JPS59192298A - 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 [Technical Field] The present invention relates to a voice message identification method for operating electronic equipment by voice messages.

[Background technology]

Figures 1 to 1 are negative phases showing the characteristics of the five Japanese vowels. The first consonant P corresponds to the phoneme, and by analyzing the frequency spectrum of the first voice as shown in Figure 1, the formant frequency corresponding to the heater of this spectrum envelope is reduced, and the lower frequency is determined. By representing the first formant F, the second formant F2, .

This formant frequency distribution 2F, examined on the F and II reference axes, is shown in Figure 5. As can be seen from Figure 5, in order to linearly identify the five Japanese vowels, %1~
It is said that this cannot be done unless the 5th formacite is determined correctly.

We investigated a method of symbolizing onomatopoeias that are approximately, if not completely, similar to five vowels, in order to produce a very small number of vulgar words or sentence words that can be identified at a low cost within the device.

FIG. 5 schematically shows the structure of a voice message identification device conventionally developed by the present inventors. In the same figure, ■ is the short poem in the voice human power, the frequency of 0 to IKHz, and the same half-equal power? It shows the energy of the past, published by μ.

During voice input, U shows a short-term average high τ in the frequency band of 5 to 12 kHz, and VB/VL, which is illusory with the energy of unvoiced sounds, oh! and V F /V B
Multiply the three-dimensional vector of kh1 by a predetermined matrix Tmk to obtain n valley vowels /a /, /i included in the voice input.
/, /u/, /e /, 10/ short-time average power V
a, Vi, Vu, V6, Vo and wide-jaw voiced, narrow-jaw;
gp sounds, old voiced sounds, retro voiced sounds, vowels /a/ and 10/
Each short-term average of intermediate voiced sounds vh, vz%
Vf, Vb%vw'z are calculated. The output of the matrix calculation unit M C0 is input to the maximum value determination unit M
z.

Which component is larger than i among vr and VW? > w IJ
The code of the largest component is input to the encoding processing unit Myo. When the code n output from the comparison means C0 is V, the symbolization processing unit MY0 converts the maximum value determination unit MX
v2.Vi% V u%V es
Vo, Vh, V/%Vf%Vb.

and Vw, and when it is 1R or S output from the comparative soft means CO, the code t is output. This symbolization processing unit MY071 is inputted and stored in the output standard pattern storage unit n, and when recognizing a drunken message, the input cover pattern is input and stored in the memory, and during the verification process, the nfC standard pattern registered in advance is Among these, the standard pattern that most closely approximates the input turn is identified as a human message.

By the way, in the conventional example shown in FIG.
Five vowels (e, a, o) on the positive side, (i
, u) is located, so the VH/VL difference 1a is V
This becomes a signal that can be called an eao/Viu difference signal. Next, V
By adjusting the power balance between F and VB, the difference signal will be at the border above the zero point, and five vowels (i, e) will be on the positive side and five vowels on the negative side.
Vowel (a, o, u) is located, V F/V B7fi
The signal el'UV1e/Vaou7i becomes a signal that should be called an ID signal. On the other hand, when the balance of the V13/VL Sho signal is adjusted, the 5th vowel ((
a), the fifth vowel (0) is located on the negative side, therefore VB
/VL difference multiplier becomes a signal called Va/Vo difference signal.

Figure 6 axis) (b) shows each voiced sound VH in the conventional example in Figure 5.
%VL, VF, VB6 average 8wtl between 9fill1
(t It is drawn as an accident-proof scale. In Fig. 6, AP is the characteristic of the adjustment amplifier described later. It shows.

FIG. 7 shows another means for realizing the same functions as the matrix calculation section MC6 and the maximum value MXo in the conventional example shown in FIG. This brown 7 figure is '? ! r difference signal Vea
The onomatopoeic order when the level of o/Vius Va/Vos Vie/Vaou is expressed in five values: high level (E-1), medium level (M), and low level (L)!
T! IJ separate foo-2 is shown. In the figure, F0- is 1
% At the first stage, Vea corresponding to the first formant F1
o/Viu difference signal 'i! In the second step, the Vie/Vaou difference signal corresponding to the second formasite F is separated, and in the third step, by the Va/Vo difference signal], (i + es a+ Oe u
-h, '+' + b+w, m) are used to encode voiced sounds into 11 types.

Figures 8 to 11 show the conventional example in Figure 5, 5 vowels/
Difference increase 1 error means S when inputting i/, /e/, /a/, 10/, /u/. ~S3 output bond waveform? It shows. In each of the above districts, U/V 4N
'74°, H/L mate@, F/B
G4 U = 4, A/U double number is yakiri amplification means S
The outputs of 0 to S3 are shown. fcSYM indicates the distinction τ of each voice, for example, in Fig. 8, 1% 1
% f% e, etc. are voiced sounds Vf, Vi, Vf
, Ve, . . . each meeting is shown. fc, and the code m is each voiced sound Va, Vi, Vu, Yes Vo, vh
% Vls Vfs Which of Vb and Vw is n?
VC also indicates a voiced sound Vmτ that does not apply. The 6th and 9th figures of the sword and figure are phase! Figure 11 shows an example of measurements taken on two male subjects named A, and two female subjects with 4N. It can be seen that each of the above-mentioned groups τ-n extracts the same characteristic &, regardless of the mother-in-law, but the /e/ sound of the vowel and /u/
The symbolization of sounds and 10/sounds is due to the know-how of the filter 11, and the symbolization is somewhat incomplete.

By the way, if we compare the above-mentioned first formant F of the vowel with Fig. 5 showing the distribution of the ``porridge 2 formant F2 and Fig. 4 showing the articulation position of the tongue, the first formant F1 is as follows.
If the jaw is wide open, the frequency will be high like /a/, and if the jaw is narrowly open, the frequency will be low like /i/, so it can be said that it corresponds to the wide and narrow of Habo'ka. Recognize.

On the other hand, it can be seen that the position of the sword 2 Polmanto F2 almost corresponds to the front and back of the tongue. '1: If you look at the second section and Figure 5, you can see that the second formant of the vowel varies greatly between men and women. However, in the conventional example, since the second formant is separated only by the VF/VB difference (i!), especially the vowel (e)
There was a problem that the separation of and (o, u) was incomplete. 1st, in VF/VB biid No. 6, among the 5 vowels in Figure 6 to No. 1, the part corresponding to / u / K on the chain side is expected to be detected as thick, but this is weak, and It turns out that the second half of the part corresponding to /e/ and the part corresponding to 10/ is also weak, which causes the notation to be incomplete. In order to eliminate such imperfections,
Previously, we proposed a method that extracts the deviation of the 7th vowel when uttering the 5th vowel as an offset for each individual.
Still, it goes without saying that it is better to make this type of offset A as rectangular as possible, and it goes without saying that the smaller the amount of offset, the better. The large second formant of V F
/ V B Since only the m signal is used, the offset of the zero-point fine soil of the output of the filter pair becomes quite large, and in some cases, it may not be possible to extract it completely. If you don't perform zero correction for the output of Kishi-no-zu like Ashitako, you'll be confused by the convulsions between the utterances of the husband and the onomatopoeic symbols. The edict against a young man had the disadvantage of preventing a decline in the balance. For this reason, it came to be known that words such as "ku" were used to describe people who were particularly cruel.

[Eleventh aspect of the invention] The present invention has been made in view of the above-mentioned points, and it is possible to extract the characteristics of the second formant of a vowel, thereby enabling a more complete symbolization of the five vowels. It is an object of the present invention to provide a system for determining the number 11 of the 11Lf outputs of the Ashida filter pair and to reduce the amount of zero point correction based on the word base of the Ashida filter pair.

[Disclosure of the invention]

FIG. 12 is a so-called claim correspondence diagram showing the collection described in claim 1 of the non-inventive claim in functional blocks. In the same area, Fvid 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. The output of the valley filter Fv%Fu is
. , and all difference signal components are extracted. It is a comparison means to Co, and the above difference w'JN1 means S. When the positive signal reading output from is smaller than the base l$1 direct Rv, allocate the sign JPj of own voice - mouth ■, and when it is larger than the base i$ILIRu, take into account the sign of the unvoiced voice UV, otherwise In the case of , the sign of silence S is determined. However, Ru ) 0 ) Rv. Next, Fa
1 is a filter that extracts the short-time average power of narrow-beat voiced sounds with a narrow jaw opening (vowels i, u, etc.), and Fu2 is a filter that extracts the short-time average power of narrow-pitched voiced sounds with a narrow jaw opening (vowels i, u, etc.).゛It is a short-term average power all-stations filter for definite voicing sounds (vowels e, a, and o). Next, Fb11I′i is a filter that extracts the short-term average power of the low sound of the first formant, such as the vowel esO, among recursive voices with a wide jaw opening, and This is a filter that eliminates the short-term average power of the first formant [6] sound, such as the vowel a, among the wide-ranging sounds of the phrase. Next, FC, i'1, is a filter that eliminates the short-term average power of low sounds in the second formant, such as the vowel 00, among the low sounds in the second formant, and FC2 is also Among the first formant's broad self-voiced sounds, the vowel e
This is a filter that eliminates the short-term average power of a 2-formant quotient such as '4'. Furthermore, Fdl is the second of vowel U-like sounds in the phlegmatic self-μ sound with a narrow opening of the jaw.
This is a filter that extracts the short-term average power of a sound with a low formant, and it is a filter that extracts the high short-term average power of a second formant such as the Fd21-1 vowel i. S0 to S4 are difference signals V/UV, Veao/Viu%Va/Veo, respectively.
SVe/Vo, ■i/Vu is calculated. Different porridge amplification means S.

In the soft means C9, the kP-values Rv, Ru(
Rv (0 (Ru)) and the output value is the standard 1
If it is smaller than 11'iRv, it is determined that the voice is V. (If the g output is larger than the reference i@Ru, it is determined that there is no sound U, and the quotient of the n1 reference value Ru and Rv is determined to be silent S. (For sound, P sound, 2 and unvoiced sound, any one of the valley marks of S, V, and V is distributed t4JI M
It is manually operated by yo. MCO is a matrix vtoJA-to which is the output of the valley pressure TMJ increasing means S1 to S, and this matrix calculation capital MCo is the output Veao/to of each running number.
Vius Va/Veo, Ve/Vo
, ViTv to calculate the total short-term average power of valley vowels 1, e, a, o, and U included in the voice input. In the configuration shown in FIG. 12, the first corner S means for calculating the ratio between the wide idiomatic sound vH and the narrow elliptical idiomatic sound VL, and the ratio between the isu ancient vocal sound VF and the isu tongue idiomatic sound VB are used. A differential@width means S6 is provided, and a matrix total p-to-MCo is provided.
Now, these valley cutting order 11141 moves W S 5, Se
] m1F+ Ml force v H/ V L s, -J: Vi ■
F/VBK predetermined matrix T ('fz ricep), wide @ own P sound (h) included in the audio input, fuku frequency practice ha 斤 (17, curved pronunciation μ sound (f), with postfix) It calculates the total power of voice (t) J, 2 and its 1st generation and restoration + 3 voice (w).

Matrix calculation city b1c: 'j] Yu'JT used in 6
v, Tc The output of the matrix calculation unit MCo is input to the maximum value judgment capital MXo, and the valley ratio I%es 1%, Os
Which is the largest component among ul hs 1.1% b%W? It is determined, and the sign of the largest component is manually entered into the symbolization process 61sMYo. When the difference between the largest component and the second largest component is small, the sign m is output. When the code output from the ratio soft means C8 is V, the symbolization processing unit MYo converts 1% es a, 0% u, h, z% output from the maximum slender foot unit MXo.
fl 1) When either of %W2 and m fully outputs one symbol, and the line output from the soft means Co 7Q is U or S, that symbol is output until its end. be. The composite code output from this symbol is input and stored in the standard pattern storage unit when registering a message, and is input and stored in the input butter-) storage unit when recognizing a voice message. , At the time of the matching process, #, which is most similar to iX among the standard patterns registered in advance, is set as the quasi-J\tashi-purple message.

In the configuration of the 2% 12 figure, the VH7VL difference number M and the VF/VB difference number are Veao/Viu, respectively.
A difference signal and a V e /V o difference signal may be used instead.

Figure 13? It is a block diagram showing an embodiment of the 18 filter system.In the above 12th configuration,
F u s F a s s F a 2 % F
b 1, F b 2, FC1% p c 2 s Fd
A total of 10 filters for I Sp d2 are required, but in the configuration of Kyō13, two of these filters are required. Eight filters are also used to extract the voice features.

In Figure 13, 2, VFh is the component of the ``Hl'' ancient vocal sound, VF is the component of the θ1'' ancient vocal sound, VB is the component of the ancient self-voiced sound, VHh is the 1 Karasugi relative portion of the wide-jawed own vocal sound, VHf is the output of a filter that extracts the low-frequency component of the in-phase voiced sound, and VL is the output of the filter that extracts the component of the voiced sound. And in the 10th worst case, the Veao/Vi difference signal and Va/Veo
For the difference signal VL for all Igawa, still V e / V
o difference signal and V i / V u difference signal VB(
or VL).

This is because the zero point of the difference signal of the filter pair corresponds to the difference point of the filter band (the frequency difference to the cross-over). Also, the band of the other filter is 2.
Depending on the type, the difference point of the filter band will change.

By the way, in the example of FIG. 13, VHh and VB
Since they are almost the same, I would like to put one in one VB, but
At 11 in Figure 13, all three signals are lost from VB, and the balanced reading of the filter pair becomes unreliable. Therefore, an example in which VB is divided into one high-frequency component VBh and a whole range component VB, and one VF is combined into one VC, and a 7-filter system is adopted is shown as the 14th example. In this case, since only two difference signals are extracted from VB, the filter paralysis becomes only 1m. From another perspective, it can be said that VBh gold is added to the conventional example shown in Fig. 5 and the conventional example shown in Fig. 1. Figure 15 (aJ (b) shows the components of the valley voice sound VL%VH, VB, VBh% when P is applied to the 7 filter method shown in Figure 14.
Special note on the number of cycles of the filter used to extract the short-term average power of VF? This figure (a) is drawn as a frequency dependence uniform scale on the horizontal axis, and in the same figure (bi is drawn from the jd wave number of Yokomoe U on a logarithmic scale.
It shows the characteristics of Karakin amplifier which will be explained later in AP.

This is an example of the 16th filter method. That is, in the above-mentioned first example, even if VH is used instead of VBll, the discrimination between the vowels i and U is qJ, so the same-wave component vectors are Uv, V, VF. , V
It can be configured with 6g of B%VH%VL (el filter). Is there another way to look at this example from Figures 1 and 6? In this case, the vF/■H polarization is added to the conventional example shown in Figure 5, and almost the same filter band can be used.

However, the VF/VB difference number is adjusted so that the vowels e and 0 can be clearly distinguished. Figures 17(a) and (b) are
In the 16th 6-filter method, each own voice component VL
, VH, VB% VF, short-term average power Total My, is a characteristic that shows the frequency characteristic of the filter used to extract the power, and Figure (a) shows the frequency of one horizontal lung as a monthly scale. , the same figure (b) is drawn with the frequency on the horizontal axis versus the drinking eye, and Z as the horizontal axis. In FIG. 15, 2 AP indicates the characteristics of the Iiyu amplifier, which will be described later.

'i! ! , the 8-filter method of the 13th example, the 7-filter method of the 1st example, and the 6-filter method of P and the 1st example. ■~7Pcs) Formulas are not used, and all elements other than the minimum required for identification are set to 0 in the 0-type transformation matrix Tm, so that 1t1-calculation can be performed faster. It is designed to provide redundancy in the part where the absolute value of In addition to making all the 5 vowel elements the same size (absolute value 14), the second formant F2
Regarding the two difference signals for the 5 vowels, the necessary grace is added to the needle 'IJ11' for each of the five vowels, and the first formant Fl is changed to the second form? It can be said that more importance was placed on this than F2. This transformation matrix Tm can be used arbitrarily depending on whether or not the area is targeted.

In particular, the components of the symbol Beta 1 - five vowels (i, e.

a, O, U), the element of the matrix corresponding to the 12th worst TV can be either (+1, 0.-1), so there is no need for the multiplication #-kai and only the code number. This allows easy symbolization. The elements of the transformation matrix of the symbol (h * lT t + b + w) corresponding to 1゛C in Figure 12 should be the same as the norm of the row vector iTv of this process. The norm of the row vector of c is smaller than the norm of the row vector of TV, and larger than the absolute value of the element of the matrix of Tv.

If you do not do this, each of the five vowel components (i, e, a, o
, u), other voiced sound components (, h ,' l
, f, b, w) will be smaller.

Next, a more specific example will be explained. The 18th example is the 7th film in Figure 14, which is a concrete example from the evening ceremony,
FIG. 19 shows a more specific example of the 6-filter system shown in FIG. 16, with only the F8 filter being far from the two. In the above-mentioned Taninu Example 2, audio is input from the microphone II), amplified by the preamplifier (2), and adjusted for gain and offset by the adjustment amplifier (3). Next, a level adjuster (5) balances the input power of the V/UV difference signal and other difference signals. (in general,
We will thoroughly emphasize the difference signal of Ike rather than V/UV Shoi Higo. )
Next, in the /UV balance adjuster t41, the filter Fv
Take the parasitic relationship between the input of the filter Fu and the input of the filter Fu. On the other hand, VB/VLJ\Lance-like headgear (tune 61 to the midpoint,
V H/V L balance reader 84171 with 1 filter F
Balance the inputs of H and filter FL, ■ F/VB balance, IAI scale (81 balance the inputs of filter F F' and filter FB (FBh). Next, VB/VL
Balance VB and VL with the balance adjustment device (6). In the 19th configuration, VB/VL balance tea utensil (6
) When all adjustments are made, VF/Vl (are balanced) at the same time.

The output of each filter is sequentially l
Convert the IJ to a logarithmic scale using a logarithmic converter (IO), and digitize it to an 8-pit binary number using an A/D converter (1 unit).Each filter is configured with a digital filter. In this case, the A/D converter (1) comes after the adjustment amplifier (3), and performs the valley filter calculation in sequence in a pipeline manner, like a kind of multi-flex filter (9). The output of each filter is calculated separately. Next, the mutual difference of this digital (lti) is calculated, and the difference signal hect Jb (UV/V, Veao/Viu, Va/ Veo
, Ve/Vo , V i/Vu ) are calculated.

Figure 20 - The second convex evil is about the example of the husband of the 18th evil,
The same voices as those shown in Figures 8 to 11 were input using a recording tape, and the characteristics of the voices were extracted. The results are shown in Figures 24 to 27. Similarly, is the sound the same as the sound in Figures 8 to '9pJl1? It shows the results of inputting all audio tapes and extracting audio features. These figures 20 to 2
In the seventh cause, the VF/VB7i signal of the conventional example is V
e/V.

There are now two signals, the difference a and the V i / V u difference signal, and the conventional VA / Vl) etc. signal is now V a / V e
o Difference G4 M (abbreviated as a/o in the figure).

In addition, in Fig. 20 to Fig. 27, in a/i, Ve
ao/Viu difference signal sound is blurred. However, in the conventional example, e%u, due to VF/VB & limit.

The detection of the difference becomes weaker as it approaches the zero point of the difference word, and esU%
Symbolization of O was difficult compared to i5 a, but the second
From 0 to 27th, V e / V o pid
The detection of e and o is not reliable in the issue, V i / V u
ii! Since the detection of U in JPj became smooth, it can be seen that the symbolization of the 5# sound can be performed more reliably. In particular, in FIGS. 24 to 27, the V l / V u difference signal distinguishes i and U more clearly than in the 20th to 23rd sections, and as far as the second example is concerned, it is more dank than that in FIG. m18. Figure 519 can be said to have five vowels in Kano.

Next, in FIG. 18 and FIG.
When it is more positive than f, it is judged as UV, when it is more negative than a certain set value RV, it is judged as ■, and all the intermediate values are judged as S. The start/end detection section (14) detects the start end of the audio based on ■ or UV judgment. Detected, while there is no sound or υ sample above a certain set value,
Symbol vector conversion unit 05 that detects the end when it joins °C
) is a matrix operation as shown in Fig. 14P and Fig. 16, and the symbol pect+V (1+ e * a+ 0+ u
+ b + l * f + b+ W) K L imitate. However, matrix operations are performed only in the interval of V. Symbol 11Z processing 8iS t161 &yo,■ is output when there is a maximum of 1 situation of revector (1a).In the UV and S sections, U
V%sf! :Output.

At f4, the formatting processing unit 1171 converts the repetition of the same symbol into a list of one symbol and its duration, and further, if the duration is less than a certain predetermined value and the symbols before and after are the same, all of these are unified. If the symbols before and after are different, include them in the song symbol.・Omit short duration symbols.

The time M-type normalization processing unit (18) calculates the total duration f so that the total duration of each list is 200 (or even 1000).

Is this the ratio between the total number of samples 200 (or 1000) and the duration, as in the conventional example? Continuation of each 11
It is best to match the hangings. This festival, the number of changes in the list is small (lo to 20), so I don't spend much time on the next division.

With the above process, a wooden sword style voice pattern can be created.

This voice pattern is registered as Seojun Butter Shiroki (Lazy 5 (19)) in the registration top.In the certified job t-do, it is matched with the standard pattern in the distance calculation section (2o), but first Primary identification is performed using Uv's song, etc., and the illumination target is limited to P〈.

Next, in the distance table (21j), find the distance (correlation 111) between the 8 points that correspond to each other at time IN, and use this as the distance between all the batas, which is the sum of all the samples.The distance table is For example, porridge with kumquats on the first side.

(Leaving space below) In Table 1, the horizontal Kusunoki P and Luo columns respectively correspond to the code of the standard pattern and the input cover pattern HIJ. For example, if the code of the standard pattern is 75≦a, and the input cover pattern is When the sign of is also a, the output of distance table +211 is 2, which means that the degree of approximation is high? Describe what you show. -1. When the sign of the standard J\ turn is UV and the sign strength of the input pattern is 2a, the output of the distance table (21) is Vi-2, indicating that the degree of approximation is low. Therefore, between the distance calculation sections, the output from the distance tailor)?2)? The degree of approximation of the entire pattern between the input cover pattern and the standard pattern can be easily calculated by simply adding the patterns in sequence.

The seasonal attraction is closer than the setting I, where the mosquitoes in the distance have a similar pattern, and even 24 days v? A certain setting 11I is better than something close to C! If it is more than V, this dragon is also close pattern and human power A turf is I mouth J! In the case of +w, the request is rejected as an audience failure. Figure 29 shows the structure of the merged invention outputted from ``Result Output Date'', which is a so-called claim vs. 28 is a block diagram showing the structure of the second example, which has become more commonplace. In each figure above, 88% Sl, S2, S3, S
, are respectively UV/V difference i=S, Ve
aO/Viu7i A horoscope, Va/Veo difference signal, Ve/Vo difference (g"l, Vi/'
This is an ax front amplification move to extract the V u f4 i lucky number. Valley Seki system increase 1 interval hand pattern S. The output of -84 is softened with a predetermined base Φ level in the Sor)C ratio softener . 251 is,
Envy J augmentation means S. If the output is above a certain positive value, it will be marked UVi, and if it is less than a certain number of shells, it will be marked V, and in other cases, it will be 99-
It is the one that assigns "S". Next VC Hii Onfu) Schiff j takes into consideration the code Veao when the output of the ax cut increases 1 and moves S1 is more than the value of the stop, and when it is less than the constant value of yellow, the code Viu f is assigned, In the case of the first generation, the code "S Zendon" is applied. Kakitatahi Software (
Ha)! 291 ids difference is assigned a sign Va when the output of the υ increase 1 quotient means S2 is above a positive constant value, is assigned a sign Veo '< when it is below a negative constant 1, and in other cases is assigned a sign ・s2 is assigned. Next, Hi Souki +3 [1) t31) is Ax control increase) Mente I
When the output of winning combination 83 is greater than or equal to the positive one-leg value, the sign V
e 'y +! a, and when it is less than a negative constant 1, the sign is V.

(i-61j, and in the case of fm, let S be θJ
It is something to be assigned. Furthermore, when the difference φυ is greater than a certain value of the output of the width means S4, the rib 193i is assigned ``i'', and when the shellfish foot 11 is less than (C code Vu 61 It is assigned the number S? to the 'f @ position of the first concave part of the zono.
]~(33j output μ person Kapit hatashi register state,
Vjs is my own place J3J'-'r
jT + c45), and the fortune-telling table tJ61 and diagnosis 1 (while turning, as in the case of the 12th evil, the code a,
eh 0% is u% hz 1% f, b%
Any one code of WS rn is loaded with stool 4. An example of the layout and encoding table l36i is shown in Table 2.

2nd: However, Table 2 shows that * can be either 0.1, and shows the case of O/liO and the case of 1. Such symbolization table A is configured on, for example, a ROM, and is stored in the input hit pattern register 13.
By accessing ROM'i as the contents of 4 as input for all addresses, the codes of each code such as 3% elo, etc. can be obtained as data output, or as shown in factor 29, Input pit pattern register +
It is an ellipse that compares with the output of the total symbolization table 1361 of 34 and outputs the code when they match.

Next, the 30th section is a flowchart showing a method of implementing the function of the symbolization processing section β5) by the sequential discrimination processing program of the microcomputer in the configuration of FIG. 28 described above. In the flowchart for this section, the first step is to divide the Veao/Viu difference signal into five groups depending on whether it is high (high level H), high level M (high level M), or low level. Then, in the second stage, when the man's first stage is H, V a / V e
o If the difference signal is H, it outputs /a/, if it is M, it outputs Jy9'/W/, and if it is L, it moves to the fifth step, checks the Ve/VO difference signal, and outputs H, 4 outputs /e/, M outputs /h/, and L outputs 10/. On the other hand, if the first stage is M, then in the second stage V
If the e/Vo difference signal is H, /f/company output output M
If it is L, /b/”k is output. Furthermore, if the first stage is L, in the second stage ■i/Vu
If the difference signal is H, /i/fa- is output, and if it is M, t
output to d/l/, and if it is L, /u/ will output the flowchart in FIG.
To explain the storm with -G, in the 7th episode, five vowels were symbolized from among the δ combinations of HlL of the five Sohakugo, so the positive and negative of each difference signal is as follows. It was also used to identify compound vowels among the 5th vowel, but since it is good to distinguish one 5th vowel in the positive and negative of the difference signal in the @ stand of 4830, a voice analysis filter is used. Band setting and difference signal balance adjustment can be performed optimally.

Here, the method of adjusting the balance of the m4H number of the filter pair will be explained as follows.

■Veao/Viu ax No. 1a uses the first formant F1 as before, with 5 vowels k (e% a, o) and (i%
Adjust so that it is divided into U).

■Va/Veo difference signal is (el 3% o) 4(a) and (e,) due to 941J1 formant F1.

).

(5) The V e /V o difference signal is adjusted so that s (e, o) is divided into (e) and (0) using the 2-formant.

■V + / V u difference is adjusted to 1 by 52 formacites so that (+, u) is divided into (1) and (u).

Figure 2 here? It can be seen that the first formant Fl and the second formant F2 have a large amount of impoverishment due to the 5th formant. On the other hand, if you look at the first convex axis, you can see that the male and female positions on the Fl axis are small, and the left side on 12M is large. In order to ensure consistent symbolization, in the case of this method, it is desirable to separate the filter bands for men and women. In order to switch the filter band in this way, you need to go out of the digital filter, write down the settings for men's, women's, P, and children's filters, respectively, and then change them accordingly. It is convenient to be able to cut it.

On the other hand, if we assume that the upper part of Figure 3 has some errors, then
All the men's and women's products are grouped together, and on the F1 axis (a
), (e, o), (i, u), and F2
It can be seen that (e) and (O), and (i) and (u) can be separated on @. In this case, (i) and (, u )
Although it is somewhat difficult to separate all men and women using a coplanar filter, as in the conventional example, (i, e) and (3
It is easier than dividing into % 0, u). removal, (e)
The boundary of the second formant F2 for and (O) is 1,5
Near K HZ, the distribution of (U) is centered, while (
Brown 2 formant F for e) and (U).

The boundary for s 1. Since the distributions overlap in the eS K Hz publication, it is necessary to divide the values into (i, e) and (aso%u) using the second formant F2 as in the conventional example, due to differences between men and women and individual differences. The Vie/Vaou difference (the offset M% of the zero point of the ig number becomes large, and there is a risk that it cannot be encoded correctly.On the other hand, in the case of an irregular method, the size I01 in the third worst distribution map is , it can be considered to be the most convenient and simple method to distinguish the sound by the difference signal, and the filter; Fl
;Area? I think it would be possible to completely symbolize the word if it were changed between men and women.

〔Effect of the invention〕

It is inventively structured as described above, and from the audio input, it can produce wide-jaw idiomatic sounds such as vowels a, ko, and sai.
With the ratio of 2 to the vowel i, tsu, which is the opening of the pronunciation, the filter pair of 1 is used, and the first formant, which is high in the first formant, such as the vowel a, is used. A second filter pair that cools the ratio between the vowel sound and the low sound of the first formant, such as the vowel sound, and the low sound of the first formant, such as the vowel sound, A fifth filter pair that calculates the ratio between a high sound in the second formant and a low sound in the double formant, such as the vowel o, and a high sound in the second formant, such as the vowel i, among narrowly voiced sounds. And the ratio of the low sound of the porridge 2 formant like the vowel? A fourth filter pair is provided, and audio characteristics are extracted from the difference signal output of the fourth filter pair.
The second formant of the vowel can now be extracted to special 41 katsuo, and among the five vowels in Japanese, it has been possible to extract the special 41 katsuo of esu and o, which were previously incomplete, and the 5th vowel. A more complete symbolization of 111IJ is possible, and when extracting the second formant, no forced extraction method is used as in the conventional method, so the amount of zero point correction by the filter pair ax I8 is reduced. ? It has the effect of reducing the amount. In addition, in the non-invention, seven I-dimensional vectors are inputted as all the difference signal output components of the pair of valley filters, and this Ushiku-Fushi vector is multiplied by a candidate matrix to obtain the five Japanese vowels and their first vowels. Short-term average power of voiced sounds? A matrix calculation unit is provided to calculate all the vectors for each category, a maximum value judgment group is provided to output all codes corresponding to the maximum component of the vectors output from column b to the calculation unit, and a comparison means is provided. Since the input cover pattern is formed by the output and the output of the maximum value 'flJ foot, the matrix calculation capital and the maximum IF
Is it a highly flexible and versatile method like the III judgment part? can be used to describe the sign of five vowels and other voiced sounds,
This also has the effect of simplifying the configuration of device a. Furthermore, in the merged invention, the differential outputs of the first to fourth filter pairs are compared with the Fuku reference value, and each of the differential outputs of the first to fourth filter pairs is compared in accordance with the magnitude relation with this base $1 value. Assign the sign of
Depending on the combination of numbers assigned to each filter pair, which of the five vowels in Japanese and the codes of the voiced sounds of the same type of voiced sound is selected? Voiced sound discrimination means that outputs by matching θJ? Since it is an advanced device, it has the effect of making it possible to identify voiced sounds at high speed by using a ROM table, etc., without having to undergo a punitive acceptance inspection. This has the effect of speeding up the movement of the gate operated by the °C-gear and being able to be constructed at low cost.

[Brief explanation of the drawing]

Is Figure 1 the quasi-spectrum of the five Japanese vowels? The second figure shows the gender difference in the vowel formants, and the third figure shows the difference between the vowel formants 1 and 2. The evil that shows,
Figure 4 is a diagram showing the relationship between the five Japanese vowels and the position of the tongue. Figure 5 is a block diagram showing the configuration of the conventional example.
(The frequency characteristics of the filter used to improve bJ are shown. The 7th problem is onomatopoeia in the conventional example @ symbolization processing procedure) 0-Chart, Figures 8 to 11 explain the production of improvement. Figure, the 12th culprit is a block diagram corresponding to claims showing the gist of non-invention @, ?, and the 13th culprit is the 5th example of Kazuo's non-invention
Figure 099, 14th edition of 弗 and the same example of the starving husband mentioned above.
The 15th problem (aJ (b) is the frequency characteristic of the filter that is improved in the month J? The problem is shown, the block diagram of the same as the 16th problem is the same as the above, and the block diagram of another husband example, Fig. 17 (a) (BL is improved The frequency of the filter used for 1st to 23rd evil
The operation theory (g) diagram of the 8th Ikuo Example, Riro 24 Iku to 27 are explanatory diagrams of the operation of the 19th Ikuo Example, and FIG. 6”) Rock Ward,
The 29th example is Kazuo's example of improvement, and the 30th example is a flowchart p-1 which shows the procedure of the onomatopoeic symbolization process as above. Fvt Fub Fa, , Fa2 % Fbl
, Fb2, F C1, F C2, F d1% Fa2
is a filter, 5o-5, is a light sword increase 11'fff means, Co ('J Hijo Tetsu, MCon row IJ calculation department, MX
oi・:JJt University 111.1 judgment capital, (he~shi33) is ratio e,, vessel,・36) is a symbolization table. Agent Patent Attorney Aburi Ishiyama Part 1 Figure 3 Figure 2 Figure 4 Figure 5 Figure 6 Frequency (KHz) Procedural amendment (voluntary) 1. Case Display Discount 1958 Special Request No. 672612, Name of the invention Voice message identification method 3, Person making the amendment Relationship with the case Patent applicant Location 1048 Bold Kadoma, Kadoma City, Osaka Name (583) Matsushita Denko Co., Ltd. Representative Iku Kobayashi 4, agent postal code 530 5, date of amendment order vlu J day”.

Claims (1)

[Claims] A difference signal output between a pair of filters that yields the short-term average power t of the cylindrical circumferential component and the low frequency component of the audio input. As an input, when the high frequency component is stronger, the sign of an unvoiced sound is t1. When the low frequency component is stronger, the sign of a voiced sound is t1. When the high frequency component and the low frequency component are approximately the same, the sign of silence + j"K is output. Ratio of broad voiced sounds with a wide jaw opening, such as the soft means fe ke, phonetic human power sword, and ra vowels, such as a, ko, and sai, and narrow voiced sounds, with a narrow jaw opening, such as the vowels i and tsu. and the first filter pair to obtain the 11th formant high tone,
A second filter pair that calculates the ratio of low sounds in the first formant such as the vowels ku and sai, and one of the low wide-mouthed sounds in the first formant, A sixth filter pair that calculates the ratio r of the high pitch and the low pitch of the second formant like the vowel, and the high pitch of the second formant like the vowel "I" among the narrowly voiced sounds, and the vowel. '44i 4 filter pairs and t are provided, and the left signal output of the 1st to brown 4 filter pairs is used as a component to create a 4-dimensional hect+b
= human power, and multiply this current-dimensional vector by a transformation matrix to make the short-term average power of the five Japanese vowels and other voiced sounds into a vector with six components. ), and the input pattern is formed by the output P of the comparing means and the output of the maximum value determining section. , This voice message identification method is characterized in that the standard pattern closest to the entering cover turn is identified as the input message by comparing it with the standard tS turns of several types of pine that have been recorded in advance. (2) Input the difference signal output of a pair of filters that extract the short-term average power of the high-frequency component P and low-frequency component of the audio input, and if the high-frequency component is more frequent, the signal of the unvoiced sound is determined. , when the low-frequency component is stronger, the sign t of a voiced sound is output, and when the high-frequency component and the low-frequency component are almost the same, a silent sign JPjτ is provided. The ratio of wide-jawed voiced sounds with a wide jaw opening, such as vowels, and bowed-jawed voiced sounds with a narrow jaw opening, such as The high sound of the first formant, such as the vowel a, and the sound of the vowel,
Comparison with the low tone of the first formant like 1<? The second filter pair that freezes, and among the low wide-jawed tones of the square l formant, the hollow sound of the second formant, which sounds like a vowel, and the low sound of the second formant, which sounds like a vowel. The ratio of the first convex filter pair of the ratio of narrow-jawed Confucian tones, such as the high sound of the second formant such as the vowel "I", and the low sound of the second formant such as the vowel "nori" xτ ice. The difference output of the first to second filter pairs is compared with a plurality of base values, and the difference output is calculated according to the magnitude relationship with this standard 1U. A voiced sound discriminating means is provided which allocates n separate notes and trays and outputs which one of the codes of the other voiced sounds of the n notes assigned to each pair of valley filters determines one code. , the output of the voiced sound discrimination means and the above J
Forming an input cover pattern r with the output of the comparison means, and comparing this human cover pattern with a plurality of pre-recorded standard patterns to identify it as the standard pattern τ human figure message closest to the input cover pattern; al- Characteristic voice message identification method.