JPH02720B2 - - Google Patents

Description

[Detailed description of the invention]

[Technical Field] The present invention relates to a voice message identification method for operating electronic equipment using voice messages. [Background Art] Figures 1 to 4 are materials showing the characteristics of the five Japanese vowels. First, speech corresponds to phoneme, and the first
It has a frequency spectrum envelope as shown in the figure,
By analyzing the frequency spectrum of the voice, the formant frequencies corresponding to the peaks of this spectral envelope are found, starting from the lowest frequency, the first formant F ₁ ,
If the second formants are represented in order as F ₂ , . . . , the five vowels can be represented by a change curve from F ₁ to _{F 4} as shown in FIG. Figure 3 shows the distribution of this formant frequency investigated on the F ₁ and F ₂ coordinate axes. It is said that in order to linearly identify the five Japanese vowels, as can be seen from Figure 3, the first to third formants must be determined correctly. We use only a few limited words or
In order to create a low-cost device for identifying sentences, we considered a method of symbolizing them into onomatopoeias that roughly, if not completely, resemble five vowels. FIG. 5 shows a schematic configuration of a conventional voice message identification device developed by the present inventors. In the figure, V indicates the short-time average power in a frequency band of 0 to 1 KHz during voice input, and corresponds to the energy of voiced sound. Also, U is inputting voice, 5
It shows the short-term average power in the frequency band ~12KHz, which corresponds to the energy of unvoiced sounds. Furthermore, VL, VH, VB, and VF are each inputting audio,
0~0.4KHz, 0.4~0.8KHz, 0.8~1.8KHz and
It shows the short-term average power in the frequency band of 1.8-3.2KHz, which corresponds to the energy of narrow jaw sounds, wide jaw sounds, back tongue sounds, and front tongue sounds, respectively.
S ₀ to _{S 3} are differential amplification means, each of which outputs a difference signal.
It calculates UV/V, VH/VL, VB/VL and VF/VB. The output of the differential amplification means _S0 is the reference value Rv, Ru (Rv<O<Ru) in the comparison means _C0 .
If the difference signal output is smaller than the reference value Rv, it is determined to be a voiced sound V. Also, if the above signal output is larger than the reference value Ru, the unvoiced sound U
If it is between the reference values Ru and Rv, it is determined that there is no sound S. Then, for each case of silence, voiced sound, and unvoiced sound, one of the codes S, V, and U is input to the encoding processing unit _MY0 . Further, MC ₀ is a matrix calculation unit which receives the outputs of the respective differential amplification means S ₁ to _{S 3} as input, and this matrix calculation unit MC ₀ receives each difference signal output VH/VL, VB/VL,
The three-dimensional vector whose components are
Short-term average power of o/Va, Vi, Vu, Ve, Vo
and short-term average powers Vh, Vl, Vf, Vb, and Vw of wide-jawed voiced sounds, narrow-jawed voiced sounds, front-tongue voiced sounds, back-tongue voiced sounds, and voiced sounds intermediate between vowels /a/ and /o/. It is used to calculate. The output of the matrix calculation unit MC ₀ is input to the maximum value determination unit MX ₀ , and each component Va,
It is determined which of Vi, Vu, Ve, Vo, Vh, Vl, Vf, and Vw is the largest component, and the code of the largest component is input to the encoding processing unit _MY0 .
When the code output from the comparison means _C0 is V, the symbolization processing unit _MY0 converts the maximum value determination unit _MX0
Va, Vi, Vu, Ve, Vo, Vh, output from
It outputs one of the codes Vl, Vf, Vb, and Vw, and when the code output from the comparing means _C0 is U or S, it outputs that code as is. This symbolization processing part
The composite code output from MY ₀ is input and stored in the standard pattern storage unit when registering a voice message, is input and stored in the input pattern storage unit when recognizing a voice message, and is input and stored in the input pattern storage unit when a voice message is recognized. Among the standard patterns, the standard pattern that is most similar to the input pattern is identified as the input message. By the way, in the conventional example shown in FIG. 5, when the power balance between VH and VL is adjusted, the five vowels {e, a, o} are on the positive side and {i, u} are on the negative side, with the zero point of the difference signal as the boundary. is located and therefore VH/VL
The difference signal becomes a signal called a Veao/Viu difference signal. Next, when adjusting the power balance of VF and VB, 5
Vowels {i, e}, 5 vowels {a, o, u} on the negative side
is located, and the VF/VB difference signal becomes a signal called the Vie/Vaou difference signal. On the other hand, when the balance of the VB/VL difference signal is adjusted, with the zero point of the difference signal as the boundary, the 5th vowel {a} is located on the positive side and the 5th vowel {o} is located on the negative side, so that the VB/VL difference signal is The VL difference signal is
This becomes a signal that can be called a Va/Vo difference signal. Figure 6 a and b show each voiced sound in the conventional example in Figure 5.
This is a diagram showing the frequency characteristics of a filter used to take out the short-term average power of VH, VL, VF, and VB. Figure a shows the frequency on the horizontal axis as a uniform scale, and Figure b shows the frequency on the horizontal axis. Frequency is plotted on a logarithmic scale. In FIG. 6, AP indicates the characteristics of the adjustment amplifier, which will be described later. Figure 7 shows the matrix calculation section in the conventional example shown in Figure 5.
This shows another means for realizing the same function as MC ₀ and maximum value determination unit MX ₀ . This figure 7 is
The onomatopoeia discrimination flow is shown when the levels of the difference signals Veao/Viu, Va/Vo, and Vie/Vaou are expressed as three values: high level (H), medium level (M), and low level (L). There is. In the flow shown in the figure, first, in the first step, discrimination is made using the Veao/Viu difference signal corresponding to the first formant F ₁ , in the second step, discrimination is made using the Vie/Vaou difference signal corresponding to the second formant F ₂ , and then By determining the Va/Vo difference signal in three stages, {i, e, a, o, u, h, l, f,
It symbolizes voiced sounds into 11 types: b, w, m}. 8 to 11 are the conventional example shown in FIG. 5,
5 vowels /i/, /e/, /a/, /o/, /u/
3 shows the output signal waveforms of the differential amplification means S ₀ to _{S 3} when the input voltage is input. In each of the above figures,
The U/V signal, H/L signal, F/B signal, and A/O signal indicate the outputs of the differential amplification means _S0 to _S3, respectively. SYM also shows the distinction between each voiced sound,
For example, in FIG. 8, l, i, f, e, . . . indicate voiced sounds Vl, Vi, Vf, Ve, . . . , respectively. However, the code m is each voiced sound Va,
A voiced sound Vm that does not fall under any of Vi, Vu, Ve, Vo, Vh, Vl, Vf, Vb, and Vw is shown. However, FIGS. 8 and 9 show examples of measurements taken on two different male subjects.
FIGS. 10 and 11 show examples of measurements performed on two female subjects. If you look at each of the above figures, you can see that almost the same features are extracted regardless of the speaker, but the vowel /e/ sound and /u/ sound /
The symbolization of the o/ sound requires a lot of know-how in filter adjustment, and the symbolization is somewhat incomplete. By the way, the first formant F ₁ of the vowel mentioned above is
Comparing Figure 3, which shows the distribution of the second formant F ₂ , and Figure 4, which shows the articulatory position of the tongue, the first formant F ₁ changes to a high frequency like /a/ when the jaw is wide open. and the jaw opens narrowly/
It can be seen that since the frequency is low like i/, it corresponds to the wideness and narrowness of the jaw. On the other hand, it can be seen that the second formant F ₂ also approximately corresponds to the front and back positions of the tongue. Furthermore, looking at Figures 2 and 3, it can be seen that there is a large variation in the second formant of vowels between men and women. However, in the conventional example, such second formant is converted to VF/
Since it is separated only by the VB difference signal,
In particular, there was a problem that the vowels {e} and {o, u} were incompletely separated. In other words, in the VF/VB difference signal, among the five vowels in Figures 8 to 11, if the part corresponding to /u/ is on the negative side, it would be detected to be large, but this is weak, and The second half of the part corresponding to /o/ is also found to be weak, and this is a factor that makes the symbolization uncertain. In order to eliminate such imperfections, we have proposed a method in which the deviation of the difference signal when the five vowels are uttered is determined and corrected as an offset for each individual, but it is still better to avoid such offset adjustment as much as possible. However, it goes without saying that it is desirable that the amount of offset be small. Nevertheless, in the conventional example, as mentioned above, since the second formant with large variations is separated only by the VF/VB difference signal, the offset of the zero point correction of the difference signal output of the filter pair becomes quite large. In some cases, there was a risk that complete correction might not be possible. In addition, if such zero point correction of the difference signal output is not performed, there is a drawback that the number of identified words decreases due to the difference between the actual utterance and the onomatopoeic symbol. For this reason, for use by unspecified speakers, there are significant limitations in terms of the number of components, etc. [Object of the Invention] The present invention has been made in view of the above-mentioned points, and makes it possible to reliably extract the characteristics of the second formant of a vowel, thereby enabling a more complete symbolization of five vowels. It is an object of the present invention to provide a voice message identification system that allows a speaker to reduce the amount of zero point correction of the difference signal output of a pair of filters. [Disclosure of the Invention] FIG. 12 is a so-called claim correspondence diagram showing the structure described in claim 1 of the present invention in functional blocks. In the same figure,
Fv is a filter that extracts the short-term average power of the low frequency component of the audio input, and Fu is a filter that extracts the high frequency component of the audio input. Each filter
The outputs of Fv and Fu are input to the differential amplification means _S0 ,
The difference signal component is extracted. C ₀ is a means of comparison,
When the difference signal component output from the differential amplification means _S0 is smaller than the reference value Rv, it is assigned the sign of a voiced sound V, when it is larger than the reference value Ru, it is assigned the sign of an unvoiced sound UV, and in other cases, is used to assign the sign of silence S. However, Ru>O>Rv. Next, FFa ₁ is a filter that extracts the short-term average power of voiced sounds with a narrow jaw opening (vowels i, u, etc.), and Fa ₂ is a filter that extracts the short-term average power of voiced sounds with a narrow jaw opening (vowels i, u, etc.). This filter extracts the short-term average power of jaw-voiced sounds (e, a, o, etc. of vowels). Next, Fb ₁ is a filter that extracts the short-term _average power of low sounds in the first formant, such as vowels e and o, among wide-jawed voiced sounds with a wide jaw opening. Among the wide-mouthed voiced sounds, the first one, like the vowel a,
This is a filter that extracts the short-term average power of high formant sounds. Next, Fc ₁ is the second formant of the low wide-mouthed voiced sounds like the vowel o.
Fc 2 is a filter that extracts the short-term average power of low formant sounds, and Fc ₂ also extracts the short-term average power of high second formant sounds such as the vowel e among low wide-mouthed first formant sounds. This is the filter to take out. Furthermore, Fd ₁ is a filter that extracts the short-term average power of a low sound in the second formant, such as the vowel u, among narrow-jawed voiced sounds with a narrow jaw opening, and Fd ₂ is a filter that extracts the short-term average power of the low sound in the second formant, such as the vowel u. This is a filter that extracts high short-term average power of 2 formants. S ₀ to _{S 4} are differential amplification means, which respectively output difference signals V/UV, Veao/Viu, Va/Veo,
This is to calculate Ve/Vo and Vi/Vu. The output of the differential amplification means S ₀ is the reference value in the comparison means C ₀
Rv and Ru (Rv<O<Ru), and if the difference signal output is smaller than the reference value Rv, it is determined to be a voiced sound V. In addition, if the above difference signal output is larger than the reference value Ru, it is determined to be an unvoiced sound U, and the reference value
If it is between Ru and Rv, it is determined that there is no sound S. Then, for each of silent, voiced, and unvoiced sounds, one of the S, V, and U codes is input to the symbolization processing unit _MY0 . MC ₀ is a matrix calculation unit that receives the outputs of each differential amplification means S ₁ to _{S 4} as input, and this matrix calculation unit MC ₀ receives each difference signal output.
By multiplying a 4-dimensional vector whose components are Veao/Viu, Va/Veo, Ve/Vo, and Vi/Vu by a predetermined matrix Tv, each vowel i, e, a included in the audio input is
This is to calculate the short-term average power of o and u. In the configuration shown in Figure 12, the wide-mouthed voiced sound VH
Differential amplification means for determining the ratio between
S ₅ and differential amplification means S ₆ for _{calculating the} ratio between the front voiced sound VF and _the back voiced sound VB are provided. The difference signal outputs VH/VL and VF/VB are multiplied by a predetermined matrix Tc to calculate the wide-mouthed voiced h, the narrow-mouthed voiced l, the frontal voiced f, and the backlogged voiced b included in the speech input. , and other wide-jaw and posterior tongue voiced sounds w. An example of the matrices Tv and Tc used in the matrix calculation unit MC ₀ is as follows. The output of the matrix calculation unit MC ₀ is input to the maximum value determination unit MX ₀ , and each component i, e, a, o, u, h,
It is determined which of l, f, b, and w is the largest component, and the code of the largest component is input to the encoding processing unit _MY0 . However, when the difference between the largest component and the second largest component is small, the code m is output. The symbolization processing unit MY ₀ is the comparison means C ₀
When the code output from _MX0 is V, i, e, a, o, u,
Any one of h, l, f, b, w and m
Furthermore, when the code output from the comparison means _C0 is U or S, that code is output as is. The composite code output from the symbolization processing unit MY ₀ is input and stored in the standard pattern storage unit when registering a voice message, is input and stored in the input pattern storage unit when recognizing a voice message, and is input and stored in the input pattern storage unit during verification processing. , the standard pattern that is most similar to the input pattern among the standard patterns registered in advance is identified as the input message. In the configuration shown in Figure 12, the VH/VL difference signal and the VF/VB difference signal are Veao/VB difference signals, respectively.
The Viu difference signal and the Ve/Vo difference signal may be used instead. FIG. 13 is a block diagram showing the configuration of an embodiment of the 8-filter system. In the configuration shown in FIG. 12 above,
Fv, Fu, Fa ₁ , Fa ₂ , Fb ₁ , Fb ₂ , Fc ₁ , Fc ₂ ,
A total of 10 filters, Ed ₁ and Fd ₂ , are required, but
In the configuration shown in FIG. 13, two of these filters are also used so that voice features can be extracted using eight filters. In Figure 13, VFh
is the high-frequency component of the front-voiced sound, VF is the component of the front-voiced sound, VB is the component of the back-voiced sound, VHh is the high-frequency component of the wide-jaw voiced sound, VHl is the low-frequency component of the wide-jaw voiced sound, VL
are the outputs of filters that extract the components of narrow-jaw voiced sounds. In the embodiment shown in FIG. 13, VL is shared between the Veao/Viu difference signal and the Va/Veo difference signal, and VB (or VL) is shared between the Ve/Vo difference signal and the Vi/Vu difference signal. be. This is because the zero point of the difference signal of a pair of filters corresponds to the intersection (crossover frequency) of the filter band, so even if the band of one filter is fixed among the pair of filters that take the difference signal, the band of the other filter is fixed. If there are two types of filter bands, the intersection of the filter bands will change. By the way, in the embodiment shown in FIG.
VHh and VB are almost the same, so one VB
I would like to summarize it as follows, but if we continue as shown in Figure 13, three difference signals will be taken out from VB, making it difficult to adjust the balance of the filter pair. Therefore, FIG. 14 shows an example in which VB is divided into a high frequency component VBh and a whole range component VB, and VF is combined into one, resulting in a 7-filter system. In this case, only two difference signals are extracted from VB, so filter balance adjustment becomes easy. From another perspective, the embodiment shown in FIG. 14 can be said to be the conventional example shown in FIG. 5 with VBh added. Figures 15a and 15b show the components VL, VH, and
This is a diagram showing the frequency characteristics of a filter used to take out the short-term average power of VB, VBh, and VF. Figure a shows the frequency on the horizontal axis as a uniform scale, and Figure b shows the frequency on the horizontal axis as a uniform scale. It is plotted on a logarithmic scale. In FIG. 15, AP indicates the characteristics of the adjustment amplifier, which will be described later. FIG. 16 shows an embodiment of the 6-filter system. That is, in the embodiment of FIG. 14 described above,
Even if VH is used instead of VBh, the vowels i and u can be distinguished, so the frequency component vector is
It can be composed of six components (six filters): UV, V, VF, VB, VH, and VL. From a different perspective, the embodiment shown in FIG. 16 is the conventional example shown in FIG. 5 in which a VF/VH difference signal is added, and almost the same filter band can be used. however,
The VF/VB difference signal is adjusted so that the vowels e and o can be reliably distinguished. Figures 17a and b show the first
In the 6-filter method shown in Figure 6, the components of each voiced sound
This is a diagram showing the frequency characteristics of a filter used to take out the short-term average power of VL, VH, VB, and VF. Figure a shows the frequency on the horizontal axis as a uniform scale, and Figure b shows the frequency on the horizontal axis. Frequency is plotted on a logarithmic scale. In FIG. 15, AP indicates the characteristics of the adjustment amplifier, which will be described later. The transformation matrix Tm of the matrix calculation unit MC ₀ in the 8-filter method of the embodiment shown in FIG. 13, the 7-filter method shown in the embodiment shown in FIG. 14, and the 6-filter method shown in the embodiment shown in FIG. Available for use. First, the transformation matrix Tm in the formula is set to 0 except for the minimum necessary elements for identification, so that calculations can be made faster. When the signal detection is weak, it is possible to widely symbolize the 5 vowels, and the formula gives the same weight (absolute value 14) to all the 5 vowel elements for the difference signal regarding the _first formant ), and regarding the two difference signals regarding the second formant F ₂ , one of the five vowels is given the necessary weight for discrimination, and the first formant F ₁ is converted to the second formant F 1 .
It can be said that it was given more importance than _F2 . This transformation matrix
Tm can be arbitrarily set depending on the word to be identified. In particular, the components of the symbol vector are divided into five vowels {i, e,
a, o, u} only, as shown in Figure 12.
Since the matrix element corresponding to Tv can be any one of {+1, 0, -1}, simple symbolization is possible without the need for multiplication and only by code conversion. On the other hand, the 12th
Code corresponding to Tc in the figure {h, l, f, b, w}
The elements of the transformation matrix are such that either the norm of the row vector of this matrix is the same as the norm of the row vector of Tv, or the value of the norm of the row vector of Tc is less than the value of the norm of the row vector of Tv, and Tv be larger than the absolute value of the elements of the matrix. If this is not done, other voiced sound components {h, l, f, b,
w} becomes smaller. Next, more specific examples will be described. FIG. 18 shows an embodiment of the 7-filter method shown in FIG. 14, and FIG. 19 shows an embodiment of the 6-filter method shown in FIG. 16. The difference between the two is the filter F _B h. It is only the presence or absence of. In each of the embodiments described above, audio is input from a microphone 1, amplified by a preamplifier 2, and adjusted for gain and offset by an adjustment amplifier 3. Next, the level adjuster 5 balances the input powers of the V/UV difference signal, other difference signals, and other difference signals. (in general,
Other difference signals are emphasized more than the V/UV difference signal. ) Next, the V/UV balance adjuster 4 balances the input of the filter Fv and the input of the filter Fu. On the other hand, adjust the VB/VL balance adjuster 6 to the middle point, and use the VH/VL balance adjuster 7 to balance the inputs of filter F _H and filter F _L.
The VF/VB balance adjuster 8 balances filter F _F and filter F _B (F _B h). Next, VB/VL
Balance VB and VL using balance adjuster 6. In the configuration shown in FIG. 19, when the VB/VL balance adjuster 6 is adjusted, VF/VH are also balanced at the same time. The output of each filter is sequentially switched by a multiplexer 9, the power is converted to a logarithmic scale by a logarithmic converter 10, and the output is digitized into an 8-bit binary number by an A/D converter 11. Note that when each filter is configured with a digital filter,
The A/D converter 11 is located at the next stage of the adjustment amplifier 3, and performs calculations for each filter in sequence in a pipeline manner, so that the output of each filter is calculated in sequence like a kind of multiplexer 9. next,
The difference between these digital values is calculated, and the difference signal vector {UV/
V, Veao/Viu, Va/Veo, Ve/Vo, Vi/
Calculate the five components of Vu}. FIGS. 20 to 23 show the results of voice feature extraction for the embodiment shown in FIG. 18 by inputting the same voices as those in FIGS. 8 to 11 from a recording tape. , and FIGS. 24 to 27
The figure shows the results obtained by inputting the same sounds as those shown in FIGS. 8 to 11 using a recording tape and extracting the features of the sounds in the embodiment shown in FIG. 19. In these FIGS. 20 to 27, the VF/VB difference signal of the conventional example becomes two, the Ve/Vo difference signal and the Vi/Vu difference signal, and the conventional example
The VA/VU difference signal has become a Va/Veo difference signal (abbreviated as a/o in the figure). Furthermore, the 20th
In Figures to Figure 27, a/i is Veao/
It shows the Vin difference signal. However, in the conventional example,
The detection of e, u, and o by the VF/VB difference signal became weaker as the difference signal approached the zero point, and symbolizing e, u, and o was difficult compared to i and a. In Figure 27, e and o can be detected reliably with the Ve/Vo difference signal, and u can be reliably detected with the Vi/Vu difference signal, so the symbolization of the five vowels can be performed more reliably. Recognize. In particular, in FIGS. 24 to 27, the Vi/Vu difference signal distinguishes i and u more clearly than in FIGS. 20 to 23, and as far as the example is concerned, the Vi/Vu difference signal in FIG. can be said to more reliably symbolize the five vowels. Next, returning to FIGS. 18 and 19, V/
In the UV judgment unit 13, when the V/UV difference signal is more positive than a certain setting value RU, it is judged as UV, and a certain setting value R _V
When it is more negative, it is determined to be V, and when it is in between, it is determined to be S. The start and end detection unit 14 detects the start of audio by determining V or UV, and detects the end if silence continues for a number of samples equal to or greater than a certain set value. The symbol vector conversion unit 15 performs matrix operations to convert the symbols into symbol vectors {i, e, a, o, u, h, l, f, b, w}, as shown in FIGS. 14 and 16. However, matrix operations are performed only in the interval of V. In the interval V, the symbolization processing unit 16 outputs the symbol if the maximum component of the symbol vector is greater than or equal to a certain set value, and outputs m if it is less than or equal to the set value. Furthermore, in the UV and S sections, UV and S are output, respectively. The formatting processing unit 17 converts the repetition of the same symbol into a list of one symbol and its duration, and further converts the repetition of the same symbol into a list of one symbol and its duration.If the duration is less than a certain setting value, if the symbols before and after are the same, these are combined into one list. If the symbols before and after are different, include them in the previous symbol, and omit those with short durations. The time axis linear normalization processing unit 18 normalizes the duration so that the total duration of each list becomes a constant value such as 200 (or 1000). This means that the total sample value is 200 (or
1000) and the duration time by multiplying each duration time. At this time, since the number of lists is small (10 to 20), it does not take much time to multiply and divide. Through the above process, the voice pattern of this method is
Can be created. This voice pattern is registered in the standard pattern storage section 19 in the registration mode. In the recognition mode, the distance calculation unit 20 matches the pattern with a standard pattern, but first, the number of UVs is used for primary identification to limit the matching target. Next, distances (correlation values) between corresponding symbols on the time axis are determined using the distance table 21, and the sum of these values for all samples is used as the distance between patterns. As the distance table 21, one shown in Table 1 is used.

[Table] In Table 1, the horizontal and vertical columns correspond to the code of the standard pattern and the code of the input pattern, respectively. For example, if the code of the standard pattern is a
, and when the sign of the input pattern is also a, the output of the distance table 21 is 2,
This indicates that the degree of approximation is high. Further, when the code of the standard pattern is UV and the code of the input pattern is a, the output of the distance table 21 is -2, indicating that the degree of approximation is low. Therefore, in the distance calculation unit 20, the degree of approximation of the input pattern and the standard pattern as a whole can be easily calculated by simply adding the outputs from the distance table 21 in sequence. . The significance test unit 22 considers that the closest pattern and the input pattern are the same if the pattern with the closest distance is closer than a certain setting value and is further away from the second closest one by more than a certain setting value,
In other cases, it is rejected as recognition failure. The recognition result is output from the identification result output section 23. Next, FIG. 28 is a so-called claim correspondence diagram that functionally shows the structure of the combined invention described in claim 2 in blocks, and FIG. 29 shows the structure of FIG. 28. FIG. 2 is a block diagram showing the configuration of a more specific embodiment. In each of the above figures, S ₀ , S ₁ , S ₂ , S ₃ , and S ₄ are UV/V
difference signal, Veao/Viu difference signal, Va/Veo difference signal,
This is a differential amplification means for extracting a Ve/Vo difference signal and a Vi/Vu difference signal. The outputs of the differential amplification means S ₀ to _{S 4} are compared with predetermined reference levels in comparators 24 to 33, respectively, and assigned different codes depending on the magnitude relationship with each reference level. First, the comparators 24 and 25 assign a code UV when the output of the differential amplifying means _S0 is above a positive constant value, and assign a sign V when the output is below a negative constant value.
, and in other cases, the code S is assigned. Next, the comparators 26 and 27 assign the code Veao when the output of the differential amplifying means _S1 is above a certain positive value, the sign Viu when it is below a certain negative value, and the sign Viu in other cases. S is assigned. Also, the comparator 28,
29 assigns a code Va when the output of the differential amplification means _S2 is above a certain positive value, assigns a code Veo when it is below a certain negative value, and assigns a code S in other cases. . Next, the comparators 30 and 31 assign a sign Ve when the output of the differential amplification means _S3 is equal to or higher than a certain positive value,
When the value is less than a certain negative value, the code Vo is assigned, and in other cases, the code S is assigned. Furthermore, the comparators 32 and 33 are differential amplification means.
When the output of S ₄ is above a certain positive value, the sign is
Vi is assigned, and when it is less than a certain negative value, a code Vu is assigned, and in other cases, a code S is assigned. The outputs of the comparators 26 to 33 are temporarily stored in the input bit pattern register 34, and are converted into codes a, e, o by referring to the symbolization table 36 in the V symbolization processing section 35, as in the case of FIG. , i, u, h, l, f, b,
The code is converted to one of w and m. An example of the symbolization table 36 is shown in Table 2.

〔Effect of the invention〕

The present invention is configured as described above, and from the audio input, wide-jawed voiced sounds with a wide jaw opening such as the vowels A, E, and O, and narrow jaw openings such as the vowels A and U are obtained. A first filter pair that calculates the ratio between narrow-jaw voiced sounds, high-pitched first formants such as the vowel A, and first formant sounds such as the vowels E and O among the wide-jaw voiced sounds.
a second filter pair for determining the ratio between a low formant sound and a high sound of the second formant, such as the phonetic E among the low wide-mouthed sounds of the first formant;
a third filter pair for determining the ratio between a low sound in the second formant, such as the vowel o, and a high sound in the second formant, such as the vowel i among narrowly voiced sounds;
A fourth filter pair is provided to determine the ratio of the second formant to a low sound such as the vowel u, and speech characteristics are extracted from the difference signal output of the first to fourth filter pairs. Because there is, the second vowel
It has become possible to reliably extract formant features, and among the five Japanese vowels, it has become possible to reliably identify e, u, and o, which were previously incomplete, and to create more complete symbols for the five vowels. Furthermore, since the second formant is not extracted in an unreasonable manner as in the conventional method, the amount of zero point correction by the speaker of the difference signal output of the filter pair can be reduced. Furthermore, in the present invention,
A 4-dimensional vector whose components are the difference signal outputs of each of the above filter pairs is input, and this 4-dimensional vector is multiplied by a transformation matrix to obtain the short-term average power of the five Japanese vowels and other voiced sounds as each component. A matrix calculation unit for calculating a vector is provided, a maximum value determination unit is provided for outputting a sign corresponding to the maximum component of each component of the vector output from the matrix calculation unit,
Since the input pattern is formed by the output of the comparison means and the output of the maximum value determination section, the five vowels and Another advantage is that codes for other voiced sounds can be obtained, and the configuration of the device can be simplified. Furthermore, in the case of the combined invention, the difference signal outputs of the first to fourth filter pairs are compared with a plurality of reference values, and different codes are assigned to each one according to the magnitude relationship with the reference values. A voiced sound discriminating means is provided which allocates and outputs one of the five Japanese vowels and other voiced sound codes in accordance with all combinations of codes assigned to each filter pair. This has the effect of making it possible to discriminate between voiced sounds with a simple configuration and at high speed using a ROM table, etc., which increases the response speed when operating electronic equipment using voice messages. It also has the effect of being able to be constructed at low cost.

[Brief explanation of drawings]

Figure 1 is a diagram showing the standard spectrum of the five Japanese vowels, Figure 2 is a diagram showing gender differences in vowel formants, Figure 3 is a diagram showing the distribution of the first and second formants of vowels, Figure 4 is a diagram showing the relationship between the five vowels in Japanese and the position of the tongue; Figure 5 is a block diagram showing the configuration of a conventional example; Figures 6a and b are diagrams showing the frequency characteristics of the filter used in the same example; FIG. 7 is a flowchart showing the procedure of onomatopoeic symbolization processing in the conventional example, FIGS. 8 to 11 are explanatory diagrams of the same operation as above, and FIG. 12 is a claim correspondence block diagram showing the configuration that is the gist of the present invention. FIG. 13 is a block diagram of one embodiment of the present invention, FIG. 14 is a block diagram of another embodiment of the same as above, FIGS. 15 a and b are diagrams showing frequency characteristics of a filter used in the above, and FIG. Block diagram of still another embodiment same as above, FIGS. 17a and 17b
is a diagram showing the frequency characteristics of the filter used in the above,
FIG. 18 is a block diagram showing another embodiment of the same as above;
FIG. 19 is a block diagram showing still another embodiment same as the above, FIGS. 20 to 23 are diagrams explaining the operation of the embodiment in FIG. 18, and FIGS. 24 to 27 are diagrams explaining the operation of the embodiment in FIG. 19. Fig. 28 is a block diagram corresponding to claims showing the gist of the combined invention, Fig. 29
The figure is a block diagram of one embodiment of the same as above, and FIG. 30 is a flowchart showing the procedure of the onomatopoeic symbolization process of the same. Fv, Fu, Fa ₁ , Fa ₂ , Fb ₁ , Fb ₂ , Fc ₁ , Fc ₂ ,
Fd ₁ and Fd ₂ are filters, S ₀ to _{S 4} are differential amplification means,
C ₀ is a comparison means, MC ₀ is a matrix calculation unit, MX ₀ is a maximum value determination unit, 24 to 33 are comparators, and 36 is a symbolization table.

Claims (1)

[Claims] 1. The difference signal output of a pair of filters that extracts the short-term average power of the high-frequency component and low-frequency component of the audio input, respectively, is input, and when the high-frequency component is stronger, the sign of the unvoiced sound is taken as the signal of the low-frequency component. A comparison means is provided that outputs the sign of a voiced sound when the voiced sound is stronger, and the sign of silence when the high-frequency component and the low-frequency component are approximately the same. Broad jawed voiced sound and vowel i,
The first filter pair calculates the ratio between a narrow-jawed voiced sound with a narrow jaw opening such as ``U'', and a high-pitched first formant sound such as the vowel A among wide-jawed voiced sounds, and the vowel ``E'' and ``O''. A second filter pair that calculates the ratio between low sounds in the first formant, such as low wide-voiced sounds in the first formant, and high sounds in the second formant, such as the vowel E, and vowel O sounds. A third filter pair that calculates the ratio between the low sound of the second formant and the high sound of the second formant of narrow voiced sounds, such as the vowel i, and the low sound of the second formant, such as the vowel u. A fourth filter pair for calculating the ratio with the sound is provided, a four-dimensional vector whose components are the difference signal outputs of the first to fourth filter pairs is input, this four-dimensional vector is multiplied by a transformation matrix, and the Japanese word 5
A matrix calculation unit is provided that calculates a vector whose components are short-term average powers of vowels and other voiced sounds, and a maximum calculation unit that outputs a sign corresponding to the maximum component of each component of the vector output from the matrix calculation unit. A value judgment section is provided, an input pattern is formed by the output of the comparison means and the output of the maximum value judgment section, and this input pattern is compared with multiple types of standard patterns recorded in advance to determine the standard pattern closest to the input pattern. A voice message identification method characterized by identifying an input message as an input message. 2 Inputs the difference signal output of a pair of filters that extracts the short-term average power of the high-frequency component and low-frequency component of the audio input, and when the high-frequency component is stronger, the sign of unvoiced sound is input, and when the low-frequency component is stronger, the signal output is the signal of the unvoiced sound. A comparison means is provided that outputs a silent code when the high-frequency component and low-frequency component are approximately the same, and it is possible to compare the voiced sounds with wide-jawed voiced sounds such as the vowels A, E, and O from the audio input. , vowel i,
The first filter pair calculates the ratio between narrow-jawed voiced sounds with a narrow jaw opening such as U, the high-pitched sound of the first formant pair such as the vowel A among wide-jawed voiced sounds, and the vowel E, A second filter pair that calculates the ratio between the low sound of the first formant and the low sound of the first formant;
A third pair of filters calculates the ratio between high formant sounds and low sounds in the second formant, such as the vowel o, and
A fourth filter pair is provided to determine the ratio between a high formant sound and a low second formant sound such as the vowel u, and the difference signal output of the first to fourth filter pairs is calculated using a plurality of reference values. By comparison, different codes are assigned to each according to the magnitude relationship with this reference value, and the five Japanese vowels and other voiced sounds are determined according to all combinations of codes assigned to each filter pair. A voiced sound discriminating means is provided which assigns and outputs one of the codes, an input pattern is formed by the output of the voiced sound discriminating means and the output of the comparison means, and this input pattern is combined with a plurality of pre-recorded patterns. A voice message identification method characterized in that the standard pattern closest to the input pattern is identified as the input message by comparing it with a standard pattern of the species.