JPS5895400A - Voice message identification system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] The present invention relates to a voice message identification method for operating a controlled device using voice messages, and its purpose is to efficiently perform verification processing based on the structure of the voice. The purpose of the present invention is to provide a voice message identification method that allows the user to perform the following tasks.

The structure of the present invention will be explained below with reference to illustrated embodiments. 1st
The figure is a block diagram showing the schematic hardware configuration of the voice message identification device according to the present invention, and FIG. In Figure 1, the sound enters from the microphone +1), is emphasized in the high range by the preamplifier (2), is logarithmically converted by the LOG amplifier (3), and becomes an output proportional to the value of A.
Only the alternating current component is amplified by the C amplifier 14). Next, a low frequency filter bank (5) extracts a low frequency component (L KHz or less), and a high frequency filter bank (8) extracts a high frequency component (5 KHz/v12 KHz). The filter bank ill +61 is composed of a bandpass filter CF+) (F2), a rectifier circuit (D+XDg), and an averaging circuit (M+) (tti2) consisting of an integrating circuit for each band. The low frequency filter bank integrator circuit (Ml) has a time constant of about 5 to 10 m5ec, and the high frequency filter bank integrator circuit (M2) has a time constant of 1 /-"l tnwwr
It is set to a high degree of purity. Both feet 1 and 6) t
Each output of 6) is input to the differential amplifier (7), and the output, that is, the difference between the high frequency component and the low frequency component, has a time constant of 2077.
The products of about 1 equation are averaged by an average circuit consisting of a circuit: 8).

This average (analog signal voltage y+t) output from the circuit □, 8) affects the ratio of the voiced sound V and the unvoiced sound U contained in the input voice. Figures 2 (a) and 2 (b) show the analog 0/) signal waveform y (t) in response to the voice input example "Massage Fage set" of our voice-controlled massage chair, and in this example, The voiced sound ■ is negative, and the unvoiced sound U has a negative value. Add this signal voltage y (t) to the V judgment circuit (81 and 15 judgment circuit (lO),
(5) =7°1 Sampling is performed at regular intervals by the OI check CK from the Rn1 leak generation circuit Ql). Both judgment circuit 91 (1
0) are applied with reference voltages Rv and Ru, respectively,
At the time of sampling, if the analog signal voltage on the positive side is less than the pressure Ru, it is a voiced sound v1. If the negative side drawing * is less than the pressure RU, it is a voiceless sound U. If the value is between Rv and Ru, it is a voiceless sound S. It is determined that In FIGS. 2(a) and 2(b), ztt) indicates the VU judgment output signal obtained from the signal processing circuit 02), which is -1 for unvoiced sound U, 0 for silent sound S, and ztt) for voiced sound. It takes a value of +l for . In addition, Fig. 2 (a) - SLOG amplifier (3
) is used, and (b) in the same figure shows the case when LOG assist (
3) is removed, and as is clear from comparing the two, the input audio is filtered using filter bank i5.
) By performing LOG conversion before tel, it is possible to reliably identify voiced sounds V and unvoiced sounds U.

The output of the preamplifier (2) for the song that is then passed through the LOG amplifier (3) is, for example, a 0-0.5 KH2 bandpass filter (F
3), a filter bank 03) consisting of a rectifier circuit (Dl), an averaging circuit (M3), and 0.5-1. OK) Characteristic components in each frequency range are extracted by a filter bank (14) consisting of 12 bandpass filters (F4), (6) 2- rectifier circuit (D4), and averaging circuit (M). The pre-low frequency Vt which is the output of the filter bank (131).

and the high frequency sound VH which is the output of the filter bank (14) are compared by the differential an-j (15+), and are averaged by the average (hi circuit α).The output of the averaging circuit (+6) is O - It is input to the ML judgment circuit θ and the VH judgment circuit &g) which operate in synchronization with the CK, and sent to the averaging circuit.If the output voltage of @ is lower than the reference voltage RL, it is judged that the VL acid component, If it is higher than RH, it is determined to be a VH component, and if it is between RL and RH, it is determined to be a midrange sound VM.

The signal processing circuit (+!J) converts each component of VH, Vt, and VM into three values into outputs of +1.0 and -1, respectively.

Each output of the 1'Zaribatari circuit (1!I) is read into the CPU 121 via the I10 port, and the read data is once input to the A comparison is made to determine which of the plurality of stored standard patterns is closest to the stored standard pattern, and the controlled device (24) is controlled based on the result of the comparison and determination. These verification and judgment operations are carried out by a microcomputer having a CpU (21+TozuO CRAM ROM+1) and a working RAM+261.

Next, FIG. 3 is a block diagram showing the voice message recognition processing function of the voice message recognition apparatus according to the present invention. As shown in the figure, the voice message identification device is composed of an acoustic processing section 1281, a wave number analysis section (9), an encoding processing section (1281), and a verification determination section (31). Of these, the acoustic processing section (28) and the wave number analysis section (2
Regarding 9), it has the configuration as detailed in the explanation of FIG. tag) is the differential anj t? ) (
15) to the signal processing circuit <1'2+ (19) t. In addition, the circuit after the I10 port (inclination) is composed of the microcomputer η as mentioned above. Yes, when a voiced sound V or an unvoiced sound U is detected, it will be reset and start the voice input operation, and if silence S continues for a certain period of time (approximately 0.2 seconds) or more, the voice input operation will be stopped. The voice input operation is 5/-2 (Iff!=e
Sashizu ring cycle of c (cycle of 5m5ec in the example)
One of the judgment circuits (from the voiced sound V
, unvoiced U, and silent S and their durations are input and stored in the memory, and the other judgment circuit (
From 38), high range sound VH1 mid range sound VM, and low range front V
Each code of t, and its duration are input and stored in the memory.The signal once stored in the memory in this way is shaped by the waveform shaping processing section.

Figure 4 (a) shows the signal waveform of a song that is waveform-shaped by the waveform shaping processing section (sake), and Figure 4 (b) shows the signal waveform after the waveform shaping process. The waveform shaping process is performed on the voiced sound V as shown by arrow a in FIG. 4(a).
Alternatively, the first process in which the short duration of the unvoiced sound U is set as the silent S, and the arrow a in FIG. 4(a).

~asVI- As shown, the duration of the voiced sound ■ or the silent sound tJ is relatively long, but the duration of the next silent S is short, and the next code following the silent S is silent SL: I) Code of the song If it is the same, then the second process is to erase the silence S. FIG. 5 is a flowchart showing the procedure of waveform shaping processing in the waveform shaping processing section +3f9.

This footset creates new list-format data by processing the data stored in the menu in a list format with the code ■, Ul S, and its duration, and stores it in the menu again. It shows 'jo jiram for. 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 a reference value r1. The duration is the standard value T.

If the code is not greater than SK, then the code V or U is
Convert and write the code S and its duration in the menu as a formatted list. Also, if the duration is larger than the reference value (direction 'rI), the code V spanning U and its duration are transcribed as is as a formatted list.Then, the duration of the next code S is greater than the standard value T2. If it is shorter than the reference value T2, it is determined whether the next code of the code S is the same as the code before the code S. If not, the code S and its duration are determined. transcribe as is into the formatted list.If they are the same, add the sign S to +
The code is changed to U or V after 4fl, and the changed code and its duration are written into the memory as a formatted list. Furthermore, it is again determined whether or not 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. When the conversion process is completed, it is determined whether the last sign of the 11 lists after formatting is S, and the last If the code is S, that code S is removed from the formatted list, and it is again determined whether the last code is S. However, when the last code is no longer S, the series of waveform shaping processing operations is completed, and at this time, the signal that has been subjected to waveform shaping processing as shown in Fig. 4(b) is displayed as a code. It is stored in a list format of U, V, S and their durations. In addition, the other waveform shaping processing section (3
Similar arithmetic processing operations are performed in 6), including: (1) converting the shorter duration of code VH and code VL into code VM; and (2) converting code VH and code V
The short-duration code VM sandwiched between H and VH
and (2) converting the short-duration code VM sandwiched between the code VL and the code Vt into the code (2) L.

Next, FIG. 6 is a flowchart showing the processing operations of the composite encoding processing section (3'71) for combining the two lists formed by the waveform shaping processing section (processing) into one list. To explain its operation,
First, the symbols U and V formed by the waveform shaping processing unit 0 wheels
, S and its duration, the first sign is U
Or determine whether it is S. If the first code is U or S, that code U or S and its duration are transcribed exactly as is as a list of embodied codes. Also, when the first code is ■ instead of U or S, the waveform shaping processing unit (examines the list of codes VH, Vv, Vj formed by lines and their durations,
First, it is determined whether or not the code VH is included in the duration of the code V. If the code VH is included, the code VH and its duration are transferred to the memory as a list of codes. Also, regarding the symbols VM and VL, the symbol V
If it is included in the duration of the code VM, VL
and its duration are each transferred to memory as a composite code list.

However, as in this embodiment, the waveforms are calculated separately for the system of codes V, U, S and the system of codes V)I, VM%Vt. As shown in Figure 7(b), as shown in Figure 7(b), the method of performing composite-soto encoding after shaping processing and then performing waveform shaping (see Figure 7(a)) There are ways to do this. In this case, the symbols U, V, VH, VT.

Composite encoding is performed by performing logical operations as shown in Table 1 based on each logical logic.

Table 1 and FIG. 8 are flowcharts showing the processing operations of the hierarchization processing section (38). The layering processing unit 13~ extracts a first layer list and a second layer list from the composite code list so that the pattern matching in the subsequent matching/judgment section @1) can be performed step by step according to the structure of the voice. It is something to create. Of these, the 11th layer list is the same as the code list consisting only of codes U, V, S and their durations, so it is best to copy the output list of the waveform shaping processing unit (as is) or use the codes in the composite code list. VH, VM, and VL are all coded V
This can be easily obtained by replacing K.

Next, regarding the thread 2-layer list, the composite is also obtained by transcribing the code list consisting of the continuation code Vl (, VL, VM and its duration from the soto code list. Therefore, the first layer list Assuming that n codes V are included in the list, n second hierarchy lists can also be obtained.

Further, FIG. 9 is a flowchart showing the processing operation of the normalization processing unit @9) for normalizing the code duration time for the first layer list and the second layer list obtained as described above. Normalization processing unit (Shun is a first layer list consisting of codes U, V, S and their durations, and codes VH, V
For the n second layer lists consisting of M, VL and their durations, normalization is performed on the time axis so that the sum of the durations is constant. Table 2 shows the first layer list. For the list, the symbols V, tJ, S and their duration TK
and normalized duration T8, where @duration TK corresponds to the number of samples.

2nd If we clarify the normalization processing operation using the O-tilt shown in FIG.
The sake bottle ΣTK (-=16623) is determined, and the normalization series Ps=1000/ΣTK is determined from this. Next, for each code tJ, V, and S, the duration time Ts(i
) in order. 1 as follows
When the normalization operation of the m duration for the hierarchical list is completed, the same operation is performed for each of the n 12th floor j-lists to normalize the duration (this is the one that performs the processing).Table 3 (a) ~ (d) shows the second layer list created for the four codes V+/C included in the first layer list (see Table 2) and their normalized durations. Table 3 (a) Table 3 (b) Table 3 (c) Table 3 (d) Table 3 (a) is the first code V (duration 3415) of the first layer list shown in Table 2. First layer list corresponding to ■
1, and in Table 3 (b) to (d) below, the duration is 3621. 1437.2637, second layer lists v2 to V corresponding to each code ■ are shown.

If the duration is normalized as described above, the recognition rate can be increased because it is less affected by the speaking speed. The duration TKO data shown in Tables 2 and 3 is based on the voice message "Senaka wo Susare.

” is analyzed using a 50 μ type Sunz Link Bar; The number) is 1 in 100 of the value in the table.

Code U of the first layer list normalized as above
%V, S, and codes Vu, Vg, V of the second layer list
Lld, 3 M sign'lt, processing i''1 part +4 (ffiKoi1'' +1.0, -1. That is, the code V in the first layer list is +11 sign. U corresponds to -l, code S corresponds to 0,
Also, the code VH in the 21st layer list is +l, and the code V
The Mba01 code VL corresponds to -1, respectively. In this way, when the distance calculation matching unit (11) compares the contents of the standard Butter Shime Seven (421) with the contents of the first layer list and the second layer list, the calculation speed will be significantly increased. In other words, the distance calculation verification unit (41) can be
+1.0, -1 ternarized data stored in z and 31i! It is designed to calculate the cross-correlation coefficient with the data output from the r encoding processing unit, 4(2), but since there are only three types of data, +1.0 and -1, The cross-correlation coefficient can be calculated at extremely high speed by simple logical operations and addition/subtraction without requiring 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 are compared in magnitude in the judgment processing section (4θ), and the cross-correlation coefficients are The larger the pattern, the more similar the pattern is determined to be.

The cross-correlation coefficient here means that the change in the standard pattern with respect to the change at time t17) is f + (t), and the change in the value of input patterns such as the primary hierarchy list and the secondary hierarchy list is f t Lt. ), then λ- is given by the following equation. Figure 10 (a) (h) The change in the value of quasi-pattern with respect to the change in time t I(t) and the input l ~ turn , respectively, and as shown in the figure, f + (t) and f t (t) are +L
, o, -1 within 3j, so the product of both t+ tt) f i tt) is also +1.0.-
Therefore, the total cross-correlation coefficient W is very generous. Such cross-correlation (number of threads f, g(τ) is mic 0 cosipi 1-
In the case of calculation using the following equation, the cross-correlation coefficient f1□(τ) is multiplied by the standard Jheturn f, where the practical soil can be calculated satisfactorily by numerical calculations as shown in the following equation. (t) and input cutter yf, (
It is a function of the phase difference τ with respect to −τ), and takes a large 1-axis value at a certain specific phase difference τ.

Therefore, the distance calculation matching unit (41) finds the point where the cross-correlation coefficient f1 and τ) is maximum, calculates the maximum value for each standard pattern, and calculates the maximum value for each standard pattern. Finally, the determination processing unit (the standard closest to the input cut-turn by comparing the magnitude relationship in the state) determines the turn.

By the way, in the present invention, when comparing the code extracted from the voice message with the standard Nortern,
After separating the code pattern into a primary hierarchy list and a secondary hierarchy list and performing a comparison on the primary hierarchy list,
By performing verification on the secondary hierarchy list,
# Texture is carried out in stages, which means that the features corresponding to the macroscopic structure of the voice are extracted first, and then the features corresponding to the microscopic features of the voice are extracted. This is because voice recognition can be performed more efficiently and reliably. Figure 11 shows the characteristics of speech in a hierarchical manner.Sounds are first broadly classified into nail sounds V, which involve vocal cord movement, and unvoiced sounds U, which do not involve vocal cord movement. It is classified into voiced sounds with a wide jaw opening (/1/crew"j) and voiced sounds with a narrow jaw opening (/i/') ru"j). A voiced sound with a wide opening of the jaw corresponds to the above-mentioned high-frequency sound■, and the frequency of the first and second parts of the voice is relatively high.
The frequency band is mostly distributed in 500 Hz "l KHz. Also, the voiced sound with a narrow jaw opening corresponds to the above-mentioned pre-low range VL, and the frequency of the first formant of the voice is relatively low, and the frequency band is It is mostly distributed in 0-500H2.

Voiced sounds with a wide jaw opening include the vowels /a/, one/, /已/, etc., and voiced sounds with a narrow jaw opening include the vowel /a/, one/, /已/, etc.
These include i/, /e/, 10/, /u/, nasal consonants, and voiced consonants. Furthermore, the unvoiced sounds U include stationary unvoiced sounds, that is, unvoiced fricatives UF, and transient unvoiced sounds, that is, unvoiced plosive sounds UB.

However, in order to clearly recognize a voice message in one 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 following is a list of the features of such speech, starting from the macroscopic features.

l) Is it a voiced sound V or an unvoiced sound tJ? Such features are found in the low range of the voice frequency spectrum (I KHz).
5+ (5KHz "12KH")
It can be determined by whether there are many z).

2) If it is a voiced sound V, then the voiced sound VH (/
q/Girouzu) or a voiced sound with a narrow jaw opening Vt,
(/i/Jirouzu) Is it? This feature is due to the presence of high-frequency sounds VH (500H2" t) in the same wavenumber spectrum of voiced sounds.
KHz) is too much, or Vt before low frequency, (0/-500Hz
) can be judged by whether there are many.

3) If it is an unvoiced sound U, it is a voiceless fricative Uv or a voiceless plosive JUa, and 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 characteristic VH, VL, UB,
The time occupied by Ur, etc., or the proportion of the duration of the voice message. Such characteristics can be determined by referring to the duration in the first and second hierarchy lists mentioned above. Although it is possible to distinguish between high-toned vowels (corresponding to /U/ and 10/), it is possible to distinguish between high-toned vowels (corresponding to /U/ and 10/). The message is fully recognizable.

For example, 112 Figure 1/Voice human power for the Anma chair/5e
This is an example of the frequency spectrum of nakaosasuτe/, which was obtained by performing 20-order LPG analysis on the voice input sampled at 20 KH2, with 200It-J"j le (lof + l castle) as one frame, but it is unvoiced. It can be seen that / and / have a concentration of noise above 5 KHz, and voiced sounds have a peak of power below l KHz.Furthermore, the voiced sounds /3/ and 10/ have a power of 500. Hz −I Concentrate on KHz, /, /, /
u/ is 0/-50oV (z It can be seen that the K power is concentrated. Furthermore, it can be seen that for voiced sounds, the same spectrum is read several times (several tens n1->) corresponding to each phoneme.

413(a) shows changes in the voiced sound component V and the unvoiced sound component U for the same audio input as above, and 413
Figure (b) shows the high frequency component VH and low frequency component Vt of voiced sound.
, which shows the changes between , and Figure 13 (
In a), the part corresponding to the unvoiced sounds /s/, /ni/ shows U, /n a /, /a o /, /a/, /
The part corresponding to u r p / clearly shows V, and in Figure 1g1a (b) /n /, /Q
The part corresponding to a / is Vt, /a/, /ao/,
The part corresponding to /e/ is VH. Therefore, as described above, there is a primary hierarchy list corresponding to the voiced sound V, unvoiced sound U, and duck sound S, and a secondary hierarchy corresponding to the high-range sound VH1, middle-range sound VM, and low-range pre-Vt among the voiced sounds. If you compare the list with the pre-memorized semi-no \ turn, the spirit of Dai JL,
Voice messages are distinguishable.

However, the following is a generalization, and if the voice message is spoken by different speakers, or if the same speaker changes the speed of voice or changes the manner of utterance, the syllables may change. The quarrelsome vocalizations of the voice medium occur due to phenomena such as the inability to detect plosive sounds in the middle of the voice, or the voiced sounds in the syllables becoming voiceless sounds. A charming lji who can handle everything for you
It is necessary to create a quasi-no-turn. In this embodiment, even if the input pattern changes due to such subtle changes in vocalization, the voice message can be correctly interpreted. That is, in this example,
As a standard no\turn, for example, as shown in the fJ14 diagram, there is a branch no\turn such as C2g, C45 in the time series of C1, C2, C3, C5, C1, etc. is added, and the input cover pattern is coded C1.
It is also possible to check with the first derived Nortern from the I time series f1j of C2, C5, and C0, and the derived J~ turn of Sou2, which consists of the time series of codes CI, C2, C1, Cps, and C1. As a result, the performance rate of voice messages is increasing.

Hereinafter, the basic no.\turn, branching pattern, and the concept of derived noturn will be explained using specific examples.

A typical example of a life in which the input pattern of a voice fluctuates is the phenomenon of missing voiceless plosives /, /, /1/, and /ni/. That is, as can be seen from FIG. 113(a), the voiceless plosive /ni/ is a transient unvoiced sound, so its duration is short and it is very difficult to express. In comparison, the unvoiced fricative /s7 is a stationary unvoiced sound, so its duration is long as shown in FIG. 13(a), and it is easy to detect. For this purpose, if the sampling period is lengthened, voiceless fricative /! I/
Even if the voiceless plosive sound /ni/ can be emitted, it may not be possible to detect the voiceless plosive sound /ni/. Fig. 15 shows a standard pattern for the first layer list of ``Personal Power'' / 5enakaoaaakure / taking this point into consideration, and is designated by the symbol U.
, S, ■8, s%ty, s, v, , S, U, S, V
,,S,tJSS, In addition to the basic Nortern consisting of time series No. 7, the code S, TJ, which is sandwiched between codes v1 and v2,
A branch pattern consisting of a code S equal to the duration of S is provided.Therefore, the encoded input character \ turn of speech is not only matched with the coded \ turn, but also the branch J \ turn. It is also matched with the derived pattern consisting of the symbols U, S, ■S%v2, S, U, S, VN, S, U%S, Vt generated by the turn, and therefore the unvoiced plosive /ni/ is included. Even if something is missing from the cover turn, the voice message can be recognized correctly. In addition, the numbers written in Figure 415■,
■ is the result of analyzing the vocal patterns of the same speaker five times.
This shows that there were four cases in which the basic pattern was matched and one case in which the derived Nortern was matched. By configuring as above, the voiceless plosive /p/
, /1/, /ni/ to prevent the color from fading.

Next, Figure 16 (a) ” (c) is voice human power / 5en
For the four voiced sounds ■, ~■, included in akaosas -ure /, each second-order Li! An example of creating a standard pattern for a list is shown. To explain the fourth voiced sound vI among them, its basic pattern is as follows: VM, Vt, VM, VHlVM
lVL(7) It is composed of time series color, and furthermore, the code V
It has three branched norterns consisting of M. Therefore, in this case, the symbols VM, VL, VM, VL
The derived pattern of 4t consisting of the time series and the codes VM, V
Derivation of 2 from the time series of L, VM, VL, VM J\
The third turn consists of a time series with codes VM, VL, and VM1VH%VM. A derived pattern is formed. The numbers ■, ■, ■, etc. written in FIG. 116 (a) indicate the number of branches in the same way as sij. Note that the duration time of each branch norturn 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 VH, VM, and vL in the second layer list vary slightly.

By the way, the code V)I in such a Sui 2-layer list
, Vv, and Vt vary in a wide variety of ways. For example, in the case of (-OVL-VH series and EVL-VM series (e.g. /5enaka/'s /na/
), (q)Vt, -Vbt series force VL-Vt, when it becomes series (e.g. /re of /5aqu-re/), θ
9 A combination of VL-VM series and VL-VH series (e.g. /re/ of /tomare/), (:)VH
- When VL series becomes VH-VM series (example / IS
en-al, /of /lk/), DI + sign V) I is VL
-VH sequence (e.g. /sa/), ('J code Vu becomes VL -VH-vL sequence (e.g. /k
Examples include a ta /'s / ka t /). If we organize the rules for the fluctuations of these codes Vo, V & 4, Vt, they can be roughly classified into the following two cases: 1) Due to the interaction of the preceding and following phonemes, the codes Vu and VM and the codes L and VM Mutual replacement occurs. That is, the VH-VL series is the VH-Vv series or VM
- It can become a VL series. and the VL-VH sequence can become the Vt-Vv sequence or the VM-VH sequence.
The code vL is added before or after H or after the power, that is, the code V)I is Vt, -VH series, VH-Vt, series, t ft1i VL -Vn -
Vt, i-column 2 is swapped.

An example of a Noh performance in which the entrance and turn of the voice fluctuates is the devoicing of vowels. For example, for Japanese people, the word "watakushi" is /wa-talzusi.
Rather than those who pronounce it correctly as /, there are more people who skip the vowel /u/ and pronounce it as /watak*i/. This is because the vowel /u/ is sandwiched between a voiceless plosive /ni/ and a voiceless fricative /S/, and generally one vowel sandwiched between a voiceless plosive Us and a voiceless plosive Ua (for example, / klQpu /no/i/) and ■Voiceless plosive U
A vowel that is rare between a and the voiceless II sound tJF (for example, /U/ in /wataku@i /), and a vowel sandwiched between a voiceless sound and a voiced consonant, have a strong tendency to be devoiced. Strong against In addition, one vowel sandwiched between voiceless U and silent S (e.g. /d-ousa / of /a/
) also tends to become silent. Therefore, for a vowel that is rare in the - throw between a voiceless consonant, between a voiceless consonant and a voiceless consonant, and between a voiceless consonant and a voiced consonant, the basic pattern where the vowel part is the voiced consonant V is created. In addition, a branch pattern in which the vowel part is an unvoiced sound U is added to the standard no-turn, and when a specific vowel in E is clearly pronounced as a voiced sound V, it is possible to make a comparison judgment using the basic pattern. In addition, if a specific vowel in E is pronounced unclearly, as if it were a voiceless sound U, it is possible to increase the recognition rate of spoken messages by making it possible to compare and judge based on derived patterns. Next, a method of creating a standard pattern having such a basic no-turn and branch l~ turns will be described. There are two ancient methods for creating standard patterns.One is to input the individual phoneme codes and their durations that make up a voice message from a board, etc., and then use a branch processing program. One method is to automatically create basic patterns and branch patterns by Party B, and the other is to register the same voice message multiple times by changing the way it is uttered or by changing the speaker, and then creating basic patterns based on common characteristics. Patterns and branching out unique characteristics that are not common
~ It is a learning registration method that registers as a turn, the former is a deductive method and the latter is an inductive method.

First, the method of the +1f1 person is, for example, as shown in FIG.
/ =1 /, /ni/, /a/, 10/, /s/
, /a/, /9/, /++/, /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, and if it is a voiced sound ■, the vowel /q/ is entered. (The code Vt+ is assigned to the vowel /i/ and the voiced consonants /m/, /
For b/, the code Vt is assigned.

For the voiced consonants and vowels /e/, /u/, and 10/ (of the dragon), we create a branch J~ turn that can be any of the symbols VH, VM, and VL. Also, the symbol U is used for unvoiced sounds. , Furthermore, l assigns the code S to silence as is.Next, enter e, the duration time, and for unvoiced sounds with short duration, that is, voiceless plosives, in addition to the basic]\ turn consisting of the code tJ, A branch no\turn consisting of the symbol S is added.Additionally, a code sequence is input, and for a school vowel sandwiched between a voiceless sound and a voiceless sound or a voiced consonant, a basic no-turn pond consisting of the symbol V is added from the symbol U. By doing the following, it is possible to automatically create a standard pattern in which a branch pattern is added to the basic nortern, which increases the recognition rate of voice messages. be.

Next, the learning registration method will be explained. Figure 18(a)~
(c) is the code VH1VM, V corresponding to the second layer list.
This shows a case where a standard l\ turn consisting of L is created, and FIG. 19 is a footset showing the creation procedure. First, as shown in Fig. 18(a), the same word is registered multiple times, the normalized time is divided into lO regions, and the portion where the sign does not change in the same time region is taken as a core pattern.
Let VM be the part where the sign changes in the same time domain. At this point, three turns of learning basics as shown in FIG. 18(b) are created. Next, a VL branch pattern is added to the portion where vM or VLK are present in the same time domain. In addition, a branch nortern of Vu is created in a portion where VM or V)I becomes in the same time domain.

Furthermore, the part that becomes both VH and Vt in the same time domain is V
Leave it as M. at this point#! A learning standard nortern having a branching nortern as shown in FIG. J18 (c) is formed. The learning standard pattern obtained in this way is i
It is designed to be registered and stored in the semi-nohetanme elt+2.

Therefore, in this embodiment, a registration process S system is provided which is a compromise between the learning v19 system and the non-learning registration method, and a flowchart thereof is shown in FIG. First, it is determined whether or not there is an S-IJ-V series or an S-■ series in the 1st layer list of data entered manually by the catalog processing unit 1461, and if so, it is determined that the 5-TJ-V series is present. S-■
A quasi-notice is formed that includes any code sequence in the sequence. Next, regarding the second layer list, it is possible to switch between creating a standard norturn in the learning mode as shown in Figure 19 in column 19, and creating a standard pattern in the non-learning mode. If you are unable to perform well in the seventh mode, you can use the other mode. Since the operation in the learning mode has already been explained using the chart 70 in FIG. If the first code of the 21st stage list to be explained is Vt, the basic pattern ST
-VL and derived pattern ST -VM)\
Create a turn.

Also, if the first code is VH, the basic pattern ST -V
t, -VH no f1. K, two primary butters: / S
Create a standard I\turn containing T -VM -Vu and ST -Vt -VM. Furthermore, if the first code is VM, a standard pattern of only basic butter:/ST-VM is created. Next, determine whether the last code is VH, VL, or V-, and whether each code included between the first code and the last code is a VL -VH sequence or VH - Vt.
, a standard pattern with branch patterns and colors as shown in the flowchart of FIG. 21 is automatically formed depending on the series.

By the way, when creating a standard pattern for the second layer list in this manner, it is necessary to correctly identify the code VH and the code VL. As mentioned above, the code Vn corresponds to a high-frequency voiced sound (/a/ri1shi-zu), and the code VL corresponds to a low-range voiced sound (/i/ 'J loop). As shown in Fig. 22, the analyzer developed by the present inventors has a balance for adjusting the output balance of the VH analysis system and the VL analysis system, a variable resistance VR for adjustment, and a variable resistance VR for adjusting the offset. By providing a resistor VR2, when the vowel /a/ is uttered, the code VH is always emitted, and the vowel /i/
When this is uttered, the code vL is always emitted. However, strictly speaking, the optimal value of this balance may differ depending on the personality of the speaker. Therefore, the present inventors found that when the vowel /e/ is naturally produced, VH/
The inventors have discovered that it is sufficient to adjust the balance so that the VL difference signal becomes zero. Figure 23 shows the principle. As shown in the figure, the first formant of the vowel /a/ is distributed at 500Hz '' l KHz, and the vowel /i
The it formacite of / is in the 0-500Hz K'y+ range, but the 41t1 lumant of the vowel /e/ is roughly in the middle. Therefore, by adjusting the balance between vH and VL using the vowel /e/ as a reference, the optimum balance value can be obtained.

Finally, regarding the secondary hierarchy list, each code VH, VM, vL
A method of matching that takes into consideration the duration of time will be explained. FIG. 24 is a diagram showing three methods of collation and identification of the secondary hierarchy list, and the most suitable one is selected and used. First, the IT method uses multiple voiced sounds V+ included in one voice message.
” For vn, the code included in it with the most codes is checked to see if it is VH, VM, or Yt. In this case, the next most codes and the code with the fewest codes are The second method is to check whether the proportion of VH included in each voiced sound ■1~Vn matches the input J~ turn and the standard pattern. This is to check whether cancer 3
The method is to restore and compare the number of cases in which VM in the input turn matches the VH or VLK of the standard pattern and the number of cases in which the VIA in the standard pattern matches the Vii or VL of the human pattern. Therefore, all voiced sounds v in a voice message,
For /i, V, match the input cut\turn with multiple standard patterns using the most appropriate one of the three matching methods above, and select the standard norturn with the most matching features. This is what I am trying to judge.

The present invention is configured as described above, and in the stage of primary matching, the outputs of a pair of filters that extract high-frequency components and low-frequency components of voice input are compared, and a section in which the high-frequency component is stronger is determined as an unvoiced speech section. , the interval in which the low homowave component is stronger is the voiced interval, and the interval in which the high-frequency component and the low-frequency component are approximately the same is the silent interval, and the first time series of the unvoiced, voiced, and silent intervals is The input cover pattern is compared with multiple types of pre-recorded standard patterns, and in the secondary matching stage, the outputs of a pair of filters that extract high-frequency components and low-frequency components of the frequency band of voiced sounds are compared. The section where the high-frequency component is stronger is called a high-voiced section, and the section where the low-frequency component is stronger is called a low-frequency voiced section. The second time series consisting of the time series of each interval of
Since the input pattern of the input pattern is compared with multiple types of standard patterns that have been recorded in advance, we first narrow down the matching range by extracting the macroscopic characteristics of the voice, such as whether it is voiced, voiceless, or silent. By extracting the microscopic features of the voice called the frequency components in the voice and identifying the input message, it is possible to limit the matching range at the color and primary matching stage, making it possible to recognize voice messages. What is the advantage of being able to do it reliably and quickly? In particular, as mentioned in the explanation of the embodiment, at the stage of secondary verification,
If you make the high-frequency voiced sound correspond to the fJl formant of the vowel /a/ and the low-range voiced sound to the 4t hon71 of the vowel /V, you can change the voiced sound to a voiced sound with a wide jaw opening. It is possible to recognize voiced sounds and voiced sounds with a small jaw opening, and it is possible to perform matching processing that matches the structure of the speech, thereby further increasing the recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a schematic configuration of a voice message identification device according to the present invention, FIGS. 2(a) and 2(b) are operation waveform diagrams of the same, and FIG. 3 is a block diagram showing the voice message identification processing operation of the same. The diagram shown,! ! 84 (a) and (b) are waveform diagrams showing the operation of the waveform shaping processing section of the same as above, FIG. 5 is a flow chart showing the operation of the same waveform shaping processing section of the above, and FIG. FIGS. 7(a) and 7(b) are flowcharts showing the operation of the encoding processing section, FIG. 8 is a flowchart showing the operation of the layering processing section, and FIG. The figure is a flowchart showing the operation of the normalization processing section same as above, 4to diagrams (a) and (b) are waveform diagrams showing the operation of the distance calculation matching section, FIG. Figure 12 is a diagram showing the frequency spectrum of audio, Figure 13 (a) and (b) are waveform diagrams of signals extracted from audio, Figure 114 is a diagram showing the principle of branch matching processing in the Shihi device, and Figure 15. 16(a) and 16(d) are diagrams showing the second layer list of voices, and FIG. 18(a), (b), and (c) are diagrams showing the principle of the learning registration method, FIG. 19 is a diagram showing the operation of the learning registration method, and FIG. 20 is a diagram showing the operation of the registration processing section in the device of the present invention. Figure 121 is a flowchart showing the operation of the non-learning registration process as above, Figure 8g22 is a circuit diagram of the speech analysis section as above, and Figure 23 is a flowchart showing the operation of the non-learning registration process of the same as above. FIG. 24, which is a diagram showing the frequency distribution, is a flowchart showing the operation of the determination processing section in the present invention. Agent Patent Attorney Ishi 1) Chief 7 Figure 22 Nail 24; C Procedural amendment band (spontaneous) 1. Indication of case Patent Application No. 193557 of 1982 2. Name of invention Voice message identification method 3. Make amendments Person Relationship to the case Patent applicant Location 1048 Bold Kadoma, Kadoma City, Osaka Name (58)
3) Matsushita Electric Works Co., Ltd. Representative Yoshi Kamimae - 4, agent postal code 530 5゜ Date of amendment order Self-corrected application number Tokusho 56-193557 No. 1, page 25, line 5 of the specification of the present application I have corrected the entire sentence as follows. [There are things that can be done. Figure 25(a) shows the vowel /
a /, / i /, / u /, / e /, 10
The figure (b) shows the frequency distribution of the first formant and second formant of the vowel.
63-p364 (cited by Eli). Furthermore, FIG. 26 shows the distribution of the first formant and second formant of Japanese vowels for male and female voices, respectively. Fig. 25(b) As is clear from the distribution of the second formant shown in Fig. 26, it is approximately 0.8 to 1.8 KH.
By analyzing the outputs of the z bandpass filter and the approximately 1.8-3.2 KHz bandpass filter, the position of the second formasite can be detected, and this can be used to detect the position of the tongue before and after the tongue position. It is also possible to extract features. 2. Extract the most microscopic features. The following sentence will be inserted after "Mono deru." in the first line of page 40 of the same page. [Furthermore, check the degree of matching between the input cover pattern and the standard pattern. It is also possible to evaluate with the corresponding score +1, 0, -1 for every 7 pulls, and to judge based on 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. Scoring is performed 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. Therefore, if the total number of samples is 1000, the total score will be 1000 when the patterns match perfectly. $4 Table" 3. In the i++ of "deru." in the third line of the 43rd Tj above, [, Figure 25 (a) is a diagram showing the articulation points of vowels, Figure 25 (b)
) We have inserted "Figure 26, which shows the frequency distribution of the first and second formants." 4. Figures 25 and 26 will be added to the attached drawings as attached. Agent Patent Attorney Ishi 1) Chief Figure 25 (b) Figure 26

Claims (1)

[Claims] (1) The outputs of a pair of filters that extract high-frequency components and low-frequency components of audio input are compared, and the section where the high-frequency component is stronger is defined as an unvoiced section, and the section where the low-frequency component is stronger is determined as an unvoiced section. is a voiced sound interval, and a silent interval is an interval in which high-frequency components and low-frequency components are approximately the same. First-order matching with the standard J~Tashi is performed, and for voiced sound sections, the outputs of a pair of filters that extract high-frequency components and low-frequency components in the frequency band of the voiced sound are compared, and the section in which the high-frequency components are stronger is determined. is a high-frequency voiced interval, an interval in which the low-frequency component is stronger is a low-frequency voiced interval, and an interval in which the high-frequency component and low-frequency component are approximately the same is a mid-range voiced interval; A standard JS turn with the minimum distance from the input cover turn is determined by secondarily comparing the input batashi of the name 2, which consists of the time series of each section of the vocal sound and the voice of Nakagusuku, with multiple types of standard patterns recorded in advance. A voice message identification method characterized by identifying it as an input message. (2) When performing the secondary comparison between the 42 input cover patterns and the standard nortern, it is checked whether the durations of the high-frequency voiced sound sections match or not. Voice message identification method (3) When performing a secondary comparison between the second input cover turn and the standard pattern, the proportion of time occupied by the high-frequency voiced sound section in the total duration of the 20th input cover turn matches. The voice message identification method according to claim 11, characterized in that the voice message identification method (
4) When performing a secondary comparison between the second input Batashi and the standard pattern, determine whether the longest duration among the high-range voiced sound section, low-range voiced sound section, and mid-range voiced sound section matches. Claim 1, which is characterized by comparing whether or not
Voice message sorting method described in Section 1. (5) When comparing the second input pattern and the standard pattern, the order in which the high-frequency voiced section, low-frequency voiced section, and Nakagusukuga voice section are arranged in descending order of duration ′+! j. A voice message identification system according to claim 1. (6) Comparing the outputs of a filter that extracts low frequency components of 1 KHz or less, where the energy peaks of voiced sounds are concentrated, and a filter that extracts high frequency components of 2 KHz and 12 KHz, where the energy peaks of unvoiced sounds are concentrated. A voice media 1 sequence identification system according to claim 1, characterized in that each section of voiced sound, unvoiced sound, and silent sound is identified. (7) Voiced sound 500) 1z -IKH in which the energy of high-frequency voice sounds such as the vowel /a/ is concentrated.
By comparing the output of the filter that extracts the component of z and the output of the filter that extracts the component of 500T (500T) where the energy f of low-frequency voiced sounds such as the vowel /i/ is concentrated, it is possible to 2. The voice message identification method according to claim 1, wherein each section of a range-voiced sound and a mid-range voiced sound is identified.