio signal is small as shown in FIG. 4B, the amplitude of the consonant region is smaller than the reference value, making it impossible to analyze the consonant region. The analysis interval that can be set from the point where the reference value and the envelope intersect is like b, and this is an analysis in the interval from the consonant to the vowel region. On the other hand, the analysis interval of the crossover area in diagram A is a'.

As mentioned above, when the amplitude of monosyllabic speech is small, the problem arises that it becomes impossible to detect the consonant region, or that the consonant region is buried in ambient noise as described above.

Journal of the Institute of Electronics and Communication Engineers, Vol J65- as a method for extracting changes in characteristics from the consonant region to the vowel region
As described in A, P1278-1285 (1982), a method of detecting consonant groups such as "voiceless consonant,""fricative," and "voiceless plosive" is carried out using a discrimination filter on the time axis. It is. This method requires a discrimination filter group in addition to the frequency analysis filter group, and in order to identify individual monosyllables, it is necessary to prepare filters for the number of monosyllables.

[Purpose of the invention]

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems in frequency analysis of an input audio signal ranging from consonants to vowels.

[Summary of the invention]

In order to achieve the above object, the present invention calculates the amount of change between the spectrum in the crossing region and the spectrum in the vowel region, uses this and the shape of the spectrum itself as data for recognition, and compares this with standard characteristics prepared in advance. To provide a speech recognition device capable of performing highly accurate recognition by calculating similarity.

Note that it is desirable that the temporal position of the analysis interval and the like with respect to the crossing interval be the same regardless of the type of syllable. For example, if you normalize the maximum amplitude of the audio envelope to 1.0. Set the analysis interval of the crossing region to around 0.5 amplitude.

Hereinafter, an outline of the present invention will be explained using FIGS. 5 and 6. FIG. 5 shows a schematic block diagram of the signal processing circuit. The audio signal inputted from the microphone 7 is made into a predetermined signal amplitude by the microphone amplifier 8, and is made into an envelope signal by the envelope forming circuit 9. When a signal with a predetermined amplitude or more is input, this becomes a trigger signal and recording to the memory 10 is started. After storing signals for a predetermined time interval, the stored data is input to the envelope parameter calculating section 11 to start calculation. The contents of the calculation include the temporal position of the maximum amplitude value in the envelope of the monosyllabic voice, the temporal position of the transition region, and the end position of the monosyllable. On the other hand, the other output signal from the microphone amplifier 8 is analyzed by the frequency analyzer 12, and the result is stored in the memory 13.
recorded in The amplitude of the trigger signal that signals the start of recording analysis data is the same as in the memory 10 and is synchronous, so
0 and the signals of the memory 13 are synchronized. When the recording for a predetermined time interval is completed, the difference spectrum calculation unit 14 performs calculation using the results in the envelope parameter calculation unit 11. The content of the calculation is to calculate the difference between the spectrum of the crossover region and the vowel region spectrum, obtain at least one spectrum before the end of a single syllable, and calculate the degree of similarity between this spectrum and the vowel region spectrum.

Next, the contents of calculation in the envelope parameter calculation section 11 will be explained in detail with reference to FIG. 6. FIG. 6 shows the envelope waveform when two monosyllables are uttered in succession. When a voice is uttered and the envelope reaches a reference value, recording of data to the memory 10 begins. The sampling period is, for example, 15 msec, and the point where the envelope curve and the vertical line intersect in FIG. 6 becomes amplitude data.
Each data interval a ₁ , a ₂ , . . . represents an analysis interval as a temporal position in the signal analysis section 12. From the data group recorded in the memory 10, it is first identified that the envelope forms two peaks consisting of two monosyllables, and then it is determined that the top of the first peak is at the position indicated by point 5. Obtain by calculation. As a result, it is determined that the vowel region is _a6 .

Also, the first crossing area is shown in Figure 6.
Corresponds to section a ₄ . Here, regarding a ₄ , the average value of the amplitude values at the beginning and end of each section from a ₁ to a ₆ is the amplitude value at point 5.
An interval close to 0.5 is selected. After confirming that _a10 is a region where two syllables are joined, the position of _a8 is used as information as the region where a single syllable ends. Analysis interval a ₄ as temporal position,
a ₆ and a ₈ specify necessary data from a series of analysis data.

From b ₁ to b ₆ is the second monosyllabic section, and the parameter calculation in this area is performed by inputting envelope data of, for example, 0.36 m seconds again to the envelope parameter calculation unit 11 with b ₁ as the first section. Execute. The analysis intervals of the crossover area and the vowel area obtained as a result of calculation are respectively b ₁
and b ₃ .

[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 inputted via the microphone 7 and the microphone amplifier 8 to the envelope forming circuit 9 and the group of bandpass filters 15. When the amplitude of the audio envelope formed in the envelope forming circuit 9 exceeds a predetermined reference amplitude value, that information is transmitted to the controller 24 by a trigger pulse, and the controller starts operating. At the same time, the audio envelope signal is recorded in the memory 10 as a digital signal via the A/D converter 18. The time constant of the envelope forming circuit 9 described above is 15 msec, and the sampling frequency of the signal in the A/D converter 18 is the same.
It is 15m seconds. Since the memory 10 can store approximately 8 seconds of audio envelope signals, it functions as a buffer memory. When this memory 10 is filled with the signal for the first 0.36 seconds, the envelope parameter calculation section 11 starts calculation.
There are usually one or two monosyllables in the 0.36 second time interval, and as will be explained in detail later, the first syllable is extracted and its parameters are calculated. The calculated parameters are sent to the controller 24
The signal is transmitted to the difference spectrum calculation section 14 via.

On the other hand, each filter output signal for the audio signal transmitted to the bandpass filter group 15 in the analysis section 12 is converted into an envelope signal by the envelope forming circuit 16. The contents of the above bandpass filter group 15 are such that one bandpass filter has a frequency band of 1/3 octave,
It consists of a group of 15 channels of filters ranging from 200Hz to 5kHz. Further, the time constant of the envelope output signal is the same as the time constant of the envelope forming circuit 9, and corresponds to the envelope parameter calculated by the envelope forming circuit 9.

The output signal of the envelope forming circuit 16 is scanned by the multiplexer 17 at a frame interval of 15 msec, and is recorded in the memory 13 as the output of the analysis section 12 via the A/D converter 19. In this case, as in the case of the A/D converter 18, the operation is started when a signal having a preset reference amplitude value or more is input. Memory 13 is memory 10
Similarly, it has a buffer memory that can store approximately 8 seconds of voice analysis data. 0.36 to memory 13
If data for a second has been recorded and the calculation by the envelope parameter calculation section 11 has been completed, the calculation in the difference spectrum calculation section 14 is started according to instructions from the controller 24. If the calculation of the envelope parameters has not been completed, wait for the calculation to be completed, and after the calculation is completed, data for 0.36 seconds is calculated in the differential spectrum calculation unit according to instructions from the controller 24. The calculation result is recorded in the pattern memory 20, and when the recording is completed, 0.36 seconds of envelope data including the second monosyllable data at the beginning is input to the envelope parameter calculation section 11 according to instructions from the controller 24.

When the pattern memory 20 is filled with data,
The pattern matching section 21 calculates the degree of similarity (distance) between the standard pattern previously stored in the standard pattern memory 22 and the sequential pattern, and identifies the monosyllable of the input signal.
The obtained result is outputted from the output terminal 25 via the input buffer 23.

Next, the contents of calculation in the envelope parameter calculation section 11 will be explained using the schematic flowchart of FIG. 7.

Calculations are performed to detect large peaks and valleys in the first 0.36 seconds of data in the uttered audio envelope signal, ignoring local irregularities in the envelope shape. As mentioned above, there are usually one or two monosyllabic signals in an interval of 0.36 seconds,
From this, the first peak of the envelope signal corresponding to the first monosyllable is detected. In the flowchart of FIG. 7, step 2 instructs that the calculation process starts from the first data among the data recorded in the memory. In the step, the data number of the peak position where the amplitude shows the maximum value in the first peak is detected. Here, if the rise of the envelope signal is rapid and the first data indicates the maximum amplitude value, the second data is set as the peak position of the first mountain. In step 2, the position of the crossing area is detected. This position exists between the first data and the peak position data, and a data number showing a value close to 1/2 of the maximum amplitude is searched. Next, in step (2), a data number indicating the position of the connecting region of the first and second peaks, that is, the valley, is searched. If this data number is n, then
When performing calculation processing on the second monosyllable with I=n, this is to notify the controller that n is the first data. In step , a data number indicating the end area of the first mountain is searched. For this area, a data number located midway between the position of the first peak and the valley (connected area) is selected. In step , the results obtained so far are transmitted to the controller 24. When the transmission is finished,
If there is no instruction to stop the calculation from the controller, the calculation for the second syllable is started from the step with I=n. When there is a stop instruction from the controller, this is the case when processing for all data in the memory 10 has been completed, and when the calculation in the difference spectrum calculating section 14 is continuing.

Next, the content of calculation in the difference spectrum calculation section 14 will be explained. The spectrum at the peak position of the envelope signal is assumed to be the vowel region spectrum. This spectrum is used as a comparison standard, and in order to improve the precision of the shape, the average value of the spectrum values of the analysis interval of the peak position and the next analysis interval (a ₆ and a ₇ in Figure 6) is calculated. This is the reference spectrum for the vowel region. That is, the analysis interval is set to 30 msec. Next, the difference between this reference spectrum and the cross-region spectrum (analysis interval 15 msec) is determined. The details of this calculation will be explained using FIG. FIG. 8A shows the spectral shape of a certain monosyllable. The vowel region spectrum is 26, and the crossing region spectrum is 27. When we find the difference between the two, spectrum 2 in diagram B is obtained.
8 is obtained. The shape of the spectrum 28 is one powerful piece of data for determining what the input monosyllable is.

In the difference spectrum calculation section 14, one more
Two calculations are performed. It is a distance calculation between the vowel region spectrum and the ending region spectrum. If the distance is smaller than the threshold, it is determined that there is no change in the spectral shape, and the spectral shape is determined to be a monosyllable containing one vowel. Furthermore, if the threshold value is exceeded, it is determined that the syllable is a monosyllable containing approximately two vowels, such as "kiya" or "kiyu". Alternatively, it is determined that the signal is added after a single syllable such as "n" which originally has a small signal amplitude.

In addition to the above calculation results, the data recorded in the pattern memory 20 is summarized as follows.

(1) Spectrum of the vowel region (2) Spectrum of the crossing region (3) Difference spectrum of (1) and (2) Based on the result of similarity calculation between the ending region spectrum and the vowel region spectrum, if necessary, (4) End Range Spectrum A calculation method in the pattern matching section 21 using the above four types of characteristics will be described below. The recognition calculation method used here is based on the conventional pattern matching method. First, the degree of similarity (distance) between the vowel region spectrum recorded in the pattern memory 20 and each standard spectrum of five types of vowels "aiueo" and "n" is calculated, and the most similar (closest distance) ) Select a vowel. When one vowel is selected,
Using the crossing region spectrum and the difference spectrum,
When the consonant series of the selected vowel, for example "a", is selected, in the series "a, ka, sa, ta, na,...", the standard spectrum of the transition area prepared in advance and the standard spectrum of the difference are used. Calculate the similarity of . Next, the results for the two characteristics are combined to select one single syllable. Here, if we obtain the information that the final region spectrum and the vowel region spectrum match as mentioned above, we can use the monosyllabic group formed by one vowel. 、Sa、Ta、Na...、Ba、
Calculations are made with 15 monosyllabic groups of 'pa'. Also, if you get information that there is no match,
Calculations are performed on 12 groups of monosyllables ``kiya, shiya, chiya, ..., piya'', which are roughly formed by two vowels, and one monosyllable is selected.

〔Effect of the invention〕

According to the present invention, the spectrum of the crossover region, which is the intermediate region between the consonant and vowel of a monosyllable, can be obtained using a relatively simple method. We were able to obtain powerful data for monosyllable recognition, and were able to improve recognition accuracy.

The overall shape of the spectrum in the region transitioning from the crossing region to the vowel shows the shape of the vowel spectrum, with the influence of the consonant spectrum added thereto. Therefore, the difference spectrum in the present invention extracts only the influence of the consonant spectrum, and is a superior method compared to the method of identifying monosyllables based on the original spectrum shape itself. It provides:

In addition, compared to the conventional method of obtaining time-series spectra from consonants to vowels and comparing them with time-series standard spectra, pattern matching of differential spectra requires much simpler calculations, making it an economical method. There is.

[Brief explanation of drawings]

Figure 1 is a block diagram of the recognition device of the present invention, Figure 2 is a block diagram of the recognition device of the present invention.
The figure shows a conventional example of a method for dividing consonants and vowels, Figure 3 shows another conventional example of a method for dividing consonants and vowels, and Figure 4 shows problems in consonant region detection. Figure 5 is a main block diagram of the present invention.
FIG. 6 is an explanatory diagram of the calculation processing method for the audio envelope, FIG. 7 is a flowchart showing the calculation processing procedure for the audio envelope, and FIG. 8 is an explanatory diagram of the difference spectrum in the present invention. 9... Envelope formation circuit, 10, 13...
Memory, 11...Envelope parameter calculation unit, 14...Difference spectrum calculation unit, 15...Band pass filter group, 16...Envelope formation circuit, 17...Multiplexer, 18, 19...
A/D converter, 20... pattern memory, 21...
Pattern matching section, 22...standard pattern memory, 23...output buffer, 24...controller.

Claims (1)

[Claims] 1 Analyze the frequency of input speech and generate multiple spectral patterns ranging from consonants to vowels,
A speech recognition device that recognizes the input speech by calculating the degree of similarity with a standard spectral pattern stored in advance, comprising: means for forming an envelope of the speech signal; In the amplitude data, a predetermined time interval from the time point showing the maximum amplitude value is defined as the vowel region, and a region approximately halfway between the start point of the single syllable and the time point showing the maximum amplitude value is defined as the crossing region. means for deriving the temporal position of the region, means for storing the temporal position, means for individually detecting at least two types of spectra of the vowel region and the crossing region, and the vowel region spectrum and the crossing region. 1. A speech recognition device comprising means for determining a difference spectrum from a region spectrum, and using at least the two types of spectra and the difference spectrum as data for recognition.