JP2006343544A - Voice recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To discriminate voices by a defined reference irrespective of the level of audio signal. <P>SOLUTION: A ratio of amplitude or power of a fundamental wave and each higher harmonic component to the total sum of amplitude or power of the fundamental wave and each higher harmonic component included in a voice frequency region is acquired as a contribution ratio, and the phonemes of a consonant and a vowel are specified by the manner in the presence of the contribution ratio not affected by the level of the audio signal. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a speech recognition method capable of recognizing a language from a speaker's voice using a simple processing device. More specifically, the present invention relates to a speech recognition method that can perform analysis based on the same reference regardless of the size of a speech signal.

  Conventionally, as a speech recognition method, a method has been proposed in which a vowel region and a consonant region are separated from a speech waveform, and the vowel and consonant can be identified and recognized from the waveform of the separated vowel region and the waveform of the consonant region. (For example, refer to Patent Document 1).

  In addition, as a method for identifying the vowel and consonant, for the separated vowel region, the time from when the speech signal level of the speech waveform passes through the voltage zero volt until it passes through the positive voltage region and again passes the voltage zero volt is detected. The vowel is identified, and for the separated consonant region, the voice signal level of the speech waveform passes through the voltage zero volt or rises from near the voltage zero volt, then moves through the positive voltage region and again passes through the voltage zero volt or near the voltage zero volt. There has also been proposed a method for identifying the consonant by detecting the time required to reach (see, for example, Patent Document 2).

Japanese Patent Laid-Open No. 9-101797 JP 2001-265379 A

  All of the above conventional speech recognition methods attempt to identify vowels and consonants separately, but identify them based on the speech waveform itself collected from a microphone or the like. For this reason, there is a problem that accurate discrimination is difficult in an environment in which conditions such as daily conversation are uncertain because it is easily affected by the magnitude of the voice (the magnitude of the audio signal).

  The present invention has been made in view of the problems of the above-described conventional speech recognition method, and an object thereof is to make it possible to identify speech on a constant basis regardless of the size of the speech signal.

  For the above purpose, the present invention performs frequency analysis on audio data groups sampled from an audio signal and A / D-converted, and from the amplitude spectrum or power spectrum obtained thereby, the fundamental wave included in the audio frequency domain and each The ratio of the amplitude or power of each of the fundamental wave and each harmonic component to the sum of the amplitude or power of the harmonic components is obtained as a contribution rate, and the phoneme of the consonant and the vowel is specified from the appearance of this contribution rate. A characteristic speech recognition method is provided.

In the present invention, the speech data group is divided into a consonant region and a vowel region, and the contribution rate is obtained by frequency analysis of the speech data group of the consonant region and the speech data group of the vowel region, and the contribution in each speech data group Identifying the phonemes of consonants and vowels from the way the rate appears,
As a contribution rate, using the ratio of the amplitude of each of the fundamental wave and each harmonic component to the sum of the amplitudes of the fundamental wave and each harmonic component included in the audio frequency domain,
The frequency analysis is sequentially performed for each analysis section of the N pieces of voice data for the voice data group, and a contribution rate is obtained for each analysis section.
Performing frequency analysis after performing window function processing on the audio data group,
Use a Hamming window for window function processing, use fast Fourier analysis for frequency analysis,
Is included as a preferred embodiment thereof.

  The speech recognition method of the present invention specifies consonants and vowels using the contribution rate.

  By the way, the contribution rate in the present invention is the ratio of the amplitude of each of the fundamental wave and each harmonic component to the sum of the amplitudes of the fundamental wave and each harmonic component contained in the speech frequency region, or the fundamental contained in the speech frequency region. The ratio of the power of each of the fundamental wave and each harmonic component to the total power of the wave and each harmonic component. Since the contribution rate is the ratio as described above, it is a value that is not affected by the magnitude of the audio signal, and the present invention performs speech recognition based on this contribution rate. Therefore, it is possible to identify with high accuracy regardless of the size of the audio signal.

  The basic procedure of the speech identification method according to the present invention will be described. In addition, description here demonstrates as an example the case where a test subject has sampled and emitted one Japanese syllabary sound for convenience of explanation.

  First, an example of the speech recognition method according to the present invention will be described with reference to FIG.

  The audio signal is collected as an analog signal using, for example, a microphone, amplified or filtered as necessary, sampled, A / D converted, and temporarily recorded as an audio data group in a memory.

  The frequency that needs to be analyzed for recognizing speech differs somewhat depending on the language, but for example, in Japanese, it may be necessary to have a frequency of about 5 to 5.5 kHz. In addition, in order to correctly obtain the frequency component included in the continuous signal as sampling data, the sampling frequency must be at least twice the upper limit of the frequency of the continuous signal, so the sampling frequency is 10 kHz or more. It is preferable. In a specific example to be described later, sampling is performed at 50 kHz, but it is not necessary to set the frequency as high as practical. Moreover, it is preferable to cut in advance a frequency component exceeding 1/2 of the sampling frequency with a filter (low-pass filter).

  The audio data stored in the memory is usually extracted and subjected to frequency analysis except for the no-signal area (silence area) that exists at the beginning and end, but in order to minimize the frequency analysis error, the pre-processing is performed. It is preferable to perform window function processing.

  As window functions when performing window function processing, there are Hanning window, Hamming window, Blackman window, rectangular window, etc., and any of them can be used. The most commonly used Hamming window is preferred.

  When the Hamming window is used, if the original audio data value is d, the data number is n, and the number of data used for frequency analysis is N, the converted data X is as follows.

X = d × [0.54-0.46 × cos {2 × π × n / (N−1)}]
n = 0 to (N-1)

  When the number N of data used for frequency analysis is small, the frequency resolution decreases, but the time resolution within the analysis interval appears greatly. Conversely, when the number of data N is large, the time resolution in the analysis interval is small, but the frequency resolution is improved. If the frequency resolution is excessively reduced, it will be difficult for the contribution rate described later to correctly reflect the frequency component contained in the speech waveform, and it will be difficult to grasp the characteristics of how the contribution rate appears, so that a resolution of 40 to 100 Hz can be obtained. It is preferable to adjust the number of data according to the sampling frequency. In the case of fast Fourier analysis, the number of data is an integer power of 2.

  The audio data group is subjected to the window function processing as necessary and then subjected to frequency analysis to obtain an amplitude spectrum and / or a power spectrum. For this frequency analysis, Fourier analysis, particularly fast Fourier analysis with a short processing time is preferable.

When Fourier analysis (fast Fourier analysis) is used, m = 1 for N pieces of voice data (for example, the number of integer powers of 2 such as 512, 1024, etc.) in the voice data group in one analysis. (m is the order) the fundamental frequency and the frequency (harmonic) of integer multiples of the fundamental frequency (the next multiple of the fundamental frequency), the coefficient b _m of the coefficient a _m and the cosine wave component of the corresponding sine wave component when the Is obtained. Then, using these coefficients, the amplitude spectrum X _m and the power spectrum X _m ² can be obtained as follows. Note that m = 0 corresponds to a DC component.

^{_{X m = √ (a m 2}} + b m 2)
X _m ² = a _m ² + b _m ²

The contribution ratio in the present invention is the ratio C of each amplitude spectrum X _m to the sum of the amplitude spectra X _m of each harmonic component of m = 2 or more from the fundamental frequency of m = 1, or m from the fundamental frequency of m = 1. = The ratio C ′ of each power spectrum X _m ² to the sum of the power spectra X _m ² of each harmonic component equal to or greater than 2 can be obtained. C or C ′ may be obtained as a ratio or as a percentage. When calculating | requiring as a ratio, it becomes a following formula. When obtained as a percentage, each value is multiplied by 100.

C = (1 / ΣX _m ) × (X _m )
C ′ = (1 / ΣX _m ² ) × (X _m ² )

  The upper limit of m depends on the frequency resolution when performing frequency analysis, but is sufficient up to the order that can be analyzed up to the frequency necessary for speech recognition. Specifically, if the sampling frequency is 50 kHz and the number of data N is 1024, the order obtained by frequency analysis is (1024/2) −1 = 511, but the frequency resolution = 50000 ÷ 1024≈48 Hz. And, as described above, if Japanese speech recognition needs to be analyzed up to about 5.5 kHz, m = 5500 ÷ 48≈114.

  The contribution ratio used in the present invention may be either C or C ′. However, in the case of C ′, the fluctuation appears almost squared, so that the magnitude of the frequency component is more easily emphasized than the amplitude spectrum. It is preferable.

  For example, the frequency analysis may be performed only once for N pieces of sound data in an appropriate region of the sound data group. However, since the number of data N and the number of samples are usually determined so that the sound data group is a collection of more than N sound data, N sound data is defined as one analysis section (one frame), and each analysis section It is preferable that the entire audio data group is divided into a plurality of times and the frequency analysis is performed while shifting the predetermined number of audio data. Thus, the accuracy can be improved by analyzing the entire audio data group. In this case, the contribution rate is obtained for each analysis section. When the analysis section number is j, the contribution rates C and C ′ can be expressed as follows.

C _j = (1 / ΣX _jm ) × (X _jm )
C _j '= (1 / Σ (X _jm ² ) × (X _jm ² )

  According to the frequency analysis, a contribution rate up to the m-th order is usually obtained for each analysis section. Then, the vowel phonemes and consonant phonemes that can be identified from the speech data group are identified by comparing the state of the determined contribution rate with a predetermined criterion. For example, when the phoneme of the consonant is “k” and the phoneme of the vowel is specified as “a”, the identification result is “K”. Further, when only “a” is specified from the audio data group, an identification result “a” is obtained.

  Next, another example of the speech identification method according to the present invention will be described with reference to FIG.

  The sampling of the audio signal and the A / D conversion are the same as in the example of FIG.

  In this example, the sound data group stored in the memory is divided into a sound data group in the vowel area and a sound data group in the consonant area prior to the window function processing. The division between the vowel area and the consonant area can be performed as follows, for example.

1. A predetermined number of audio data are sequentially compared from the beginning position of the signal area of the audio data group, and the value of the maximum peak (maximum peak P _max ) in the audio data and its position (the maximum peak P _max is the In the middle part).

2. An appropriate number of audio data sections are set, and the largest peak (interval peak P _n ) is sequentially obtained from the position of the maximum peak P _max toward the beginning of the audio data group.

3 Since there is no sudden peak drop in the vowel region, the maximum peak P _max , the section peak P ₁ in the section adjacent to this, and the section peak P ₂ in the next section adjacent to the section of P ₁ For example, when the section peak P ₁ is 60% or more of the maximum peak P _max , it can be determined that it is a vowel area, and the section peak P _n is the previous section peak P _{n−. When} it is 60% or more of ₁ , it can be determined that it is a continuation of the vowel region.

4 performs the comparison, determine the position to be lowered significantly in comparison with the interval peak P _n-1 immediately preceding the section peak P _n, which can be determined overall if the head position is a vowel area, which is audio data If it is the middle position of the group, it can be determined that it is the boundary between the consonant region and the vowel region. Even when the position at which the peak value has sharply decreased is not the head position, if the number of data from the head position to the position is extremely small, it can be determined that the region is a vowel rising region.

5. Further, when the same peak value is compared from the position of the maximum peak P _max toward the tail of the voice data group, the position of the tail of the vowel area can be detected. When a plurality of sounds are generated continuously, the boundary between the sounds can be detected by detecting this position.

  The above vowel region and consonant region segmentation method is one example, and any conventionally known method can be applied as the vowel region and consonant region segmentation method in the present invention. For example, it can also carry out by the method of patent document 1 quoted by background art. In addition, a plurality of sorting methods can be used in combination.

  After dividing the voice data group into a voice data group in the vowel area and a voice data group in the consonant area, each is subjected to window function processing similar to that described above as necessary, and frequency analysis is performed in the same manner as described above to determine the amplitude. A spectrum and / or power spectrum is determined.

  The frequency analysis is performed for each of the speech data group in the vowel region and the speech data group in the consonant region, and a contribution rate up to m-th order is obtained for each analysis section. Then, the vowel and the consonant are specified by comparing the obtained contribution rate state with a predetermined criterion. For example, when the consonant is “k” and the vowel is specified as “a”, the identification result is “K”. In addition, if only the vowel area is identified as “a” from the voice data group, the identification result is “a”. In particular, in the case of this example, the contribution rate obtained from the consonant speech data group is a comparison of only the criteria for identifying the consonant phoneme, and the contribution rate obtained from the vowel speech data group is the vowel phoneme. The comparison can be made only by the determination criterion for specifying the vowel, and the comparison can be simplified by dividing the vowel area and the consonant area in advance.

  The criterion for identifying phonemes of consonants and vowels is to determine the contribution rate of the fifty sounds from as many subjects as possible in advance, and organize how the contribution ratios of the respective phonemes of each subject appear. Can be prepared. Specifically, how many contribution ratios appear in what frequency domain, frequency domain that produces the maximum contribution ratio, contribution ratio of a specific frequency domain and other specific frequency domain It can be prepared by creating a database of the phoneme of the Japanese syllabary, such as the magnitude relationship with the rate.

  In the case where a result corresponding to a plurality of phonemes is obtained by comparison with the judgment criterion, for example, a priority order is set for the phonemes, and either is specified by specifying the order or referring to the original waveform. It is possible to select.

  The recognition accuracy can be further improved by introducing a neural network or the like when creating a criterion or identifying an unknown voice. In addition to the computer, the object can be achieved by using an appropriate electronic circuit.

  Next, an example in which the contribution rate is actually obtained will be described.

―About “A” ―
Phonemes were determined according to the procedure shown in FIG.

  First, ask the test subject to say “a” with a single sound, collect the sound with a microphone, sample it, A / D convert it, and add the data number in chronological order starting from No. 1. Stored. The sampling frequency was 50 kHz, and frequency components exceeding 25 kHz were cut with a low-pass filter when A / D conversion was performed.

  The collected voice waveform is shown in FIG.

  The audio signal area (area excluding the no-signal area) of the audio data group stored in the memory was taken out, subjected to window function processing using a Hamming window function, and subjected to fast Fourier transform. The number of fast Fourier transform data N is 1024, and the frequency analysis order m is 114. In this case, spectra from (1024/2) -1 = 511th order are obtained, but all the spectra from the 115th order to the 511th order are negligible values (close to 0).

  Tables 1 to 18 show the contribution ratios obtained as percentages.

  Table 1 shows contribution ratios obtained by performing fast Fourier transform on 1024 audio data with data numbers 314 to 1338 as one analysis section (one frame), and Table 2 shows 1024 with data numbers 714 to 1738. The contribution rate calculated | required by carrying out the fast Fourier transform as the audio | voice data 1 analysis area is shown. The data number in Table 1 is from 314, while the data number in Table 2 is 714 indicates that the analysis was performed while shifting each analysis section by 400 audio data. The same applies to the other tables hereinafter, in which 1024 audio data are defined as one frame and each frame is shifted by 400 audio data.

  The description in the “determination” column at the end of each table indicates the phoneme of the specified vowel or consonant, and the description in the “determination criteria” column indicates “phoneme” shown in tables 311 to 322 described later. Corresponding to the reference numerals shown in parentheses in the "" column. If the “determination” and “determination criteria” fields are blank, it indicates that the data was not used for the determination (data that did not correspond to the determination criteria described later). The same applies to other tables below.

―About “I” ―
The subject was asked to “a” with a single sound, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected speech waveforms are shown in FIG. 4, and the contribution rates obtained as a percentage are shown in Table 19 to Table 43. The data number in Table 1 starts from 314, whereas the data number in Table 19 starts from 21 because in Table 1, up to 313 was a no-signal state (silent state). It is excluded from the processing target, and in Table 19, it is up to 20. The same applies to the deviation of the data numbers in the other sound tables thereafter.

-About "U"-
The subject was asked to pronounce “U” with a single note, and the contribution rate was obtained in the same manner as the measurement of “A”.

  The collected speech waveforms are shown in FIG. 5, and the contribution ratios obtained as a percentage are shown in Table 44 to Table 68.

―About “E” ―
The subject was asked to pronounce “e” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected speech waveforms are shown in FIG. 6 and the contribution rates obtained as percentages are shown in Tables 69 to 93.

―About “O” ―
The subject was asked to pronounce “o” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected speech waveforms are shown in FIG. 7, and the contribution ratios obtained as percentages are shown in Tables 94 to 123.

―About “K” Line―
The subject was asked to pronounce “ka” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected voice waveforms are shown in FIG. 8, and the contribution ratios obtained as percentages are shown in Tables 124 to 144.

  Note that “ki”, “ku”, “ke”, and “ko” are omitted because the phoneme discrimination of the consonant is the same as “ka”.

―About the “sa” line―
The subject was asked to pronounce “sa” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected voice waveforms are shown in FIG. 9 and the contribution ratios obtained as percentages are shown in Tables 145 to 173.

  Note that “shi”, “su”, “se”, and “so” are omitted because the phoneme discrimination of the consonant is the same as “sa”.

―About “Ta” line―
The subject was asked to pronounce “ta” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected voice waveforms are shown in FIG. 10, and the contribution ratios obtained as percentages are shown in Tables 174 to 194.

  Note that “chi”, “tsu”, “te”, and “g” are omitted because the phoneme discrimination of the consonant is the same as “ta”.

―About “Na” Line―
The subject was asked to pronounce “na” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected speech waveforms are shown in FIG. 11, and the contribution ratios obtained as percentages are shown in Tables 195 to 223.

  Note that “ni”, “nu”, “ne”, and “no” are omitted because the phoneme discrimination of the consonant is the same as “na”.

―About “C” line―
The subject was asked to pronounce “ha” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected speech waveforms are shown in FIG. 12, and the contribution rates obtained as percentages are shown in Tables 224 to 250.

  Note that “hi”, “fu”, “he”, and “ho” are omitted because the phoneme discrimination of the consonant is the same as “ha”.

-About "Ma" line-
The subject was asked to pronounce “ma” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected speech waveforms are shown in FIG. 13 and the contribution rates obtained as percentages are shown in Tables 251 to 280.

  Note that “mi”, “mu”, “me”, and “mo” are omitted because the phoneme discrimination of the consonant is the same as “ma”.

―About “Ya” line―
“Ya”, “Yu”, and “Yo” are omitted because they are considered to be equivalent to “ia”, “iu”, and “io”.

―About “La” line―
The subject was asked to pronounce “ra” with a single note, and the contribution rate was obtained in the same manner as the measurement of “a”.

  The collected speech waveforms are shown in FIG. 15 and the contribution rates obtained as percentages are shown in Tables 281 to 310.

  Note that “li”, “le”, “le”, and “ro” are omitted because the phoneme discrimination of the consonant is the same as “la”.

―About “Wa” line―
“Wa” and “Wo” are omitted because they are considered to be equivalent to “ua” and “uo”.

―About “N” ―
“N” is omitted because it is considered to be equivalent to “un”, “n” or “m”.

―About judgment criteria―
Examples of determination criteria obtained as a result of measuring the Japanese syllabary from a plurality of male and female subjects are shown in Tables 311 to 322.

  In Tables 311 to 322, in order to simplify the display, the value obtained by adding the contribution ratios of the first harmonic (49 Hz) and the second harmonic (98 Hz) is defined as the contribution ratio of 98 Hz and the third harmonic ( 147 Hz) and the contribution ratio of the fourth harmonic (196 Hz) are added as the contribution ratio of 196 Hz, and in the same manner, the contribution ratio of the m-1st harmonic and the contribution ratio of the mth harmonic are added together. The value is expressed as a contribution rate at the m-th order frequency (where m is an integer of 2 or more). However, the criterion should not be based on the value obtained by adding the contribution ratio of the m-1st harmonic and the contribution ratio of the mth harmonic as the contribution ratio at the mth order frequency. It is also possible to use the one that directly represents the contribution rate from the first order to the m-th order in each analysis section.

  In Tables 311-322, the upper and lower numbers in the “frequency” item mean numbers that should not be multiplied by 98 Hz, the upper number indicates the tens place, and the lower number indicates the first place. Further, symbols such as A, B, C,... Shown in the “section” column mean areas indicated by arrows in the “frequency” column, but are attached for convenience of the following description. The symbols attached to do not mean the same frequency region.

  Hereinafter, Tables 311 to 322 will be supplementarily described.

(1) Determination criteria for “a” As shown in Table 311, “a” can be determined when either one of the two determination criteria A-1 and A-2 is satisfied.

A-1 is determined as “a” when all of the following conditions are satisfied.
-The spectrum in the section A (1 × 98 to 4 × 98 Hz) should not have a contribution ratio of 10 or more.
-The spectrum in the section B (5 × 98 to 9 × 98 Hz) should have less than 2 contribution factors with a magnitude of 10 or more.
-In the spectrum in the section C (8 × 98 to 15 × 98 Hz), there must be more than three with a contribution ratio of 3 or more.
-The spectrum in the section D (13 × 98 to 25 × 98 Hz) should not have 0 contribution factor of 3 or more.

A-2 is determined as “a” when all of the following conditions are satisfied.
-The spectrum in the section A (1 × 98 to 4 × 98 Hz) should not have a contribution ratio of 10 or more.
-In the spectrum in the section B (2 × 98 to 7 × 98 Hz), there should be more than one with a contribution ratio of 3 or more.
-The spectrum in the section C (5 × 98 to 9 × 98 Hz) should have less than 2 contributions with a magnitude of 10 or more.
-In the spectrum in the section D (9 × 98 to 15 × 98 Hz), there must be more than two having a contribution ratio of 10 or more.
-The spectrum in the section E (13 × 98 to 25 × 98 Hz) must not have 0 contribution factor of 3 or more.

(2) Determination criteria for “i” As shown in Table 312, it is possible to determine “i” when either of the two determination criteria I-1 and I-2 is satisfied.

  The way of reading the table of I-1 is based on the table 311 showing the determination criteria of the above “a”.

I-2 is determined as “i” when all of the following conditions are satisfied.
-In the spectrum in the section A (2 × 98 to 4 × 98 Hz), there should be no zero having a contribution ratio of 9 or more.
-The spectrum in the section B (11 × 98 to 15 × 98 Hz) has zero contribution factor of 2.5 or more.
-The spectrum in the section C (17 × 98 to 26 × 98 Hz) should have less than six with a contribution ratio of 2.5 or more.
-The spectrum in the section D (17 × 98 to 20 × 98 Hz) should have zero contribution factor of 1.5 or more.
-The spectrum in the section E1 (28 × 98 to 41 × 98 Hz) has eight or more contribution factors having a magnitude of 0.5 or more, or the section E2 (28 × 98 to 41 × 98 Hz). A certain spectrum has three or more contribution factors having a magnitude of 1 or more, or a spectrum in a section F (28 × 98 to 41 × 98 Hz) has a contribution rate of 0.5 or more. The spectrum in the section G (28 × 98 to 41 × 98 Hz) must have no more than one with a contribution ratio of 1 or more.
-The spectrum in the section H (35 × 98 to 46 × 98 Hz) has zero contribution factor of 2.5 or more.
・ In sections 1 × 98 to 10 × 98 Hz, those with a contribution ratio of 3 or more are 7 × 98
Does not exist above Hz.

(3) “u”, “e”, “o”, “s”, “t” Determination Criteria “u” is Table 313, “e” is Table 314, “o” is Table 315, “s” Table 317, “t” can be determined according to the determination criteria shown in Table 318. The reading of the T-1 table of “u”, “e”, “o”, “s”, and “t” is in accordance with Table 311 showing the determination criteria of “a”. The view of the T-1 table of “t” is in accordance with the view of K-2 described below.

(4) About the judgment criterion of “k” As shown in Table 316, it is judged as “k” when any one of the three judgment criteria of K-1, K-2, and K-3 is satisfied. Can do.

  The way of viewing the tables of K-1 and K-3 is in accordance with Table 311 showing the determination criteria of “a”.

K-2 is determined as “k” when all of the following conditions are satisfied.
-The spectrum in the section A (1 × 98 to 5 × 98 Hz) must have zero contribution ratio of 6 or more.
-The spectrum in the section B (16 × 98 to 20 × 98 Hz) has zero contribution factor of 2.5 or more.
-The spectrum in the section C1 (36 × 98 to 40 × 98 Hz) has one or more contribution magnitudes of 2 or more, or is in the section C2 (46 × 98 to 55 × 98 Hz). There must be at least one spectrum with a contribution ratio of 2 or more.
-The spectrum in the section D (41 × 98 to 45 × 98 Hz) must have at least one contribution factor of 3 or more.

(5) Determination criteria for “n” As shown in Table 319, “n” can be determined when all of the following conditions are satisfied.
-The spectrum in the section A (1 × 98 to 6 × 98 Hz) must have zero contribution factor of 30 or more.
-The spectrum in the section B (1 × 98 to 6 × 98 Hz) must have more than one contribution factor of 10 or more.
-The spectrum in the section C (1 × 98 to 6 × 98 Hz) should have more than 2 contributions with a contribution ratio of 5 or more.
The maximum contribution ratio of the spectrum in the section D (7 × 98 to 9 × 98 Hz) is p0, the maximum contribution ratio of the spectrum in the section E (10 × 98 to 15 × 98 Hz) is p1, and the section F (16 × The maximum contribution ratio of the spectrum in the range 98 to 21 × 98 Hz is p2, and the maximum contribution ratio of the spectrum in the section G (22 × 98 to 30 × 98 Hz) is p3. At least one is larger than p1, and the contribution of at least one of p0, p2, and p3 is 2 or more.
-The spectrum in the section H (31 × 98 to 55 × 98 Hz) should have zero contribution factor of 2 or more.

(6) Determination criteria for “h” As shown in Table 320, “h” can be determined when any one of the four determination criteria H-1 to H-4 is satisfied.

  The way of reading the table of H-2 is based on K-2 of the above table 320, and the way of reading the table of H-3 is based on the table 311 showing the judgment criteria of the above “a”.

H-1 is determined as “h” when all of the following conditions are satisfied.
-The spectrum in the section A1 (1 × 98 to 5 × 98 Hz) has zero contribution factor of 7 or more, or the spectrum in the section A2 (21 × 98 to 26 × 98 Hz) Is that there are no zero contributions of 3 or more.
-The spectrum in the section B (6 × 98 to 10 × 98 Hz) must not have zero contribution factor of 3 or more.
-The spectrum in the section C (11 × 98 to 15 × 98 Hz) must not have 0 contribution factor of 3 or more.
-The spectrum in the section D (16 × 98 to 20 × 98 Hz) must not have 0 contribution factor of 3 or more.
-The maximum contribution p0 of the spectrum exists in the section E (6 × 98 to 30 × 98 Hz), and the size of this p0 is 8 or more.

H-4 is determined as “h” when all of the following conditions are satisfied.
-The spectrum in the section A (1 × 98 to 5 × 98 Hz) has zero contribution ratio of 20 or more.
-The maximum contribution p0 of the spectrum exists in the section C (1 × 98 to 26 × 98 Hz), and the size of this p0 is 8 or more.
-There must be at least one of two or more of the sections B1 to B8 excluding the section to which the maximum contribution rate p0 belongs and having a contribution rate of 4 or more.

(7) Judgment criteria of “m” As shown in Table 321, it can be judged as “m” when one of the two judgment criteria of M-1 and M-2 is satisfied.

M-1 is determined to be “m” when all of the following conditions are satisfied.
-The spectrum in the section A (1 × 98 to 6 × 98 Hz) must have more than one contribution factor of 10 or more.
-The spectrum in the section B (1 × 98 to 6 × 98 Hz) must have more than 2 contribution factors with a magnitude of 5 or more.
The maximum contribution ratio of the spectrum in the section C (7 × 98 to 10 × 98 Hz) is p0, the maximum contribution ratio of the spectrum in the section D (11 × 98 to 15 × 98 Hz) is p1, and the section E (16 × P1 is p0, p2, and p3, where p2 is the maximum contribution ratio of the spectrum in the range 98 to 21 × 98 Hz) and p3 is the maximum contribution ratio of the spectrum in the section F (22 × 98 to 30 × 98 Hz). And p1 is 2 or more.
-The spectrum in the section G (31 × 98 to 55 × 98 Hz) has zero contribution factor of 4 or more.

  The way to read the table of M-2 is based on the table 311 showing the determination criteria of the “a”.

(8) Determination criteria for “r” “r” can be determined according to the determination criteria shown in Table 322. The way of reading this table conforms to M-1 in Table 321 above.

It is a block diagram which shows an example of the speech recognition method which concerns on this invention. It is a block diagram which shows the other example of the speech recognition method which concerns on this invention. It is a figure which shows the audio | voice waveform of "a". It is a figure which shows the audio | voice waveform of "I". It is a figure which shows the audio | voice waveform of "U". It is a figure which shows the audio | voice waveform of "d". It is a figure which shows the audio | voice waveform of "o". It is a figure which shows the audio | voice waveform of "K". It is a figure which shows the audio | voice waveform of "sa". It is a figure which shows the audio | voice waveform of "ta". It is a figure which shows the audio | voice waveform of "na". It is a figure which shows the audio | voice waveform of "c". It is a figure which shows the audio | voice waveform of "ma". It is a figure which shows the audio | voice waveform of "La".

Claims (6)

  A frequency analysis is performed on a group of audio data sampled from the audio signal and A / D converted, and the amplitude spectrum or power spectrum obtained thereby is used to calculate the amplitude or power of the fundamental wave and each harmonic component included in the audio frequency domain. A speech recognition method characterized in that the ratio of the amplitude or power of each of the fundamental wave and each harmonic component is obtained as a contribution rate, and phonemes of consonants and vowels are specified from the appearance of the contribution rate.   The speech data group is divided into a consonant region and a vowel region, and the contribution rate is obtained by frequency analysis of the speech data group of the consonant region and the speech data group of the vowel region, and from the appearance of the contribution rate in each speech data group, the consonant 2. The speech recognition method according to claim 1, wherein phonemes of vowels are specified.   The ratio of the amplitude of each of the fundamental wave and each harmonic component to the sum of the amplitudes of the fundamental wave and each harmonic component included in the audio frequency region is used as the contribution rate. Speech recognition method.   The voice according to any one of claims 1 to 3, wherein the voice data group is sequentially subjected to frequency analysis for each analysis section of N pieces of voice data, and a contribution rate is obtained for each analysis section. Recognition method.   5. The speech recognition method according to claim 1, wherein a frequency analysis is performed after a window function process is performed on the speech data group.   The speech recognition method according to claim 1, wherein a Hamming window is used for window function processing, and fast Fourier analysis is used for frequency analysis.

Images