JPS6331798B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a voice message identification method for operating a controlled device using voice messages. Figure 1 shows the schematic configuration of a conventional voice message identification device developed by the present inventors. In the figure, audio first enters from microphone 1, is high-frequency emphasized by preamplifier 2, and logarithmically processed by LOG amplifier 3. It is converted into an output proportional to the decibel value, and the AC
Only the AC component is amplified by the amplifier 4. Next, use low frequency filter bank 5 to filter out low frequency components (1KHz or less).
is taken out, and a high frequency component (5KHz to 12KHz) is taken out using a high frequency filter bank 6. The filter banks 5 and 6 include band filters F ₁ , F ₂ ,
Averaging circuit consisting of rectifier circuits D ₁ , D ₂ and an integrating circuit
It consists of M ₁ and M ₂ . Integrating circuit M ₁ of the low frequency filter bank has a time constant of about 5 to 10 msec, and integrating circuit M ₂ of the high frequency filter bank has a time constant of 1 to 10 msec.
It is set to about 2msec. The respective outputs of both filter banks 5 and 6 are input to a differential amplifier 7, and the output thereof, that is, the difference between the high frequency component and the low frequency component, is an averaging circuit 8 consisting of an integrating circuit with a time constant of about 20 msec.
is averaged. The analog signal waveform y(t) output from this averaging circuit 8 represents the ratio of voiced sound V to unvoiced sound U included in the input speech. Figures 2a and b show the analog signal waveform y(t) for the voice input example "kneading set" of our voice-controlled massage chair.
In this embodiment, the voiced sound V takes a positive value, and the unvoiced sound U takes a negative value. This signal voltage y(t) is applied to the V determination circuit 9 and the U determination circuit 10, and is sampled at regular intervals by the clock CK from the clock generation circuit 11. Both judgment circuits 9 and 10 each have a reference voltage.
R _V and R _U are added, and if the analog signal voltage is above the positive reference voltage R _V during sampling, it is voiced V, and if it is below the negative reference voltage R _U , it is voiceless U, R _V and R _U If the value is between , it is determined that there is no sound S. In Figure 2 a and b, z(t)
indicates the VU judgment output signal obtained from the signal processing circuit 12, which takes the value -1 for silent U, 0 for silent S, and +1 for voiced sound. There is. Figure 2a shows the case where the LOG amplifier 3 is used, and figure 2b shows the case where the LOG amplifier 3 is removed.As is clear from comparing the two, it is clear that the input audio is filtered by the filter bank 5, If LOG conversion is performed before step 6, voiced sounds V and unvoiced sounds U can be reliably distinguished. Next, the output of the preamplifier 2 before passing through the LOG amplifier 3 is filtered by a band filter of 1.5KHz to 2.5KHz, for example.
A filter bank 13 consisting of F ₃ , a rectifier circuit D ₃ and an averaging circuit M ₃ and a filter bank 14 consisting of a 2.5KHz to 3.5KHz band filter F ₄ , a rectifier circuit D ₄ and an averaging circuit M ₄ Characteristic components in the frequency domain are extracted. The low frequency sound V _L which is the output of the filter bank 13 and the high frequency sound V _H which is the output of the filter bank 14 are compared by a differential amplifier 15.
It is averaged by the averaging circuit 16. The output of the averaging circuit 16 is V _L which operates in synchronization with the clock CK.
The output voltage of the averaging circuit 16 is input to the determination circuit 17 and the V _H determination circuit 18, and the output voltage of the averaging circuit 16 is the reference voltage.
If it is lower than R _L , it is determined to be a V _L component, if it is higher than the reference voltage R _H , it is determined to be a V _H component, and if it is between R _L and R _H , it is determined to be a midrange sound V _M. The signal processing circuit 19 converts each component of V _H , V _L , and _VM into three values into outputs of +1, 0, and -1, respectively. Each output of the signal processing circuits 12 and 19 is read into the CPU 21 via the I/O port 20, and the read data is temporarily stored in the input pattern memory 22, and is then transferred to the plurality of data stored in the standard pattern memory 23. A comparison is made to determine which of the standard patterns is closest to the standard pattern, and the controlled device 24 is controlled based on the comparison and determination result. These verification and judgment operations are carried out by the CPU 21, program ROM 25, and working RAM.
This is carried out by a microcomputer 27 having 26. However, in such conventional examples, since there is only one standard pattern to be compared and determined with the input pattern, it is necessary to change the utterance rate or the manner of utterance even when the utterances of the voice message are different, or even for the same utterance. If this was changed, there was a problem that voice messages could no longer be recognized correctly. Conventionally, the voices of multiple speakers are recorded, or multiple utterances of the same speaker are registered to prepare multiple standard patterns for the same voice message. A voice message recognition method has been proposed that controls a controlled device according to the matching output when the message matches one standard pattern. If a variety of standard patterns were to be prepared in the standard pattern memory 23, there was a problem in that the capacity of the standard pattern memory 23 would become significantly large. The present invention has been made to solve these problems of the conventional example, and it is possible to reliably recognize a voice message even if there are various changes in the manner of utterance of the voice message. It is an object of the present invention to provide a voice message identification method that allows memory capacity to be kept as small as possible. The configuration of the present invention will be described below with reference to illustrated embodiments. FIG. 3 shows a block diagram of a voice message recognition device that implements the method of the present invention. As shown in the figure, the voice message recognition device includes an acoustic processing section 28, a frequency analysis section 29, and an encoding processing section. 3
0, and a comparison determination section 31. Of these, the acoustic processing section 28 and frequency analysis section 29 have the same configuration as the conventional device,
Filter banks 5, 6 and filter bank 1
Judgment circuits 32 and 33 connected to the subsequent stage of 3 and 14
is constituted by circuits from differential amplifiers 7 and 15 to signal processing circuits 12 and 19 in FIG. However, the pass frequency bands of band filters F ₃ and F ₄ are 0 to 500 Hz and 0 to 500 Hz, respectively, as described later.
It is set between 500Hz and 1KHz. These I/O
The circuit at the stage before the port 20 is composed of an analog IC, and the circuit at the stage after the I/O port 20 is composed of a microcomputer. In the encoding processing unit 30, 34 is an S counter for detecting a silent section, and when a voiced sound V or an unvoiced sound U is detected, it is reset and starts the voice input operation. If it continues for more than 0.2 seconds, the count-up will count up and the voice input operation will stop. The voice input operation is performed according to a sampling period of 5 to 20 msec (in the embodiment, a period of 5 msec), and one of the determination circuits 32 outputs each symbol of a voiced sound V, an unvoiced sound U, and a silent sound S and their durations. The other judgment circuit 33 inputs the codes and durations of the high-range sound V _H , middle-range sound V _M , and non-range sound V _L and stores them in the memory. It is becoming more and more common. The signals once stored in the memory in this way are shaped by waveform shaping processing sections 35 and 36. FIG. 4a shows the signal waveform before waveform shaping in the waveform shaping processing section 35, and FIG. 4b shows the signal waveform after the waveform shaping process. The waveform shaping process includes a first process in which a voiced sound V or an unvoiced sound U with _a short duration is treated as silence S, as shown by arrow a1 in FIG.
As shown in a ₂ to _{a 6} , the duration of the voiced sound V or the unvoiced sound U is relatively long, but the duration of the silence S that follows is short, and the next code following the silence S is the same as the code before the silence S. If they are the same, then the second process erases the silence S.
FIG. 5 is a flowchart showing the procedure of waveform shaping processing in the waveform shaping processing section 35. This flowchart is a program for calculating data stored in memory in a list format of codes V, U, S and their durations to create new list format data and storing it in memory again. It shows. First, it is determined whether the first code of the list before formatting is S, and if the first code is not S, it is determined whether its duration is greater than the reference value _T1 . Duration is standard value
If not greater than T ₁ , its sign V or U
is converted into a code S, and the code S and its duration are written into the memory as a formatted list. If the duration is larger than the reference value _T1 , the code V or U and the duration are directly transferred as a formatted list. Then, it is determined whether the duration of the next code S is shorter than the reference value _T2 , and if it is shorter than the reference value _T2 , the code next to that code S is the same as the code before the code S. Determine whether or not. If they are not the same, the code S and its duration are transferred as they are to the formatted list. If they are the same, change the code S to the previous or next code U or V,
The changed code and its duration are written into memory as a formatted list. Furthermore, it is determined again whether the duration of the next code S is shorter than the reference value _T2 . The above arithmetic processing operation continues until all the lists before formatting have been converted, and when the conversion process is finished, it is determined whether the last sign of the list after formatting is S, and the last sign is If is S, that code S is removed from the formatted list, and it is again determined whether the last code is S. However, the last sign is S
If not, the series of waveform shaping operations is completed, and at this time, the waveform shaped signals as shown in FIG. 4b are stored in the memory with symbols U, V,
It is stored in the form of a list of S and its duration. Similar arithmetic processing operations are also performed in the other waveform shaping processing section 36, including an operation of converting the code V _H and the code V _L with the shorter duration into the code V _M , and a code V _H
and the short-duration code sandwiched between the code V _H
Operation of converting V _M to code V _H , code V _L and sign
It performs the operation of converting the short-duration code V _M between the code V _L and the code V _L into the code V L. Next, FIG. 6 shows a composite encoding processing section 37 for combining the two lists formed by the waveform shaping processing sections 35 and 36 into one list.
3 is a flowchart showing the processing operation of FIG. To explain its operation, first, the waveform shaping processing section 35
Examine the list of codes U, V, S and their durations formed by and determine whether the first code is U or S. If the first code is U or S, that code U or S and its duration are transferred directly to memory as a composite code list. Further, when the first code is V instead of U or S, the waveform shaping processing unit 35
Examine the list of codes V _H , V _M , V _L and their durations formed by , and first determine whether the code V _H is included in the duration of the code V. If so, the codes V _H and their durations are transferred to memory as a composite code list. Also, regarding the codes V _M and V _L , if they are included in the duration of the code V, the codes V _M and V _L
and its duration are each transferred to memory as a composite code list. A composite code list is obtained in the above manner, but as in this embodiment, waveform shaping is performed separately for the system of codes V, U, S and the system of codes V _H , V _M , V _L. In addition to the method of performing composite encoding after processing (see Figure 7a), there is also a method of first performing composite encoding and then performing waveform shaping processing, as shown in Figure 7b. It is something. In this case, the symbols U, V,
Composite encoding is performed by performing logical operations as shown in Table 1 based on each logical value of V _H and V _L.

[Table] Next, FIG. 8 is a flowchart showing the processing operation of the hierarchization processing section 38. The layering processing section 38 creates a first layer list and a second layer list from the composite code list so that pattern matching can be performed step by step according to the structure of the voice in the matching/judging section 31 at the subsequent stage. It is.
Among these, for the first layer list, the symbols U,
Since it is the same as a code list consisting only of V, S and their duration, either the output list of the waveform shaping processing section 35 can be transcribed as is, or all the codes V _H , V _M , and V _L in the composite code list can be changed to the code V This can be easily obtained by replacing . Next, regarding the second layer list, the consecutive codes V _H , V _L ,
It is obtained by transcribing a code list consisting of V _M and its duration. Therefore, if it is assumed that n codes V are included in the first layer list, n pieces of the second layer list are also obtained. Furthermore, FIG. 9 is a flowchart showing the processing operation of the normalization processing section 39 for normalizing the code durations for the first layer list and the second layer list obtained as described above. The normalization processing unit 39 calculates the durations of the first layer list consisting of codes U, V, S and their durations, and the n second layer lists consisting of codes V _H , V _M , V _L and their durations. Normalization on the time axis is performed so that the sum of the sums is constant. Table 2 shows the codes V, U, S and their durations T _K for the first layer list.
and normalized duration T _S , where duration T _K corresponds to the number of samples.

[Table] To explain _the normalization processing operation using _the flowchart in FIG. Find the coefficient P _S = 1000/ΣT _K. Next, for each code U, V, S, its duration T _K (j) is multiplied by the normalization coefficient P _S to obtain the normalized duration.
The purpose is to find T _S (j) in order. When the duration normalization operation for the first hierarchical list is completed as described above, the duration normalization process is performed for each of the n second hierarchical lists by the same operation. Table 3 (a) to (d)
The second layer list created for the four codes V included in the layer list (see Table 2) and its normalized duration are shown, respectively.

【table】

【table】

【table】

[Table] Table 3 (a) shows the second layer list V ₁ corresponding to the first code V (duration 3415) of the first layer list shown in Table 2, and the following Table 3 (b) ) to (d) show second layer lists V ₂ to V ₄ corresponding to codes V having durations of 3621, 1437, and 2637. By normalizing the duration as described above, the recognition rate can be increased because it becomes less susceptible to the influence of the speaking speed. The duration T _K data shown in Tables 2 and 3 is simulation data when the voice message ``Senaka wo Susure.'' is analyzed using a 50 μsec sampling pulse. When performing analysis using sampling pulses, the duration (ie, the number of samples) is 1/100 of the value in the table. Codes U, V, S of the first layer list and codes of the second layer list normalized as above
V <sub>H</sub> , <sub>VM</sub> , and V <sub>L</sub> are converted into ternary codes of +1, 0, and -1 in the ternary encoding processing section 40 . That is, first, the code V in the first layer list corresponds to +1, the code U corresponds to -1, and the code S corresponds to 0, and the code V <sub>H</sub> in the second layer list corresponds to
corresponds to +1, the sign V <sub>M</sub> corresponds to 0, and the sign V <sub>L</sub> corresponds to -1, respectively. By doing this, the standard pattern memory 42 in the distance calculation matching unit 41
The calculation speed can be significantly increased when comparing the contents of the first layer list and the second layer list with the contents of the first layer list and the second layer list. That is, the distance calculation matching section 41 calculates the cross-correlation coefficient between the ternary data of +1, 0, -1 stored in the standard pattern memory 42 and the data output from the ternary encoding processing section 40. Although it is becoming
Since there are only three types of data: +1, 0, and -1,
The cross-correlation coefficient can be calculated at an extremely high speed by simple logical operations and addition/subtraction without the need for numerical multiplication. The cross-correlation coefficients calculated for each standard pattern are stored in the primary layer identification section 43 and the secondary layer identification section 44, and compared in magnitude in the determination processing section 45. It is determined that the pattern is Here, the cross-correlation coefficient is defined as f <sub>1</sub> (t), which is the change in the value of the standard pattern with respect to the change in time t, and f <sub>2</sub> (t), which is the change in the value of the input pattern such as the primary hierarchical list or the secondary hierarchical list. t) is given by the following equation. f <sub>12</sub> (τ) = ∫ <sup>∞</sup> <sub>-∞</sub> f <sub>1</sub> (t) f <sub>2</sub> (t - τ) dt Figures 10a and b show the change in the value of the standard pattern f <sub>1</sub> (t) with respect to the change in time t, and the input pattern f <sub>2</sub> (t), and as shown in the figure, f <sub>1</sub> (t) and f <sub>2</sub> (t) are +1, 0,
Since there are only three values, −1, the product of both f <sub>1</sub>
The value of (t)f <sub>2</sub> (t) also takes only one of +1, 0, and -1, which makes calculation of the cross-correlation coefficient very easy. When such a cross-correlation coefficient f <sub>12</sub> (τ) is calculated using a microcomputer, it can practically be sufficiently calculated by numerical calculations as shown in the following equation. f <sub>12</sub> (τ) = <sub>N</sub> 〓 <sup>t=0</sup> f <sub>1</sub> (t) f <sub>2</sub> (t - τ) By the way, the cross-correlation coefficient f <sub>12</sub> (τ) is the standard pattern f <sub>1</sub> (t) and the input pattern f <sub>2</sub> that are multiplied together. (t-
It is a function of the phase difference τ with respect to τ), and takes the maximum value at a specific phase difference τ. Therefore, the distance calculation matching unit 41 finds the point where this cross-correlation coefficient f <sub>12</sub> (τ) is maximum, and
The maximum value is calculated for each standard pattern, and 1
The next and second layer identification sections 43 and 44 respectively store the pattern, and finally, the judgment processing section 45 compares the magnitude relationship to judge the standard pattern closest to the input pattern. By the way, in the present invention, when comparing a code pattern extracted from a voice message with a standard pattern, the code pattern is separated into a primary layer list and a secondary layer list, and the comparison is performed on the primary layer list. Later, the verification process is performed in stages by performing verification on the secondary hierarchy list, but this is done by first extracting the features corresponding to the macroscopic structure of the audio, and then extracting the features corresponding to the macroscopic structure of the audio. It is better to extract features that correspond to microscopic features.
This is because voice recognition can be performed efficiently and reliably. Figure 11 shows the characteristics of speech in a hierarchical manner.The speech is first divided into a voiced sound V accompanied by vocal fold vibration, and an unvoiced sound U without vocal fold vibration.
The voiced sounds V are classified into a voiced sound /a/ group with a wide jaw opening and a voiced sound /i/ group with a narrow jaw opening. A voiced sound with a wide jaw opening corresponds to the above-mentioned high-frequency sound V <sub>H</sub> , and the frequency of the first formant of the voice is relatively high, and the frequency band is mostly distributed from 500 Hz to 1 KHz. In addition, voiced sounds with a narrow jaw opening correspond to the above-mentioned low-frequency sound V <sub>L</sub> ,
The frequency of the first formant of speech is relatively low, and its frequency band is mostly distributed between 0 and 500 Hz.
Voiced sounds with wide jaw opening include vowels /a/, /
〓/, /ε/, etc., and voiced sounds with narrow jaw opening include vowels /i/, /e/, /o/, /u/.
, nasal consonants, and other voiced consonants. In addition, the unvoiced sound U includes a stationary unvoiced sound, that is, a voiceless fricative U <sub>F</sub> , and a transient unvoiced sound, that is, a voiceless plosive U <sub>B</sub>
There is. However, in order to accurately recognize voice messages word by word, it is necessary to identify all of the characteristics of these voices. In the case of control, it is not necessary to completely identify all consonants and vowels, and it is sufficiently practical to extract macroscopic features. The features of such speech are listed in order from macroscopic features as follows. 1 Is it a voiced sound V or an unvoiced sound U? Such characteristics may be due to the fact that there are many low-frequency components (1KHz or less) in the frequency spectrum of the voice, or high-frequency components (5KHz to 12KHz).
It can be determined by whether there are many Hz). 2 If it is a voiced sound V, it is a voiced sound with a wide jaw opening.
V <sub>H</sub> (/a/ group) or a voiced sound with narrow jaw opening V <sub>L</sub> (/i/ group). This characteristic is due to the fact that there are many high-frequency sounds V <sub>H</sub> (500Hz to 1KHz) in the frequency spectrum of voiced sounds, or there are many low-frequency sounds V <sub>L</sub> (0
~500Hz). 3 If it is a voiceless U, is it a voiceless fricative U <sub>F</sub> ?
Is it a voiceless plosive U <sub>B</sub> ? Such characteristics can be determined depending on whether the unvoiced sound is stationary or transient. That is, the determination can be made based on whether the duration of the unvoiced sound U is long or short. 4 In the voice message, each feature V <sub>H</sub> , V <sub>L</sub> ,
The time occupied by U <sub>B</sub> , U <sub>F,</sub> etc., or the proportion of the duration of a voice message. Such characteristics can be determined by referring to the durations in the first hierarchical list and second hierarchical list described above. In addition, front vowels (corresponding to /i/, /e/) and high tongue vowels (corresponding to /u/, /o/) can be distinguished depending on whether the second formant of the voice is high or low. It is something that can be done. Figure 25a shows the articulation points of the vowels /a/, /i/, /u/, /e/, /o/, and Figure 25b shows the frequency distribution of the first and second formants of the vowels. (Quoted from pages 363-364 of ``New Edition of Hearing and Speech'' by the Institute of Electronics and Communication Engineers (supervised by Dr. Miura)). FIG. 26 shows the distribution of the first formant and second formant of Japanese vowels for male and female voices, respectively. As is clear from the distribution of the second formant shown in FIG. 25b and FIG.
If we analyze the output of the 3.2KHz band filter, we get
The position of the second formant can be detected, and thereby features corresponding to the front and back of the tongue position can also be extracted. Of course, even without extracting such microscopic features, voice messages for controlling equipment can be fully recognized. For example, Figure 12 shows the voice input/
This is an example of the frequency spectrum of senakaosasure/, which is the audio input sampled at 20KHz.
200 samples (10msec) as one frame, 20
The following LPC analysis was performed, and the power of the unvoiced sound /s/ is concentrated above 5KHz.
It can also be seen that voiced sounds have a peak of power below 1KHz. Furthermore, it can be seen that in voiced sounds, the power of /a/ and /o/ is concentrated between 500Hz and 1KHz, and the power of /n/ and /u/ is concentrated between 0 and 500Hz.
Furthermore, it can be seen that for voiced sounds, the same spectrum continues for several times (number + msec) corresponding to each phoneme. Furthermore, Fig. 13a shows the changes in the voiced sound component V and the unvoiced sound component U for the same voice input as above, and Fig. 13b shows the changes in the high-frequency component V _H and low-frequency component V _L of the voiced sound. This shows a change in
First, in Figure 13a, there is no voice /s/, /k/
The part corresponding to indicates U, /na/, /
The parts corresponding to ao/, /a/, and /ure/ clearly indicate V. Also, in Figure 13b, /
The part corresponding to n/, /sa/ is V _L , /
The parts corresponding to a/, /ao/, and /e/ are _VH . Therefore, as described above, the first layer list corresponds to voiced sound V, unvoiced sound U, and silent sound S, and corresponds to high-range sound V _H , middle-range sound _VM , and low-range sound V _L among voice sounds. Most voice messages can be identified by comparing the secondary hierarchy list with pre-stored standard patterns. However, the above is just a generalization, and voiceless plosives in syllables can occur when the speakers of a voice message are different, or when the same speaker changes the rate of speech or the manner of vocalization. Phenomena occur in which a sound cannot be detected or a voiced sound sandwiched between unvoiced sounds in a syllable becomes voiceless, so it is necessary to create a standard pattern that can accommodate all the various ways in which voice messages are uttered. There is. The present invention enables voice messages to be correctly recognized even when input patterns fluctuate due to such subtle changes in vocalization. That is, in the present invention,
As standard patterns, for example, as shown in Fig. 14, in addition to the basic pattern consisting of a time series of codes C ₁ , C ₂ , C ₃ , C ₄ , C ₅ , C ₆ , there are also patterns such as codes C ₂₄ and C ₄₅ . Add a branch pattern and encode the input pattern
Match with the first derived pattern consisting of the time series of C ₁ , C ₂₄ , C ₅ , C ₆ and the second derived pattern consisting of the time series of codes C ₁ , C ₂ , C ₃ , C ₄₅ , C ₆ The recognition rate for voice messages has been increased. The concepts of the basic pattern, branch pattern, and derived pattern that constitute the standard pattern in the present invention will be explained below with specific examples. The most typical example of variations in speech input patterns is, first of all, voiceless plosives /p/, /
Examples include the phenomenon of missing t/ and /k/. That is, as can be seen with reference to FIG. 13a, the voiceless plosive /k/ is a transient unvoiced sound and therefore has a short duration, making it extremely difficult to detect. Compared to this, the unvoiced fricative /s/ is a stationary unvoiced sound, so its duration is long as shown in FIG. 13a, and it is easy to detect. For this reason, by slightly lengthening the sampling period, voiceless fricatives/
There may be cases where s/ can be detected but voiceless plosive /k/ cannot. Taking this point into consideration, FIG. 15 shows a standard pattern for the first layer list of the audio input /senakaosasure/, with the symbols U, S, V ₁ , S, U, S, V ₂ , S ，
In addition to the basic pattern consisting of the time series U, S, V ₃ , S, U, S, V ₄ , a code S equal to the duration of the code S, U, S sandwiched between the codes V ₁ and V ₂ A branch pattern consisting of the following is provided. Therefore, the encoded input pattern of speech is not only checked against the above basic pattern, but also the codes U, S, V ₁ , S, V ₂ , S,
It is also compared with a derived pattern consisting of U, S, V ₃ , S, U, S, V ₄ , so even if the voiceless plosive /k/ is missing from the input pattern, it will still be recognized as a voice. Messages can be recognized correctly. The numbers shown in Figure 15 indicate that, as a result of analyzing the same speaker's utterance patterns five times, there were four cases where the basic pattern was matched and one case where the derived pattern was matched. It shows. By configuring as above, voiceless plosive /p/, /
This can prevent a situation where t/ and /k/ are opened and dropped. Next, Figures 16a to 16c show voice input/
4 voiced sounds included in senakaosasure/ V ₁ ~
For V ₄ , an example of creating standard patterns for each secondary hierarchy list is shown. To explain the first voiced sound V ₁ ,
Its basic pattern is coded V _M , V _L , V _M , V _H , V _M ,
It consists of a time series of V _L and further has the sign V _M
It has three branch patterns consisting of: Therefore, in this case, the first derived pattern consisting of the time series of codes V _M , V _L , V _M , V _L and the codes V _M ,
A second derived pattern consisting of a time series of V _L , _VM , V _L , _VM and a third derived pattern consisting of a time series of symbols V _M , V _L , _VM , V _H , _VM are It is something that is formed. The numbers written in Figure 16a,
etc. indicate the number of branches as before. Note that the duration time of each branch pattern is equal to the duration time when the basic pattern is continued as it is without entering the branch pattern. By configuring the standard pattern as described above, voice messages can be reliably recognized even if the codes V _H , V _M , and V _L in the second layer list vary slightly. By the way, there are many different ways of changing the codes V _H , V _M , and V _L in such a second layer list.
To give an example, (a) when the V _L -V _H series becomes the V _L -V _M series (e.g. /na/ of /senaka/), (b) V _L -V _M
When the series becomes a V _L −V _L series (e.g. /sesure/
/re/), (c) V _L - V _M series becomes V _L - V _H series (e.g. /tomare/'s /re/), (d) V _H - V _L series becomes V _H - V _M series (e.g. /senaka/
/ak/), (E) When the code V _H becomes a V _L - V _H sequence (e.g. /sa/), (F) When the code V _H becomes a V L - _{V H -} V _L sequence ( Examples include /kat/ of /kat/). When the rules for fluctuations of these codes V _H , V _M , and V _L are summarized, they are roughly classified into the following two cases. 1 Due to the interaction of the preceding and following phonemes, the signs V _H and V _M
and that the signs V _L and V _M are interchanged. That is, the V _H - V _L sequence can become the V _H - V _M sequence or the V _M - V _L sequence, and the V _L
-V _H series can become V _L -V _M series or V _M -V _H series. 2 The code V _H is affected by the unvoiced sound before and after it, and the code V _L appears before, after, or both before and after the code V _H.
shall be added. That is, the code V _H is V _L
-V _H series, V _H -V _L series, or V _L -V _H -V _L
To be replaced in a series. Another example of variations in speech input patterns is the phenomenon of vowel devoicing. For example, in the case of Japanese people, the word "watakushi" is
Rather than someone who correctly pronounces watakusi/,
Many people skip the vowel /u/ and pronounce it as /wataksi/. This is because the vowel /u/ is sandwiched between a voiceless plosive /k/ and a voiceless fricative /s/, and generally one vowel sandwiched between a voiceless plosive U _B and a voiceless plosive U _B (e.g. /kiQpu/
/i/), voiceless plosives U _B and voiceless fricatives U _F
One vowel in between (for example /watakusi/
/u/) and a single vowel sandwiched between a voiceless consonant and a voiced consonant have a very strong tendency to be devoiced. In addition, there is also a slight tendency for one vowel sandwiched between the unvoiced sound U and the silent sound S (for example, /a/ in /dousa/) to become silent. Therefore, in general, for a vowel sandwiched between voiceless sounds, between voiceless sounds and voiceless sounds, and between voiceless sounds and voiced consonants, in addition to the basic pattern of making the vowel part a voiced sound V, Add a branching pattern that makes the vowel part the unvoiced sound U to the standard pattern,
If the above-mentioned specific vowel is pronounced clearly as a voiced sound V, it can be compared and determined by the basic pattern, and if the above-mentioned specific vowel is pronounced unclearly as if it were an unvoiced sound U, then The recognition rate of voice messages can be increased by making it possible to perform comparison and determination based on derived patterns. Next, a method of creating a standard pattern having such a basic pattern and a branch pattern will be described. There are two main ways to create standard patterns. One is to input the individual phoneme codes and their durations that make up the voice message from a keyboard, etc., and then use a branch processing program to automatically create the basic pattern and branch pattern. The other method is to register the same voice message multiple times by changing the way it is uttered or by changing the person who is saying it, and using the common characteristics as the basic pattern and the unique characteristics that are not common. This is a learning registration method in which branch patterns are registered, and the former is a deductive method and the latter is an inductive method. First, in the former method, as shown in FIG. 17, from the keyboard 50 /s/, /e/, /
n/, /a/, /k/, /a/, /o/, /
s/, /a/, /s/, /u/, /r/, /e/
This method involves sequentially inputting each phoneme code and its duration. First, it is determined whether each phoneme code is a voiced sound V. If it is a voiced sound V, it is a vowel/
The code V _H is assigned to a/, and the vowel /i/
and voiced consonants /m/, /b/ are assigned the code V _L , and other voiced consonants and vowels /e/, /
For u/ and /o/, branch patterns are created assuming that they can be any of the codes V _H , V _M , and V _L . Also, the code U is assigned to no voice, and the code S is assigned to no voice. Next, the duration is input, and for unvoiced sounds with short durations, that is, unstretched plosives, in addition to the basic pattern consisting of the symbol U, a branch pattern consisting of the symbol S is added. Furthermore, by inputting a code sequence, a branch pattern consisting of another code U to the basic pattern consisting of the code V is added to a monophthong sandwiched between an unvoiced sound and a voiceless sound or a voiced consonant.
By doing the above, it is possible to automatically create a standard pattern in which a branch pattern that increases the voice message recognition rate is added to the basic pattern. Next, the learning registration method will be explained. 18a to 18c are symbols V _H corresponding to the second layer list,
This shows the case where a standard pattern consisting of V _M and V _L is created, and FIG. 19 is a flowchart showing the creation procedure. First, as shown in Figure 18a, the same word is registered multiple times, and the normalized time is
Divided into 10 regions, the part where the sign does not change in the same time domain is defined as a core pattern, and the part where the sign changes in the same time domain is defined as _VM . At this point the 18th
A learning basic pattern as shown in Figure b is created. Next, in the same time domain, the part where V _M becomes V _L is
Add V _L branch pattern. In addition, a V _H branch pattern is created in a portion where V _M or V _H occurs in the same time domain. Furthermore, the portion that becomes V _H and V _L in the same time domain is left as V _M. At this point, a learning standard pattern having a branching pattern as shown in FIG. 18c is formed. The learning standard pattern thus obtained is registered and stored in the standard pattern memory 42. However, in the present invention, the learning registration method and the non-learning registration method are a compromise between the registration processing unit 46.
A flowchart is shown in Fig. 20. First, the data input to the registration processing unit 46 is entered into the SUV series or SUV series in the first layer list.
- It is determined whether there is a V series, and if so, S
- A standard pattern is formed that includes both the UV and SV code sequences. Next, regarding the second layer list, it is possible to switch between creating a standard pattern in the learning mode and creating a standard pattern in the non-learning mode as shown in FIG. 19 above. If that doesn't work, you can use the other mode. Since the operation in the learning mode has already been explained using the flowchart in FIG. 19, the operation when creating a standard pattern for the second layer list in the non-learning mode will be explained using the flowchart in FIG. 21. . First, if the first code in the second layer list is V _L , a standard pattern including a basic pattern ST-V _L and a derived pattern ST-V _M is created. Also, if the first sign is V _H , the basic pattern
In addition to ST―V _L ―V _H , there are two derived patterns ST―
A standard pattern including _VM - V _H and ST - V _L - _VM is created. Furthermore, if the first code is _VM , a basic pattern ST--a standard pattern of only _VM is created. Next, determine whether the last code is V _H , V _L , or V _M , and whether each code included between the first code and the last code is a V _L -V _H sequence or V _H -V Depending on _{the L} series, a standard pattern with a branch pattern as shown in the flowchart of FIG. 21 is automatically formed. By the way, when creating a standard pattern for the second layer list in this way, it is necessary to correctly identify the code _VH and the code _VL . As mentioned above, the code V _H is a high-frequency voiced sound (/a/ group)
, and the code V _L corresponds to a low-frequency voiced sound (/i/ group). However, in the present invention, as shown in FIG. 22, the V _H analysis system and the V _L analysis system are A variable resistor VR ₁ for adjusting the output balance and a variable resistor VR ₂ for offset adjustment are provided, so that when the vowel /a/ is uttered, the sign V _H is always detected and the vowel /i/ is uttered. When this happens, the code V _L is always detected. However, strictly speaking, the optimal value of this balance may differ depending on the personality of the speaker. Therefore, the inventors of the present invention have discovered that it is sufficient to adjust the balance so that the V _H /V _L difference signal becomes zero when the vowel /e/ is naturally generated. 23rd
The figure shows the principle. As shown in the figure, the first formant of the vowel /a/ is distributed from 500Hz to 1KHz, and the first formant of the vowel /i/ is distributed from 0 to 500Hz.
Hz, and the first formant of the vowel /e/ is roughly located in the middle. Therefore, by adjusting the balance between V _H and V _L using the vowel /e/ as a reference, the optimum balance value can be obtained. Finally, regarding the secondary hierarchy list, each code V _H ,
A verification method that takes into consideration the duration of V _M and V _L will be explained. FIG. 24 is a flowchart showing three methods of checking and identifying the secondary hierarchy list, and the most appropriate method is selected and used. The first method is to determine whether the most common code among the plurality of voiced sounds V ₁ to V _o included in one voice message is V _H or _VM or V _L The second method is to check whether each voiced sound V ₁ ~
The third method is to check whether the proportion of V _H included in V _o matches the input pattern and the standard pattern.
The number of cases in which _VM matches the standard pattern _VH or _VL and the number of cases in which _VM in the standard pattern matches the input pattern _VH or _VL are listed and compared. Therefore, for all voiced tones V ₁ to _{V o} in a voice message,
The input pattern is compared with a plurality of standard patterns using the most appropriate one of the three types of matching methods mentioned above, and the standard pattern with the most matching characteristics is determined. . Furthermore, the degree of agreement between the input pattern and the standard pattern is determined by the corresponding score +1, 0, - for each sample.
It is also possible to evaluate by 1 and judge by the total score. Table 4 shows this scoring method, and the basic idea is almost the same as when calculating the cross-correlation coefficient between ternary codes described above. Then, scores are assigned according to the rules in Table 4, and when the total score calculated for each sample is greater than or equal to a predetermined value, it is determined that they match, and when it is less than or equal to the predetermined value, it is determined that they do not match. It is something to do. Therefore, if the total number of samples is 1000, the total score will be 1000 when the patterns match perfectly.

[Table] In the present invention, in addition to the basic pattern, a branch pattern that branches from the basic pattern is provided as a standard pattern to be matched with the input pattern, so that multiple derived patterns can be formed. It is possible to determine whether any of the patterns match the input pattern, but on the other hand, if the input pattern is varied, the standard pattern becomes unique. It is also possible to make it a thing.
In other words, by performing operations such as filling in voiceless plosives that are missing from an input pattern or restoring devoiced vowels, multiple derived input patterns are created from one input pattern, and these are combined into one. If it is configured to match with a standard pattern on the street, the recognition rate can be increased as in the case where the sign pattern is varied. As described above, the present invention uses a filter to extract low-frequency components where the energy of voiced sounds is concentrated and high-frequency components where the energy of unvoiced sounds is concentrated from a voice input, and the difference signal output level between the signals extracted by the filters. The voice message is converted into an input pattern consisting of a time series of first, second, and third codes according to the size of the code, and this input pattern is compared with multiple types of pre-recorded standard patterns and input. In a message identification method, when the same voice message is input multiple times in different ways, the input pattern that occurs with the highest probability is defined as the basic pattern, and the input pattern that occurs with a lower probability than the basic pattern is defined as the derived pattern. The part where the pattern does not match with the above basic pattern is recorded in advance as a branch pattern that branches from the basic pattern, and when the input pattern is checked against the basic pattern and is not coded, the derivation caused by the combination of the basic pattern and the branch pattern is recorded. Since it is designed to provide a standard pattern that allows for branch matching processing such as matching a pattern with an input pattern, input data may be input due to the gender, age difference, speaking speed difference, or dialect accent of the speaker. Even if the characteristics of a voice message vary slightly, input patterns that are slightly different from the basic pattern can be sufficiently recognized by comparing them with various derived patterns created by the combination of the basic pattern and branch patterns. In addition, in the present invention, an input pattern that occurs with the highest probability is taken as a basic pattern, and derived patterns that occur with a lower probability than the basic pattern are recorded in the form of a branch pattern that branches from the basic pattern. Because of this, a very large number of derived patterns can be stored with a very small memory capacity, which has the effect of realizing an inexpensive voice message identification method that has a very high recognition rate. In other words, in the present invention, when the characteristics of an input voice message change slightly and do not match the basic pattern, the derived pattern to be matched is created by combining the basic pattern and a pre-stored branch pattern. This branch pattern is data only for the part where the derived pattern and the basic pattern do not match, and compared to the case where the derived pattern is stored as is, the basic pattern part is overlapped. Since the data has a small number of bits corresponding to the amount of data that is not stored, the memory capacity can be significantly reduced when preparing a large number of derived patterns to increase the recognition rate. This has the effect of realizing a voice message recognition method. In addition, in the combined invention described in claim 9, a plurality of input patterns formed by inputting the same voice message multiple times in different modes during the recording operation of the standard pattern are inputted at fixed time intervals. For each time-divided section, the code that does not change is assigned as a core pattern for the section where the code does not change.
All the third codes are assigned to the interval where the sign changes, and a basic pattern is created using this third code and the core pattern, and the code can be the first code in the interval where the sign changes. For intervals where the code cannot be the second code, a branch pattern consisting of the second code is added to the basic pattern, and for an interval where the code cannot be the second code, a branch pattern consisting of the first code is added to the basic pattern. Since patterns and branching patterns are recorded in advance as standard patterns, it is possible to register the same voice message multiple times while changing the speaker, or to register the same voice message multiple times while changing the way the voice is uttered. By registering the same voice message multiple times, characteristics that are common to each voice message can be automatically extracted as basic patterns, and unique characteristics that are not common to each voice message can be automatically extracted as branch patterns. This has the advantage that it can be added to the basic pattern in any case, and that it can significantly simplify the task of creating a standard pattern for performing branch matching processing. In the voice message identification method described above, codes corresponding to voiced sounds, unvoiced sounds, and silence are used as the first, second, and third codes, and codes corresponding to high-range voiced sounds, low-range voiced sounds, and mid-range sounds are used. If codes corresponding to vocal sounds are used, matching operations can be performed according to the structure of the voice, thereby significantly increasing the recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the conventional example, FIG. FIG. 5 is a waveform diagram showing the operation of the waveform shaping processing section; FIG. 5 is a flowchart showing the operation of the same waveform shaping processing section;
This figure is a flowchart showing the operation of the composite encoding processing section same as above, FIGS. 7a and 7b are block diagrams of the encoding processing section same as above, and FIG. Fig. 9 is a flowchart showing the operation of the normalization processing section same as above, Fig. 10 a and b are waveform diagrams showing the operation of the distance calculation matching section, Fig. 11 is a diagram showing the audio characteristics in a hierarchical manner, and Fig. 12 The figure shows the frequency spectrum of the voice, Figures 13a and 13b are waveform diagrams of signals extracted from the voice, Figure 14 is a diagram showing the principle of the branch matching process of the present invention, and Figure 15 shows the first waveform of the voice. Figures 16a to 16d are diagrams showing the hierarchy list, Figures 16a to d are diagrams showing the second hierarchy list of audio, Figure 17 is a flowchart showing the operation of the device for creating a standard audio pattern, and Figures 18a, b, and c are A diagram showing the principle of the learning registration method, FIG. 19 is a diagram showing the operation of the learning registration method, FIG. 20 is a flowchart showing the operation of the registration processing unit in the present invention, and FIG. 21 is an operation of the non-learning registration processing same as above. 22 is a circuit diagram of the same speech analysis section as above, and FIG. 23 is a flowchart showing vowels /a/, /i/, /
FIG. 24 is a flowchart showing the operation of the determination processing unit of the present invention; FIG. 25a is a diagram showing the articulation points of vowels; FIG.
and FIG. 26 is a diagram showing the frequency distribution of the first formant and the second formant. 5,6,1
3 and 14 are filter banks, 42 is a standard pattern memory, and 46 is a registration processing section.

Claims (1)

[Claims] 1. A filter extracts a low frequency component where the energy of voiced sounds is concentrated and a high frequency component where the energy of unvoiced sounds is concentrated from the audio input, and the difference signal output of the signals extracted by the filter is When the value is equal to or greater than the first reference value, the first code is assigned, when it is equal to or less than the second reference value, the second code is assigned, and when it is equal to or less than the first reference value and equal to or greater than the second reference value, the third code is assigned. By doing so, an input pattern consisting of a time series of the first, second, and third codes is created for the input voice message, and this input pattern is compared with a plurality of pre-recorded standard patterns. In a voice message identification method that identifies a standard pattern having the minimum distance from an input pattern as an input message, an input pattern that occurs with the highest probability when the same voice message is input multiple times in different ways is used as a basic pattern,
An input pattern that occurs with a lower probability than the basic pattern is defined as a derived pattern, a portion where this derived pattern does not match the basic pattern is stored in advance as a branch pattern that branches from the basic pattern, and the input pattern is compared with the basic pattern. 1. A voice message identification method comprising: providing a standard pattern capable of branch matching processing in which a derived pattern generated by a combination of a basic pattern and a branch pattern is matched with an input pattern when the input pattern is not coded. 2 The difference signal output between a filter that extracts low frequency components of 1KHz or less, where the energy of voiced sounds is concentrated, and a filter that extracts high frequency components of 2KHz to 12KHz, where the energy of unvoiced sounds is concentrated, is used to distinguish voiced, unvoiced, and silent sounds. 2. The voice message identification method according to claim 1, wherein an input pattern consisting of a time series of three types of codes is created and compared with a standard pattern. 3 In the voiced sound section, the energy of high-frequency voiced sounds such as the vowel /a/ is concentrated between 500Hz and 1KHz.
The signal output of the difference between the filter that extracts the component of 500 Hz and the filter that extracts the component of 500 Hz or less where the energy of low-frequency voiced sounds such as vowel /i/ is concentrated allows high-frequency voiced sounds, low-range voiced sounds, and medium 3. The voice message identification method according to claim 2, wherein an input pattern consisting of a time series of three types of codes of voiced sounds is created and compared with a standard pattern. 4. If the time series of voiced sounds, unvoiced sounds, and silence that make up the basic pattern includes unvoiced sounds with short durations, a branch pattern that replaces the silent parts with silence is added to the basic pattern. 3. A voice message identification method according to claim 2, characterized in that: 5 If there is a time series of transitions from high-range voiced sounds to low-range voiced sounds in the time series of high-range voiced sounds, low-range voiced sounds, and mid-range voiced sounds that make up the basic pattern of the voiced sound section. The basic pattern is a branching pattern that can replace the time series with either a time series that transitions from high-range voiced sounds to mid-range voiced sounds or a time series that transitions from mid-range voiced sounds to low-range voiced sounds. 4. The voice message identification method according to claim 3, further comprising the step of: 6 If there is a time series of transitions from low-range voiced sounds to high-range voiced sounds in the time series of high-range voiced sounds, low-range voiced sounds, and mid-range voiced sounds that make up the basic pattern of the voiced sound section. The basic pattern is a branching pattern that can replace the time series with either a time series that transitions from low-range voiced sounds to mid-range voiced sounds or a time series that transitions from mid-range voiced sounds to high-range voiced sounds. 4. The voice message identification method according to claim 3, further comprising the step of: 7. If a high-range voiced sound is included in the time series of high-range voiced sounds, low-range voiced sounds, and mid-range voiced sounds that make up the basic pattern of a voiced sound section, the low-range voiced sound is Claim 3, characterized in that the basic pattern is provided with a branching pattern that creates a time series that is added before, after, or both before and after a high-frequency voiced sound. Voice message identification method. 8 In the time series of voiced sounds, unvoiced sounds, and silence that make up the basic pattern, if there is a short voiced sound sandwiched between unvoiced sounds and unvoiced sounds, or if there is a short voiced sound sandwiched between voiceless sounds and silence, 3. The voice message identification method according to claim 2, further comprising adding to the basic pattern a branching pattern that replaces the voiced part with an unvoiced sound. 9 When the difference signal output of a filter that extracts different frequency components from the audio input is equal to or greater than the first reference value, the first sign is assigned, when it is equal to or less than the second reference value, the second symbol is assigned, and when the difference signal output is equal to or less than the first reference value, When the value is equal to or higher than the second reference value, a third code is assigned to each of them, thereby creating an input pattern consisting of a time series of each of the first, second, and third codes for the input voice message. In a voice message identification method that compares an input pattern with multiple kinds of pre-recorded standard patterns and identifies the standard pattern with the minimum distance from the input pattern as the input message, when the same voice message is detected during the recording operation of the standard pattern. A plurality of input patterns formed by inputting multiple times in different manners are time-divided at fixed time intervals, and among the time-divided sections, for the section where the sign does not change, the code that does not change is is assigned as a core pattern, a third code is assigned to the section where the sign fluctuates, a basic pattern is created using this third code and the core pattern, and a For intervals where the code cannot become a code, a branch pattern consisting of the second code is added to the basic pattern, and for an interval where the code cannot become the second code, a branch pattern consisting of the first code is added to the basic pattern. The basic pattern and the branch pattern are recorded in advance as a standard pattern, and if the input pattern is not coded by comparing it with the basic pattern, the derived pattern generated by the combination of the basic pattern and the branch pattern is recorded as the input pattern. A voice message identification method characterized by performing branch processing such as verification.