JPS59195287A - Voice analyzer - 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 an improvement in a simplified speech analysis device.

Normally, a speech recognition device analyzes a speech waveform and performs discrimination calculation between a time series of characteristic parameters, which is the output of the analysis, and a pre-stored pattern to obtain a recognition result.

Conventionally, speech analyzes used in this speech recognition device include bandpass filter analysis, cepnutrum analysis, and modified cepnutrum analysis.

Speech waves can be considered to be the radiation output from the vocal tract excited by the vibration of the vocal cords, and the speech signal G(1) is (
]) It is expressed by the convolution of the vocal tract impulse 1ft) and the sound source waveform 5(t).

G (t) = R, (tl rice 8 (1...
・・・・・・・・・・・・・・・・・・・・・・・・(1
) Convolution operation When formula (1) is Fourier transformed, Gf (wl = Rftw) x 8 f (b)...
・・・・・・・・・・・・・・・・・・・・・・・・(
2). The sound source characteristic Sf (w) is a periodic line spectrum, and the vocal tract characteristic Rf (wl is the envelope of the voice spectrum Of (w). One way to obtain this envelope is to use a method that has a bandwidth above a certain value. There is a band-pass filter analysis in which multiple band-pass filters are arranged within the audio band.By having a bandwidth above a certain value, the influence of the line spectrum, which is a sound source characteristic, is weakened, and by arranging multiple band-pass filters, the entire envelope can be In other words, we were able to obtain the characteristics of the vocal tract.

Incidentally, in order to obtain more precise vocal tract characteristics, it is necessary to narrow the bandwidth of the bandpass filter, but when the band width is narrowed, the influence of the line spectrum, which is the sound source spectrum, becomes greater. Therefore, the band ll of the bandpass filter
@(Ha) It was not possible to reduce Ml+ beyond a certain degree, and it was not possible to obtain more precise vocal tract characteristics through band-pass filter analysis.
Cepstrum analysis is a method for obtaining more precise vocal tract characteristics. In 6° cepstrum analysis, equation (2) is linearly transformed, and 1oglGr(vν)l = log l R
f(w)l+≧glS f(w)l−-(3
Then, the cepstrum is obtained by inverse Fourier transform.

G c (ql = Rc (ql + SC(q)
・・・・・・・・・・・・・・・・・・・・・・・・
(4) As shown in equation (4), the product in the spectral domain becomes the sum in the cepstrum domain. The periodic line spectrum care strum 5c(q), which is the sound source spectrum 8 ((wl), appears only in the vicinity of the period Tp of the sound source.

−10,000, the vocal tract spectrum Rf(w) appears as the envelope of Gf(wl, and its cepstrum T(, c(q
) appears in the low quefrency area. That is, by performing cepstral analysis on the audio signal, it is possible to obtain vocal characteristics that are separated from the sound source characteristics into components with low frequency.

Furthermore, patent application No. 169031 (% cut 5
As described in Japanese Patent No. 7-185098, by using a cutting section that cuts out only the frequency components within the band from the voice spectrum and shifts them to zero frequency,
The influence of out-of-band characteristics of the transmission path can be removed.

In addition, scale conversion of the frequency axis using a mapping function that compresses the high frequency part, such as 11g scale conversion, 'Me
l! By using a school conversion section that performs scale conversion, etc., a modified cepstrum can be obtained in which weight is placed on the high frequency band and on the low frequency band, that is, the characteristics are close to the human auditory characteristics.

The scale conversion is performed using a mapping function Sm= as shown in FIG.
M(Se) allows data only within the band of the transmission path to be converted to a mausoleum scale or Mel! It is to rearrange to scale. That is, the spectrum of the Be-th scan is set as the Sm-th modified spectrum.

However, the aforementioned band-pass filter analysis, cepstrum analysis, and modified cepstrum analysis are based on Fourier transform, and require multiplication with a trigonometric function system, resulting in a large-sized apparatus.

On the other hand, since the Walsh transform, which is an approximation of the Fourier transform, is a transform using a binary orthogonal function system of ±1, the Walsh spectrum can be obtained by addition and subtraction. By using this Walshy transformation, it is possible to
It is known to realize a small pseudo bandpass filter as described in Japanese Patent No. 700. However, the pseudo bandpass filter has the disadvantage that it is not possible to obtain finer and more precise vocal tract characteristics.

The purpose of the present invention is to provide a compact device for approximating the cepstrum by replacing the 7-lier transform and inverse Fourier transform in cepstrum analysis with the co-button Worn-transform. An object of the present invention is to provide a speech analysis device for a hut.The speech analysis device according to the present invention includes a first Walsh transform section that performs Walshy transformation of an input signal, and a Walshy-Passault spectrum obtained from the output of the Walshy transform section. a conversion unit that performs the line conversion; a scale conversion unit that maps the output of the line conversion unit to a modified alternating number axis using a mapping function 8m=M (s e ) that performs scale conversion of the Walshy alternating number axis; It has a second Walshy transformer that performs Walshy transform on the output of the scale transformer.

Next, the Walshy transform used in the present invention will be explained. The Walsh transformation is defined by equation (5), where W″H;
-X・・・・・・・・・・・・・・・・・・・・・
. . . (5) However, the human column vector Hl of the Walsh spectrum X is a Walsh transformation matrix and is determined by the formula Hi=Hi-1 H1 1 Hl = ()-1.

(■ is the Kronecker product) The fast Kuorunyu transform (abbreviated as FWT), which is a method for quickly calculating the above equation (5), was published in December 1976, Journal of the Institute of Electronics and Communication Engineers Vow 59-A, I612, No. 11.
It is described in ``A Study on Fourier Transform and Worunyu Transform d″'' in No. 1135 Negative 1 from page 34.

Input: X is a column vector arranged in reverse binary order in a 5-column.

If the Walsh spectrum is W1 and the transformation matrix is C, then the FWT is W=C-X =Gn-Gn-1...G]・X...song...song (6)
It can be expressed as the product of n matrices. Here each G1d
+71, +81, (determined according to formula 91.

Gi=Ei■In-1・・・・・・・・・・・・・・・
・・・・・・・・・Between songs...songs・(2 pieces)
2 However, I is an identity matrix with 2' rows and 2'' columns, and dtag
() is a diagonal matrix with diagonal elements inside parentheses.

Here, reverse binary ordering means expressing natural numbers in binary, thinking of numbers with their digits reversed, and arranging them in the order of that number.
When H=3, X = (L X4 X2 X6XI X
5X3X7). Furthermore, each G becomes.

Each row of the matrix Gi has only two non-zero elements, representing an operation isomorphic to the butterfly operation used in fast Fourier transform. However, since the element is only ±1, it can be executed only by addition and subtraction. Also, the Walsh spectrum is W = (W.

W2W4W6W, W5W, W, ).

The speech analysis device of the present invention has the advantage that by replacing the Cepus 1 Ram analysis with the Fourier transform and the inverse 7-lier transform, the simple adder/subtractor can be constructed with an arithmetic unit. Furthermore, it has the advantage of obtaining more detailed and precise vocal tract characteristics than a pseudo bandpass filter analysis device that uses Walsh transform.

Next, the specific configuration of the apparatus of the present invention will be explained with reference to the drawings.

The first embodiment of the present invention, as shown in FIG.
It is composed of a second buffer memory section 6, a second Walshy calculation section 7, and a second Walshy conversion control section 8. First, input time series data is input to the first buffer memory section 1 and temporarily stored.

The calculation proceeds from the first stage to the nth stage by a control signal from the first Walshy transformation control unit 3 according to the calculation flowchart.

The processing of the i-th stage is to execute 2n-1 butterfly operations of the i-th stage shown in FIG.
This means multiplying the matrix of .

Butterfly operation is Ya-X8+Xb Yb=Xa-Xb ・・・・・・・・・・・・・・・
・・・・・・・・・・・・・・・・・・ aCS or Ya-X8-Xb Yb-Xa+Xb ・・・・・・・・・・・・・・・・・・
. . . aυ, which is obtained by the first Walshy calculation unit 2 shown in FIG. 3. In the butterfly operation, Xa and Xb are first read out from the first buffer memory section 1 and stored in registers 2 (11 and 202, respectively). and the result is Ya,
Yb is written to the location where Xa and Xb were stored in the first buffer memory section 1, and is written to the location where Xa and
The output of the adder 21 goes to the location of Xb, and if the tuυ formula is selected, the output of
1, the output of subtractor 212 is 1 to X, 1lCI)
It will be destroyed. The selection i, i of this uQ or the 09 formula is performed by the first wall conversion fUi control unit 3. When the above-mentioned processing is completed at the n-th stage, which is the final stage, the first batch of A/Usvectra is obtained.

After the Umerso conversion is completed, the conversion unit 4 shown in FIG.
The pupil power Walsh spectrum is obtained by the school conversion unit 5, and the mapping function Sm- that performs scale conversion is calculated.
M(87) performs mapping onto the deformed alternating number axis. That is, the scale conversion control section 51 emits a control signal according to the time chart NC shown in FIG. Wzl+1 is sequentially read out, squared by the multiplier 41 of the paddy conversion unit 4, and then sent to the adder 4.
2 and the accumulator 43, the powerWal7Us vector (p, =W2j+W2i+S) is obtained,
Subsequently, the mapping function value M(i), which is line-converted in the mausoleum conversion unit 4 and designated using the signal a2 as an address, is read out from the mapping function chifur memory unit 52, and the output M(il) is sent to the second buffer memory unit 6. The second Walsh transform operates in the same way as the first Walsh transform, and the second Walsh transform operates in the same way as the first Walsh transform.
Note that this is executed by the buffer memory unit 6, the second Qualsi transformation unit 7, and the second Qualsi transformation control unit 8. By the way, in normal speech recognition, only the low-order terms of Guvnutrum are conveniently used, so the second Walsh transform only needs to calculate the low-order terms. Therefore, the second Walsh transformation only needs to be calculated for the low-order terms of the transformation matrix of Equation (5), that is, for k with a small number of alternations in Equation (121), only N=2'' needs to be calculated.

A second embodiment of the invention converts the second Walsh transform to 01
The second Walsh conversion control section 8, the second Waln, and the stagnation calculation section 7 in the first embodiment are replaced by the seventh
This is a modification to the one shown in the figure. The second Ume A sea conversion control unit 7 issues a control signal in accordance with the tights and chi six shown in FIG. 8, and the accumulator 7
2 is cleared, and the deformed line power Walsh spectrum x7g=i is obtained from the second buffer memory section 6 according to the signal l)1.
ngp MIj f.

Readout +1 according to tRunyu transformation matrix HkIIK
Further, addition or subtraction is performed between the adder/subtracter 711i and the accumulator 72 using the signal b2 of C.mar-1. That is, when the signal b2 is +1, ACC+Xl-+ACC is performed, and when the signal b2 is -1, ACC-Xl→ACC is performed.3. When signal b1 becomes N-1, accumulator 7
A 2-Hewell transformation value Wk, that is, a pseudo-kegnutrum is obtained.

The present invention has been explained above based on Example 1b, but these d
Self-control does not limit the scope of the present invention.In particular, in the embodiment of the present invention, as an FWT algorithm (6)
As shown in the formula, the human time series is arranged in reverse binary Jl, and from G1 to O
The product was calculated sequentially up to n, but the time series X' in the normal order as shown in the t13 formula is Gn%! 1) It is obvious that it is also possible to adopt the method of taking the products sequentially up to Gt'' and obtaining the Walsh spectrum W in reverse binary order as the result 2. u3
1 Also, the power spectrum is PI-W2I×W2. +1
However, since a multiplier is required, Pi
It is obvious that a method of approximately obtaining the power spectrum as the sum of absolute values, such as ilW2 r l-1-IWz l+11, can also be adopted.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing scale conversion, FIG. 2 is a block diagram of the first embodiment of the present invention, FIG. 3 is a block diagram of the first Walshy calculation unit 2, and FIG. is a flowchart of the calculation of the first Walsh transformation, and FIG.
FIG. 6 is a scale conversion time chart, FIG. 7 is a block diagram of the second Walshy calculation section 7 in the second embodiment, and FIG. 8 is a block diagram of the scale conversion section 5. This is the timenayat of the second Walsh transformation. In the figure, l is the first buffer memory section, 2 is the first Walsh calculation section, 3 is the first Walsh conversion control section, 4 is the conversion section, 5 is the scale conversion section, 6 is the second buffer memory section, and 7 is the a second Walsh calculation unit; 8 is a second Walsh conversion control unit 201, 202 is a register; 211 is an adder;
212 is a reducer. @5 In Figure 5, 41 is a multiplier, 42 is an adder, 43 is an accumulator, 44 is a speech converter, 51 is a school conversion control section, 52 is a mapping function table memory, 71 is an adder/subtractor, and 72 is an accumulator. Gi 1 Kuchi etc. 2 Fig. 3 5 8 Gi 4 Fig.

Claims (1)

[Claims] a first Walshy transformer that performs Walsh transform of an input signal; a line transformer that obtains a Walshy Powers vector from the output of the Walsh transformer and performs line transformation thereof; and a mapping function Sm= that performs scale transformation of the Walshy alternating number axis. M(
se) a school transformer that maps the output of the boundary transformer onto a modified alternating number axis; and a second Walsh transformer that performs Walsh transform of the output of the scale transformer. Device.