JPH0462596B2 - - Google Patents

Description

[Detailed description of the invention]

[Technical Field] The present invention relates to a voice message identification method for operating electronic equipment using voice messages. [Background Art] Since phonological information depends on the articulation method, it is mainly included in the envelope of the frequency spectrum that corresponds to changes in cross-sectional area, and is especially contained in the envelope of the frequency spectrum corresponding to changes in cross-sectional area. ) and its bandwidth (50-110Hz). The frequency spectrum of speech is almost determined by the vocal tract transfer characteristics and the shape of the sound source waveform, but the vocal tract transfer characteristics include resonance points determined by the vocal tract cross-sectional area and resonance points determined by the vocal tract length. Articulation, or phonology, is determined mostly by the cross-sectional area of the vocal tract, and the length of the vocal tract varies between men, women, children, and individuals. Furthermore, the sound source waveform (particularly due to vocal fold vibration of voiced sounds) is thought to depend on the pitch and strength of the voice. A short interval analysis method is also known as a method for determining the spectral envelope and Holt mant. This is a linear predictive analysis performed within an interval that is slightly longer than the period of a voiced sound or as short as 3 msec (especially the glottal closure interval), and is not affected by the vocal fold frequency.
It is said that Holtmant can be calculated, but there is a problem in that it requires a large number of equal multiplications to calculate a correlation function for linear prediction and to calculate a covariance matrix for short interval determination. [Object of the Invention] The present invention has been made in view of the above-mentioned problems, and its purpose is to provide a voice message identification method that requires less calculation and is less affected by individual differences between speakers and vocal cord vibration. There is something to do. [Disclosure of the Invention] (Example) Fig. 1 is a circuit configuration diagram based on the processing flow of the present invention. In the figure, 1 is a high frequency emphasis section, and this high frequency emphasis section 1 enhances the high frequency range of input audio. This is for emphasis. Reference numeral 2 denotes an A/D converter that performs A/D conversion of the high-frequency emphasized input audio, and the output from this A/D converter 2 is input to a section compensator 3 and a pitch detector 4. The pitch detection unit 4 detects the absolute value IPOW of the amplitude of the time waveform of the A/D converted audio within one frame.
, and detects a large downward amplitude point IPN1 where the average value of a predetermined number of samples is the lowest in the first half of the frame, and also detects a second large downward amplitude point IPN that follows this point IPN1.
3, these downward large amplitude points IPN1 and IPN3
This is to find the pitch IPit from . The interval compensator 3 is for removing a DC component in a short interval, which will be described later, in consideration of the case where the amplitude of the time waveform is small. Reference numeral 5 denotes an analysis interval determination unit, which determines the large upward amplitude point IPN0 before the large downward amplitude point IPN1, and determines the large amplitude point IPN0 of both large amplitude points IPN1, IPN1,
The point of amplitude 0 in the middle of IPN0 is set as the large amplitude point IMiD,
The large amplitude point IPN points upward with this large amplitude point IMiD as the center.
A short section including one cycle of a half cycle of 0 and a half cycle of the large downward amplitude point IPN1 is determined. Here, the number of samples for fast Fourier transform is generally a power of 2 such as 64, 128, and 256, but due to window calculations, the length of the short interval is 32 points and 64 points. It was adopted. Reference numeral 7 denotes a fast Fourier transform section. This fast Fourier transform section 7 is used to obtain the envelope of the frequency spectrum by fast Fourier transform, and during calculation, the spectrum window determined by the analysis window calculation section 6 is multiplied. It will be done. The analysis window calculation unit 6 optimizes the length and position of the spectral window to be applied to the fast Fourier transform so that the fast Fourier transform unit 7 can extract the spectral envelope more accurately and with fewer calculations (number of multiplications). It is for the purpose of Reference numeral 8 denotes a frequency band division section, and this frequency band division section 8 extracts the short-term average power of each frequency component, such as UV, V, VH, from the output after converting the frequency spectrum extracted by the fast Fourier transform section 7 into a logarithmic power spectrum. , VL, VF, and VB. Here, V is inputting voice,
It shows the short-term average power in the frequency band of 0 to 1 KHz, and corresponds to the energy of voiced sounds.
UV indicates the short-term average power in the frequency band of 5 to 12 KHz during audio input, and corresponds to the energy of unvoiced sound. Also, voiced sounds VL, VH, VB,
VF is during audio input, 0~0.5KHz, 0.5~1.0K
Hz, 1.0~2.0KHz, and 2.0~4.0KHz frequency bands show the short-term average power, respectively.
It corresponds to the energy of wide jaw sounds, back tongue sounds, and front tongue sounds. Reference numeral 9 denotes a difference signal vector converter, and this difference signal vector converter 9 converts the five phonemes (i, e, a, o, u) into eao/
UV/V, VH/VL, VF/VB, VB/VL,
This is to find the difference signal vector of VF/VH.
Reference numeral 18 is a formant locus converter for obtaining a formant vector, whereas the frequency band dividing section 8 and the difference signal vector converting section 9 are for obtaining a difference signal vector by frequency band division. The peak frequency (formant frequency) of the spectrum envelope is determined and used as a formant vector, and the components of the formant vector are the differences with respect to the average value for each formant, and the frequency axis is expressed on a logarithmic or linear scale. The recognition rate can be improved by switching the above-mentioned average value, which serves as the reference frequency for each formant, to classes such as male, female, and child through pitch detection. Figures 9a and 9b show the formant distribution of five vowels and the positions of the peaks. Reference numeral 10 denotes a symbol vector converter, which converts the symbol vector {i, e, a, o, u,
h, l, f, b, w}, and the value of the transformation matrix only needs to have row elements corresponding to the magnitude of each component of the difference signal vector or formant vector corresponding to the symbol. Reference numeral 11 denotes a start/end detection section, and this start/end detection section 11 determines UV when the UV/V difference signal is more positive than a certain set value Ru, and determines V when it is more negative than a certain set value Rv. Equipped with a voiced/unvoiced determination function that determines the middle as S, and detects the beginning of the voice by determining UV and V.
If silence continues for a number of samples equal to or greater than a certain set value, it is detected as the end. Reference numeral 12 denotes a symbol conversion processing unit, which outputs the symbol if the maximum component of the symbol vector in the interval V is greater than or equal to a certain set value, and if it is less than or equal to the set value, it outputs the symbol.
Output. In addition, in the UV and S sections, respectively
Outputs UV and S. Reference numeral 13 denotes a formatting processing section, which converts repetitions of the same symbol into a list of one symbol and its duration, and omits symbols with short durations. 14 is a word standard pattern storage unit, and this word standard pattern storage unit 14
is for registering a voice pattern in registration mode and using it as a standard pattern for recognition and verification. The preliminary selection unit 15 is used in the recognition mode to perform preliminary selection to limit the objects of comparison by performing primary identification based on the number of UVs or the like before comparison. 16
is the time axis normalization/verification section, and this time axis normalization/
The matching unit 16 has a time axis normalization function for normalizing the duration so that the total duration time in the above list becomes a constant value such as 200 (or 1000), and a corresponding one on the time axis. It consists of a distance calculation function that calculates the distance (correlation value) between symbols and matches it with a standard pattern using the distance table shown in Table 1, in which the sum of all samples is the distance between patterns.

[Table] In Table 1, the horizontal and vertical columns correspond to the standard pattern symbol and input pattern symbol, respectively.For example, if the standard pattern symbol is a, and the input pattern symbol is also a When , the output of the distance table is -2, indicating that the degree of approximation is low. Therefore, in the distance calculation function, the degree of approximation of the input pattern and the standard pattern as a whole can be easily calculated by simply adding the outputs from the distance table in sequence. Reference numeral 17 denotes a significant difference testing unit, which detects the closest pattern when it is closer than a certain setting value and is further away from the second closest pattern by more than a certain setting value. It is equipped with a significant difference test function that considers the input patterns to be the same and rejects them as recognition failure in other cases, and a result output function that outputs the recognition result. 19 is an optimization feedback section, and this optimization feedback section 1
9 is the speaker's /i, e,
The time series of the occurrences of a, o, and u/ is stored, and the dividing frequency is varied in the vicinity of a preset standard dividing frequency, and the direction and amount of variation are determined according to the sensitivity characteristics of the symbol vector. In this case, the slope of the spectrum is compensated by the offset of the difference signal vector, and in particular, when the input voice is an A note, the i component is prominent, and when the input voice is an A note, the a note is prominent. /e/, /u/
The gain balance of the difference signal input is adjusted so that identification of the difference signal becomes more reliable. In this case, first VH/VL
Then, the optimum adjustment of VF/VB is performed, and then the optimum adjustment of VB/VL is performed. In the example, the sampling period is 80 μsec (sampling frequency 12.5 KHz), and the frame length is
There were 512 samples. The period of the fundamental frequency is the lowest
If it is 90Hz, there will be 139 samples, and in order to calculate the frequency spectrum of 256 points, normal fast Fourier transform will require calculation of 512 points, and the number of multiplications will be ^29.
× (2 ⁴ + 2 ⁵ ) = 512 × (16 + 32) = 24576 times,
When 64 samples in an interval shorter than the fundamental period are extracted from a frame of 512 samples and analyzed, 128
Fast Fourier transform of a point is sufficient, so 2 ⁷ × (2 ³ + 2 ⁴ ) =
128×(8+16)=3072 multiplications are enough. Furthermore, since the multiplication of the analysis window in the preprocessing of the fast Fourier transform is the same as the number of samples of the frequency spectrum, it can be seen that the short interval analysis is effective as a simple method. FIG. 2 shows a flowchart of the characteristic part consisting of the pitch detection section 4 and the analysis interval determination section 5 of the embodiment of FIG.
Calculate IPOW in (1), and in (2) find the downward large amplitude point where the average value of each 30 samples is the lowest in the first half of the frame.
Detect IPN1, and then proceed to the next downward large amplitude point in (3)
Detect IPN3, and in (4), from these downward large amplitude points IPN1 and IPN3, pitch IPit=IPN3−
Find IPN1. After detecting the pitch (5), the previous upward large amplitude point IPN0 is detected from the downward large amplitude point IPN1, and from the point with amplitude 0 between the two large amplitude points IPN0 and IPN1, the large amplitude point IMiD is determined in (6). Large amplitude point
A short section including one cycle consisting of a half cycle of large upward amplitude and a half cycle of large downward amplitude is determined with IMiD as the center. Next, perform DC component compensation in (7), perform analytical windowing in (8), perform fast Fourier transform in (9), select mode between difference signal vector or formant vector in (10), and (11) ) to find the frequency band division, and (12) to find the formant locus. FIG. 3 shows a specific circuit diagram of the present invention, in which audio is input from a microphone 18, amplified by a preamplifier 19, and adjusted for gain and offset by an adjustment amplifier 20. Next, the A/D conversion circuit 21 digitally converts the audio input, and the digitally converted audio frame is stored in the audio frame memory 23. Reference numeral 24 denotes an FFT processor, and this FFT processor 24 is a general type having a control section 24a, an arithmetic register 24b, a built-in RAM 24c, and a coefficient ROM 24d in which coefficients are stored.
Consists of FET chip, audio frame memory 23
The audio frame read from is taken in and fast Fourier transform is applied to the window. 25 is a spectrum frame memory, and FET processor 2
This is for storing the spectrum frame calculated in step 4. 22 is audio frame memory 2
3. A timing circuit that provides operating timing for the FFT processor 24 and spectrum frame memory 25. 26 is a CPU that performs control calculations based on the operation program written in advance in the program ROM 27, and in the verification mode operates the verification calculation circuit 30 to encode data stored in the spectral frame memory 25 and register it in advance. In the mode, it is possible to perform a comparison calculation with the standard pattern stored in the standard pattern RAM 31, or in the registration mode, the input voice pattern can be stored as a standard pattern in the standard pattern RAM 31, and furthermore, in the learning mode, the above-mentioned optimization can be performed. We send feedback. In the figure, 32 is the terminal section, 33 is the microcomputer bus, and 28 is the working section.
RAM 29 is a control input/output section. Figure 4a shows a frame length of 128 points, and Figure 4b shows a frame length of 128 points.
Fig. 5 shows an analysis example of the fundamental period corresponding to the large amplitude position of the time waveform shown in Fig. 5, and Fig. 5 shows an analysis example using this method, and Fig. 5 a shows the time waveform of the /I/ sound shown in Fig. 5 b. The frame length is 64 points (fast Fourier transform is
This is the simulation result of fast Fourier transform multiplied by a window of 128 points). Fig. 6 shows an example of analysis using this method, and Fig. 6 a shows the time waveform of the /I/ sound shown in Fig. This is a simulation result using Fourier transform. FIG. 7 shows the symbolization process. In the figure, V indicates the short-time average power in the frequency band of 0 to 1 KHz during voice input, and corresponds to the energy of voiced sound. Also, UV is inputting audio,
It shows the short-term average power in the frequency band of 5 to 12 KHz, which corresponds to the energy of unvoiced sounds. Furthermore, VL, VH, VB, and VF are respectively input during voice input.
It shows the short-term average power in the frequency bands of 0~0.4KHz, 0.4~0.8KHz, and 1.8~3.2KHz, which corresponds to the energy of narrow jaw sounds, wide jaw sounds, back tongue sounds, and front tongue sounds, respectively. ing. S ₀ to _{S 4} are differential amplification means,
Difference signals V/UV, Veao/Viu, Va/Veo, respectively
This is to calculate Ve/Vo and Vi/Vu. _C0 is a comparison means, and when the difference signal component output from the differential amplification means _S0 is smaller than the reference value Rv, it is assigned the sign of a voiced sound V, and when it is larger than the reference value Ru, it is assigned the sign of an unvoiced sound V. In other cases, it is determined to be silent S. However, Ru>O>Rv. MY ₀ is a symbolization processing unit, and this symbolization processing unit MY ₀ inputs one of the SVUV codes for each case of silence, voiced sound, and unvoiced sound. MC ₀ is each difference signal output Vea/Viu, Va/Vea, Ve/Vo, Vi/Vu
Each vowel i, e,
a, o, u, and other voiced sounds h, i, f, b,
The short-term average power of w is calculated, and the output of the matrix calculation unit MC ₀ is input to the maximum value determination unit MX ₀ to calculate each component i, e, a, o, u, h, l, f,
It is determined which component is the largest among b and w, and the sign of the largest component is sent to the encoding processing unit MY ₀
is input. However, when the difference between the largest component and the second largest component is small, the code m is output. When the code output from the comparison means _C0 is V, the symbolization processing unit _MY0 operates as a maximum value determination unit.
i, e, a, o, u, h, output from MX ₀
It outputs any one of the codes l, f, b, w, and m, and when the code output from the comparing means _C0 is U or S, it outputs that code as is. It should be noted that as the transformation matrix Tm of the matrix calculation unit _MC0 , equations (1) to (3) can be used. [Tm] = -17, 17, 17, 17, -17, 18, -18, 0, 0, 13, 0, 0, 17, 0, 0, 0, 0, 0, 0, 0, 0, 17 , 0, -17, 0, 0, 0, 18, -18, 0,17 0 0 0 0 0 0 0 0 -13 ...(1) [Tm] = -16, 16, 16, 16, -16, 18, -18, 0, 0, 13, -8, -8, 16, -8, -8, 0, 0, 0, 0, 0, 0, 16, 0, -16, 0, 0, 0, 18, -18, 0,16 0 -8 0 -16 0 0 0 0 -13 ...(2) [Tm] = -14, 14, 14, 14, -14, 18, -18, 0, 0, 13 −14, −14, 14, −14, −14, 0, 0, 0, 0, 0,0, 14, 0, −14, 0, 0, 0, 18, −18, 0,14 0 −14 0 −14 0 0 0 0 −13 ...(3) First, the transformation matrix Tm in equation (1) is set to 0 except for the minimum necessary elements for identification to speed up the calculation. ) formula has redundancy in the part where the absolute value of the element is 8, so that if the detection of the difference signal is weak, it is possible to symbolize a wide range of five vowels. The elements of the 5 vowels for the difference signal for the 1st formant F ₁ are all given the same weight (absolute value 14), and for the 2 difference signals for the 2nd formant F ₂ , for the 5 vowels, either The first formant F 1 is given the necessary weight for identification one by one, and the first formant F ₁ is changed to the second formant F1.
It can be said that it was given more importance than _F2 . This transformation matrix
Tm can be arbitrarily set depending on the word to be identified. AP in FIG. 3 shows the characteristics of the adjustment amplifier 20 described above. In addition to the above-mentioned verification method, it is also possible to create a binary signal from the difference signal, encode it in combination, and successively perform verification. Examples of this method include: In other words, UV/V difference signal, Veao/Viu difference signal, Va/Veo difference signal, Ve/Vo difference signal, Vi/Vu obtained from the short-time average power vector.
Extract the difference signal and assign the symbol Veao if the Veao/Viu difference signal is above a certain positive value, assign the symbol Viu if it is below a certain negative value, and assign the symbol S in other cases. When the difference signal is above a certain positive value, the symbol Va is assigned, when it is below a certain negative value, the symbol Veo is assigned, and in other cases, the symbol S is assigned,
When the Ve/Vo difference signal is above a certain positive value, the symbol Ve is assigned, when it is below a certain negative value, the symbol Vo is assigned, and in other cases, the symbol S is assigned, and the Vi/Vu difference signal is positive. When it is above a certain value, the symbol Vi is assigned, when it is below a certain negative value, the symbol Vu is assigned, and in other cases, the symbol S is assigned. Then, while storing these symbols in the temporary storage means and referring to the symbolization table shown in Table 2, the symbols a, e, o,
Convert to one of the symbols i, u, h, l, f, b, w, and m.

【table】

[Table] However, in Table 2, * indicates that it can be either 0 or 1, and 0/1 indicates 0 or 1.
The case is shown below. Such a symbolization table is constructed using, for example, a ROM, and by accessing the ROM by using the temporarily stored contents as an address input, codes for each symbol such as a, e, o, etc. can be obtained as data output. Alternatively, it is preferable to compare the temporarily stored contents and the contents of the symbolization table using exclusive OR, and output a symbol when they match. FIG. 9 is a flowchart showing a method for realizing the matching section by a sequential discrimination processing program of a microcomputer,
First, as a first step, the Veo/Viu difference signal is divided into three groups depending on whether it is a high level H, a medium level M, or a low level L. And in the second stage, when the first stage is H,
If the Va/Veo difference signal is H, it outputs the symbol /a/, if it is M, it outputs the symbol /Vo/, and if it is L, it moves to the third step, checks the Ve/Vo difference signal, and if it is H, it outputs the symbol /Vo/. It outputs the symbol /e/, if it is M it outputs /h/, and if it is L it outputs the symbol /o/. On the other hand, the first
If the stage is M, in the second stage, if the Ve/Vo difference signal is H, the symbol /f/ is output, and if it is M, the symbol /f/ is output.
It outputs m/, and if it is L, it outputs the symbol /b/.
Furthermore, if the first stage is L, the second stage is Vi/Vu
If the difference signal is AH, the symbol /i/ is output, if it is M, the symbol /l/ is output, and if it is L, the symbol /u/ is output. Although the above-described embodiment uses fast Fourier transform as the orthogonal transform, Walsh transform may also be used. [Effects of the Invention] The present invention is configured as described above, and maintains the positive and negative phases of the time waveform of the voice input so that the positive and negative phases are constant with respect to the direction of the vocal cord vibration, and the time waveform of the voice that is the differential waveform of the vocal cord vibration. The large downward amplitude point is defined as the upward large amplitude point, and the downward large amplitude point next to the upward large amplitude point corresponds to the differential waveform of vocal fold vibration due to the fall of vocal fold vibration. Means for detecting a point with a maximum downward value as a point with a size larger than the frame average value, and detecting the intermediate point between the large downward amplitude point and the large upward amplitude point before this large amplitude downward point as a large amplitude point, and detecting the large amplitude point as a large amplitude point. means for determining a short interval analysis interval centered on
The DC component of this determined short-term analysis section is removed, an analysis window is applied, and the envelope of the frequency spectrum of the audio input is extracted by short-term analysis. Because short-term analysis is performed, macroscopic features can be easily detected.In particular, even if there is an influence of the sound source waveform, there is no influence of the fundamental period, so the spectral envelope can be easily determined, and individual differences between speakers can be detected. This method has the advantage of being less affected by vocal fold vibration and requiring less calculations, allowing for high-quality recognition.

[Brief explanation of the drawing]

FIG. 1 is a schematic circuit configuration diagram of an embodiment of the present invention, FIG. 2 is a flowchart for explaining the operation of the main parts of the same, FIG. 3 is a specific circuit diagram of the same, and FIGS.
7 is a circuit block diagram for explaining the symbolization process of the embodiment of the present invention; FIG. 8 is a flowchart for explaining the operation of another collation example of the present invention; FIG. The figure is a waveform diagram for explaining the Holtmann locus, where 3 is an interval compensation section, 4 is a pitch detection section, 5 is an analysis section determination section, 6 is an analysis window calculation section, 7 is a fast Fourier transform section, and 8 is a frequency band. 10 is a symbol vector conversion section; 12 is a symbolization processing section; 16 is a time axis normalization/verification section; and 18 is a Holtmal locus conversion section.

Claims (1)

[Claims] 1. A means for extracting the frequency spectrum of a voice input, and dividing the logarithmic power spectrum into frequencies to extract short-term average powers for each frequency band and extracting five vowels i, e, a from these short-term average powers. ,o,u
Extract the difference signal vector so that it is divided into the ratio of e, a, o and i, u, the ratio of a and e, o, the ratio of e and o, and the ratio of i and u, or extract the formant vector from the formant locus. matrix calculation means for calculating symbol vectors of five vowels and other voiced sounds by multiplying the difference signal vector or formant vector by a transformation matrix; A means for outputting the largest component as an onomatopoeic symbol of an analysis frame and a standard pattern in which input patterns consisting of symbols and durations are stored in advance based on the onomatopoeic symbol are compared on a time axis or by symbol. In a voice message identification method that identifies the standard pattern closest to the input pattern as the input message, the positive and negative phases of the time waveform of the voice input are maintained so that they are constant with respect to the direction of vocal cord vibration. , the temporal waveform of the voice corresponding to the differential waveform due to the rise of vocal fold vibration is set as the upward large amplitude point, and the downward water amplitude point next to the upward large amplitude point is set as the falling point of vocal fold vibration. means for detecting a downward large amplitude point for a voiced sound as a point with a maximum downward value of a short-time average value of a voice sample and a point having a magnitude larger than a frame average value in correspondence with a differential waveform of vocal fold vibration; means for detecting the middle point between the large amplitude point and the previous upward large amplitude point as the large amplitude point, and determining a short period analysis section centered on this large amplitude point; A voice message identification method characterized in that the envelope of the frequency spectrum of a voice input is extracted by short-term analysis by applying an analysis window and removing the frequency spectrum of voice input. 2. The voice message identification system according to claim 1, wherein the width of the short section includes one cycle at the time of large amplitude of the voiced sound.