JPH02717B2 - - 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 technology] Conventionally, a spectral pattern obtained by frequency analysis of speech extracted using a bandpass filter is converted using a conversion matrix, etc., a characteristic pattern of the speech is extracted using symbols or parameters, and speech is identified using this pattern. In order to simplify and reduce costs, we investigated a method for identifying voice messages by forming micro-feature patterns corresponding to the differences between voiced and unvoiced sounds and formant trajectories. Figure 1 shows the earlier application related to the present invention (Japanese Patent Application No. 1983-
67261) is a block diagram showing the configuration of the device. The voice message identification device shown in the figure can be roughly divided into an acoustic analysis section A, a pattern conversion section B, and a word identification section C. The acoustic analysis section A converts the output of the preconditioning amplifier 2, which amplifies the sound detected by the microphone 1, emphasizes the high frequency component (equivalent circuit), and removes the DC component, into frequency spectrum components using the frequency analysis section 3. The frequency spectrum pattern is expanded and converted into a logarithmic scale by the logarithmic conversion unit 4. FIG. 2 shows an embodiment of a conventional acoustic analysis section A. In Figure 2, audio frequency analysis is performed using six filter banks (V, UV, VL,
VH, VB, VF). Here, to explain the frequency band of each filter bank, V is the band where the power of voiced sound V is concentrated (approximately 0 to
1KHz), UV is the band where the power of silent UV is concentrated (approximately 5 to 10KHz), and VL is the band where the first formant of constrictor sound VL is concentrated (approximately 0 to 0.4K
Hz), VH is the band where the first formant of the broad jaw sound VH is concentrated (approximately 0.4 to 0.8 KHz), and VB is the band where the second formant of the back tongue sound VB is concentrated (approximately 0.8 to 0.8 KHz).
1.8KHz), and VF extracts the band (approximately 1.8 to 3.2KHz) where the second formant of the frontal sound VF is concentrated. Figure 3 a and b show the frequency characteristics of each filter bank VL, VH, VB, and VF. Figure a shows the frequency on the horizontal axis as a uniform scale, and Figure b shows the frequency on the horizontal axis. is plotted on a logarithmic scale. Note that in FIG. 3, AP indicates the characteristics of the adjustment amplifier 2b, which will be described later. As shown in the block diagram of FIG. 2, the preconditioning amplifier 2 includes a preamplifier 2a that amplifies the audio signal detected by the microphone 1, and a preamplifier 2a that amplifies the audio signal detected by the microphone 1.
an adjustment amplifier 2b that is connected to the output of the output and adjusts the gain and offset value;
It has c. The level adjuster 2c balances the power of the signals supplied to the filter banks V and UV with the power of the signals supplied to the other filter banks. Next, the V/UV balance adjuster 3a balances the input of the filter bank V and the input of the filter bank U. For other filter banks, adjust the VB/VL balance adjuster 3b to the midpoint, and then adjust the filter bank VH and filter bank using the VH/VL balance adjuster 3c.
Balance the VL input, and use the VF/VB balance adjuster 3d to balance the filter bank VF and filter bank.
Balance VB. Further, the VB/VL balance adjuster 3b balances the filter bank VB and the filter bank VH or VL. The output of each filter bank is multiplexer 3e
The signals are sequentially switched in a time-division manner and input to the logarithmic conversion section 4. The logarithmic conversion unit 4 converts the input power into a logarithmic scale. The output of the logarithmic converter 4 is input to an A/D converter 4a and digitized into an 8-bit binary number. Note that when each filter is configured with a digital filter, the A/D
Converter 4a is the next stage after adjustment amplifier 2b. The output of the A/D converter 4a is input to the difference signal vector converter 5 of the pattern converter B, and a difference signal vector consisting of the difference signals of the outputs of each filter bank is extracted. FIG. 4 shows an example of the configuration of the difference signal vector converter 5. In FIG. As shown in figure a, the difference signal vector is obtained by averaging (integrating) the difference between the two filter outputs. In other words, when the filter bank output is digitized by the A/D converter 4a at a sampling period of 10 msec, the difference signal of the previous sample multiplied by the coefficient α by the coefficient unit 5a and the difference signal of the current sample. The value stored in the sum register 5b becomes the difference signal vector component of the current sample. Coefficient α
is approximately 0.6 to 0.8. The difference signal vector converter 5 is simplified as shown in FIG. 4b, and FIG. 5 shows an embodiment of the pattern converter B in the prior application. As shown in FIG.
Then, UV/V difference signal, V _eap /V _iu difference signal, V _a /V _ep
A total of five component difference signal vectors are extracted: the difference signal, the V _e /V _p difference signal, and the V _i /V _u difference signal. First, the UV/V difference signal is used to distinguish between voice presence and no voice, and is a difference signal between the outputs of filter bank UV and filter bank V. next,
The V _eap /V _iu difference signal is used to distinguish between vowels e, a, o and i, u, and is a difference signal between the outputs of filter bank VH and filter bank VL. Further, the V _a /V _ep difference signal is used to identify vowels a, e, and o, and is a difference signal between the outputs of filter bank VB and filter bank VL. The V _e /V _p difference signal is used to distinguish between vowels e and o, and is a difference signal between the outputs of filter bank VF and filter bank VB. Furthermore, the V _i /V _u difference signal is used to identify the vowels i and u, and is a difference signal between the outputs of filter bank VF and filter bank VH. Among the difference signals, the V _eap /V _iu difference signal and the V _a /V _ep difference signal extract the characteristics of the first formant of the vowel.
Further, the V _e /V _p difference signal and the V _i /V _u difference signal extract the characteristics of the second formant of the vowel. The UV/V difference signal created as described above is input to the V, UV, S determining section 18, and is divided into a section of voiced sound V, a section of unvoiced sound UV, and silent S.
It is used to determine the interval. The V, UV, S determination unit 18 converts the UV/V difference signal into predetermined reference values Rv, Ru.
(Rv<O<Ru), and if the UV/V difference signal is smaller than the reference value Rv, it is determined to be a voiced sound V,
If it is larger than the reference value Ru, it is determined to be an unvoiced sound U, and if it is between the reference values Ru and Rv, it is determined to be a silent sound S. The start end end detection unit 6 detects that the audio input is a voiced sound V.
Alternatively, if it is determined that the voice input is an unvoiced sound U, it is determined that it is a voice section, and if the voice input is determined to be a silence S, it is determined that it is a silent section. Next, the difference signal vectors of the outputs of the filter banks VF, VB, VH, and VL are input to the symbol vector converter 7. The symbol vector converter converts each difference signal
A four-dimensional difference signal vector whose components are V _eap /V _iu , V _a /Veo, V _e /V _p and V _i /V _u is multiplied by the transformation matrix [Tm] to be included in the audio input. Short-term average power V _i of each vowel i, e, a, o, u,
V _e , V _a , V _p , V _u , and each short-term average power V of wide-mouthed sounds, narrow-mouthed sounds, front-mouthed sounds, back-mouthed sounds, and voiced sounds intermediate between vowels a and o. _h , _Vl , Vf,
This is to calculate V _b and V _w . An example of the components of the transformation matrix [Tm] in the symbol vector transformation unit 7 is as shown in the following equation. The output of the symbol vector converter 7 is input to the maximum value determination unit 8a, and each component V _i , V _e , V _a ,
It is determined which of V _p , V _u , V _h , V _l , V _f , V _b , and V _w has the largest component. Symbol output section 8b
When the V, UV, and S determining section 18 determines that the audio input is silent or unvoiced, it outputs the symbols S and UV, respectively. Also, V, UV,
When the S determining section 18 determines that the audio input is a voiced sound, the symbol output section 8b outputs the symbol of the voiced sound determined to be the largest component in the maximum value determining section 8a. However, each voiced sound V _i , V _e , V _a ,
If the maximum component among V _p , V _u , V _f , V _b , V _h , V _l , and V _w does not reach the predetermined reference value,
Voiced sound V _n that does not correspond to any of the above voiced sounds
Outputs the symbol. Therefore, the symbol output section 8b
From, S, UV, V _i , V _e , V _a , V _p , V _u , V _h ,
Any one of the 13 types of symbols V _l , V _f , V _b , V _w , and V _n will be output. The time series of each symbol output from the symbol output section 8b is input to the formatting section 9a, where it is formatted. That is, the formatting processing unit 9a converts the repetition of the same symbol into a list of one symbol and its duration, and furthermore, if the duration is less than a certain set value, if the symbols before and after are the same, then these are If the symbols before and after are different in one list,
Include them in the previous symbol, and omit short duration ones. The output of the shaping processing section 9a is input to the time axis linear normalization processing section 9b. The time axis linear normalization processing unit 9b normalizes the duration so that the total duration of each list becomes a constant value such as 200 (or 1000). This can be done by multiplying each duration by the ratio of the total sample value 200 (or 1000) to the duration. Through the above process, a voice pattern for the input voice message can be created. This voice pattern is registered in the standard pattern storage section 10 in the registration mode. In the recognition mode, the correlation calculation unit 12 matches the standard pattern, but first the preliminary selection unit 11 performs primary identification to limit the matching target. The preliminary selection section 11 selects unvoiced sounds.
The objects to be matched are limited by the number of UVs, the number of voiced sounds V, the number of pauses P indicating a rather long silent section (for example, before a break), the number of symbols, the total duration, etc. The correlation calculation unit 12 uses the correlation table 13 shown in Table 1 to determine the correspondence between the input pattern and the standard pattern using the DP matching method so that the correlation is maximized (the distance is the shortest). dynamically collate the data.

[Purpose of the invention]

The present invention has been made in view of the above-mentioned points, and it is possible to reduce the difference in speech patterns for the same word depending on how it is uttered, and to more reliably perform symbolization even when the power of the speech is low. It is an object of the present invention to provide a voice message identification method which can reliably encode vowels following consonants and unvoiced sounds. [Disclosure of the Invention] FIG. 8 is a block diagram of an embodiment of the present invention. In the figure, an acoustic analysis section A and a word identification section B
The configuration is the same as that of the conventional example shown in FIG. 1, except that the configuration of the pattern conversion section C is different. First, in FIG. 8, reference numeral 16 denotes a differential vector converter, which calculates a differential signal of the difference signal component output from the difference signal vector converter 5. A difference signal emphasizing section 17 adds the differential signal calculated by the differential vector converting section 5 to the original difference signal to emphasize the rising and falling edges of the difference signal. Reference numeral 18 denotes an offset calculation section, which calculates an offset value according to the positive/negative sign of the differential signal outputted from the differential vector conversion section 16 and its absolute value. Furthermore, 19 is an offset compensator which adds the offset value calculated by the offset calculator 18 to the original difference signal. The offset compensating section 19 and the offset calculating section 18 create an offset processing section D.
It consists of The circuit diagram in FIG. 9 shows the configuration of a specific circuit that realizes the functions of the difference signal vector converter 5, the differential vector converter 16, and the difference signal enhancer 17. To explain the operation of the circuit in the same figure, first, the current difference signal is stored in register t _o , the previous difference signal is stored in register t _o-1 ,
The differential average of this register t _o and register t _o-1 (the short-term average power of the variation of the difference) is formed as a differential vector component, and this differential vector component and the register
This is a method in which the short-term average power of the sum with t _o-1 or the contents of register t _o is used as the emphasized difference signal vector component. The output of the differential vector converter 16 is input to an offset calculator 18, where an offset value is calculated. This offset value is added to the original difference signal in the offset compensator 19 as described above, but equivalently, the matrix calculation formula in the symbol vector converter 7 is changed as shown below. Therefore, offset compensation can be performed. However, in the following equation, Of ₁ , Of ₂ , Of ₃ , and Of ₄ indicate offset values. Figures 10 and 11 show the differential waveforms of the difference signal waveforms in Figures 6 and 7, respectively, and these figures show the differential waveform itself before the short-term average power is integrally calculated. ing. 1st
2 and 13 show the waveform of the emphasized difference signal vector which is the output of the difference signal emphasizing section 17, and the encoding pattern. It shows. Comparing the waveform diagrams in Figures 6 and 7 with the waveform diagrams in Figures 12 and 13, the rising and falling edges of the waveforms of the emphasized difference signal vectors in Figures 12 and 13 are emphasized. It can be seen that the waveform is overshot, and that the unevenness of the waveform is becoming clearer. By the way, Japanese syllables are consonants.
There are many CV syllables (consonants + vowels) consisting of the combination of a word and a vowel, and they are recognized as syllables when the mouth goes from a closed state to an open state. It seems effective to emphasize and weight the differential vector component. Figures 14 and 15 and Figures 16 and 17 show the waveforms of the emphasized difference signals and their symbolization patterns when two different male subjects utter the voice message "Start motion". It shows. In this measurement example, the time constant for short-time averaging of the emphasized difference signal vector is set longer than in the case of FIG. 13. Figures 15 and 17
In the figure, V1, V2, V3, V4, etc. indicate voiced sound sections. For example, in the V1 section, V _l ,
This shows that the symbolization patterns of each of the voiced voices V _n , V _p , V _a , V _p , and V _n have been obtained. Further, the symbol F indicates that a fricative sound was obtained from the unvoiced sound UV. In this measurement example, the "s" in "dousa" and the "s" in "Kaisi" are each detected as voiceless fricatives F.
Furthermore, the symbol P means a pause before the voiceless plosive "K". If we look at the symbolization patterns in Figures 15 and 17, we can see that “dousa
It can be seen that the vowel sequence o, (u), a, a, i, i of "Kaisi" is detected regardless of the speaker. Figures 18 to 21 show two different women and two different men. Regarding the subject of the name, Japanese 5
The waveform of the emphasis difference signal and its symbolization pattern are shown when the vowel "ye-a-oh" is uttered. Looking at each of the above figures, it can be seen that the symbolization of the five vowels ``ye-a-ou'' is done uniformly regardless of the speaker. By emphasizing the difference signal using the differential waveform component as described above, it is possible to emphasize the rising and falling edges of the difference signal, and the vowel that follows the consonant can be well symbolized regardless of the speaker. It is something that can be done. By the way, the offset processing section D in FIG.
The offset compensator 18 performs dynamic offset compensation, and when the differential vector component is positive, the offset compensator 19 converts the positive offset into an emphasized difference signal vector (or
(difference signal vector) to form an offset-compensated difference signal vector, so to speak, so that the zero point of the difference signal moves to the negative side by the offset amount, and when the differential vector component is negative, By adding a negative offset to the emphasized difference signal vector, the original state without offset is restored. However, when detecting a change from a consonant to a vowel, that is, from a narrow jaw opening to a wide jaw opening, as with the CV syllable mentioned above, the difference signal vector (V _eap It can be said that it is effective to emphasize the rise of V _iu , V _a /V _ep ). On the other hand, the difference signal vectors (V _e /V _p , V _i /
In the case of V _u ), in the above-mentioned CV syllable, we want to emphasize the change from the front tongue state to the back tongue state, so if the differential vector component of the difference signal vector is negative, a negative offset is applied to the offset compensator 19. In addition to the emphasized difference signal vector (or difference signal vector), if the differential vector component is positive, a positive offset is added to the emphasized difference signal vector (or difference signal vector) to restore the offset to the original value. Therefore, it can be said that the encoding of sounds corresponding to the front tongue and the rear tongue becomes reliable. That is, the offset compensator 19 gives the offset values O ₁ to Of ₄ in the equation to the emphasized difference signal vector (or difference signal vector).
In this case, the difference signal vector for offset compensation is correctly divided into positive and negative, or high _and low, with respect to the phonetic symbol, so that the encoding process is sequentially determined (branch and limit) as shown in Figure 22. ), it becomes possible to perform symbolization more easily and quickly than by performing a linear conversion operation such as an equation, and it can be applied to a simpler voice input device. The sequential discrimination processing as shown in FIG. 22 easily realizes the functions of the symbol vector conversion section 7 and the maximum value determination section 8a described above, and can be realized by a sequential discrimination processing program of a microcomputer. can. In the flowchart of the same figure, in the first step, the signal is divided into three groups depending on whether the V _eap /V _iu difference signal is at a high level H, a medium level M, or a low level L. It's divided. Then, in the second stage, when the first stage is H, if the V _a /Veo difference signal is H, then
Output the symbol a, if M, output the symbol h, if L, move to the third step, check the V _e /V _p difference signal,
If H, outputs e, if M, outputs w, and if L, outputs o. On the other hand, if the first stage is M,
In the second stage, if the V _e /V _p difference signal is H, it outputs f, if it is M, it outputs m, and if it is L, it outputs b. Furthermore, if the first stage is L, then in the second stage
If the V _i /V _u difference signal is H, it outputs i, if it is M, it outputs l, and if it is L, it outputs u. [Effects of the Invention] The present invention is configured as described above, and receives as input the difference signal output of a pair of filters that extracts the short-term average power of high-frequency components and low-frequency components of audio input. 5 Japanese vowels and other vowels according to the magnitude relationship of the difference signal outputs of multiple pairs of filters that extract the short-term average power of different frequency regions from the voice input. Voiced sound analysis means is provided which assigns any one of the codes to voiced sounds, and among the outputs of the comparison means, the code of the voiced sound is replaced with the code output from the voiced sound analysis means, , unvoiced sounds, five vowels, and other voiced sounds, and each input pattern formed when multiple types of voice messages are uttered in a standard manner is registered in advance as a standard pattern. , a voice message identification method that identifies a standard pattern that is most similar to a pattern as an input message, includes a differential vector converter that calculates a differential signal of the difference signal output of each filter pair, and a differential vector converter that calculates a differential signal of the difference signal output of each filter pair, and adds the differential signal to the original difference signal. Since the present invention is equipped with a difference signal emphasizing section that emphasizes the rising and falling edges of the difference signal, it has the advantage of being able to emphasize the rising and falling edges of the difference signal and ensuring reliable symbolization. In particular, in the case of a syllable consisting of a combination of a consonant and a vowel, as described in the explanation of the embodiment, the difference signal corresponding to the first formant includes a difference signal at the rise when changing from a narrow jaw state to a wide jaw state. A positive weight can be added to the difference signal corresponding to the second formant, and a negative weight can be added to the difference signal corresponding to the second formant at the trailing edge when changing from the front tongue state to the back tongue state, so that differences between speakers and It has the advantage that even if there are differences in speaking style, it can be encoded well. Furthermore, in the case of the combined invention, a differential vector conversion unit that calculates a differential signal of the difference signal output of each filter pair, and an offset calculation unit that calculates an offset value according to the positive/negative sign of the differential signal and its absolute value. A positive offset value is added to a rising difference signal, and a negative offset value is added to a falling difference signal. By adding the offset value of In particular, in the case of syllables consisting of consonant and vowel combinations, the difference signal corresponding to the first formant has a range from narrow to wide. A positive offset is added at the rising edge when transitioning to the chin state and returned to the original value at the falling edge.A negative offset is added to the difference signal corresponding to the second formant at the falling edge when transitioning from the anterior tongue condition to the posterior tongue condition. It has the advantage that it can be restored to its original state at the time of rise, and that the characteristics of the vowel following the consonant can be extracted well.

[Brief explanation of drawings]

Figure 1 is a block diagram of the conventional example, Figure 2 is a block diagram showing the configuration of the acoustic analysis section used in the same example, and Figure 3 is a block diagram showing the configuration of the acoustic analysis section used in the same.
Figures a and b are diagrams showing the characteristics of the filter bank used in the above, Figures 4 a and b are schematic circuit diagrams showing the configuration of the difference signal vector converter used in the same, and Figure 5 is a pattern converter used in the same. 6 and 7 are explanatory diagrams of the same operation as above, FIG. 8 is a block diagram of an embodiment of the present invention, FIG. 9 is a schematic circuit diagram of a difference signal enhancement processing section used in the same, and FIGS. FIG. 21 is an explanatory diagram of the same operation as above, and FIG. 22 is a flowchart showing the sequential discrimination process used in the same. 3 is a frequency analysis section, 5 is a difference signal vector conversion section, 7 is a symbol vector conversion section, 16 is a differential vector conversion section, 17 is a difference signal emphasis section, 18 is an offset calculation section, and 19 is an offset compensation section.

Claims (1)

[Claims] 1. The difference signal output of a pair of filters that extracts the short-term average power of the high-frequency component and low-frequency component of the audio input, respectively, is input, and when the high-frequency component is stronger, the sign of the unvoiced sound is taken as the signal of the low-frequency component. A comparison means is provided that outputs the sign of voiced sound when the signal is stronger, and outputs the sign of silence when the high-frequency component and the low-frequency component are approximately the same, and multiple sets of short-term average powers in different frequency regions are extracted from the voice input. 5 in Japanese depending on the magnitude relationship of each difference signal output of the filter pair.
A voiced sound analysis means is provided which assigns one of the codes of vowels and other voiced sounds, and among the outputs of the comparison means, the code of the voiced sound is assigned to the code output from the voiced sound analysis means. Replace it with
Forms an input pattern consisting of a time series of symbols for silence, unvoiced sounds, five vowels, and other voiced sounds, and pre-registers each input pattern formed when multiple types of voice messages are uttered in a standard manner as a standard pattern. In a voice message identification method that identifies a standard pattern that is most similar to an input pattern as an input message, a differential vector conversion unit that calculates a differential signal of the difference signal output of each filter pair, and a differential vector converter that calculates a differential signal of the difference signal output of each filter pair, and 1. A voice message identification system comprising: a difference signal emphasizing section for adding up and emphasizing rising and falling edges of a difference signal. 2 Inputs the difference signal output of a pair of filters that extracts the short-term average power of the high-frequency component and low-frequency component of the audio input, and when the high-frequency component is stronger, the sign of unvoiced sound is input, and when the low-frequency component is stronger, the signal output is the signal of the unvoiced sound. Comparing means for outputting a silent code when high-frequency components and low-frequency components are substantially the same is provided, and each difference signal of a plurality of filter pairs extracts short-term average power in different frequency regions from the voice input. Japanese 5 depending on the size of the output
A voiced sound analysis means is provided which assigns one of the codes of vowels and other voiced sounds, and among the outputs of the comparison means, the code of the voiced sound is assigned to the code output from the voiced sound analysis means. Replace it with
Forms an input pattern consisting of a time series of symbols for silence, unvoiced sounds, five vowels, and other voiced sounds, and pre-registers each input pattern formed when multiple types of voice messages are uttered in a standard manner as a standard pattern. In a voice message identification method that identifies a standard pattern that is most similar to an input pattern as an input message, a differential vector conversion unit that calculates a differential signal of the difference signal output of each filter pair, and a differential vector converter that calculates a differential signal of the difference signal output of each filter pair, and 1. A voice message identification system comprising: an offset calculation unit that calculates an offset value according to an absolute value; and an offset compensation unit that adds the offset value to an original difference signal.