JPH0219477B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to an improvement in a simplified speech analysis device. Normally, speech recognition devices analyze speech waveforms and
Discrimination calculations are performed between the time series of feature parameters that are the output of the analysis and pre-stored patterns to obtain recognition results. Conventionally, speech analyzes used in this speech recognition device include band filter analysis, cepstrum analysis, and modified cepstrum analysis. It can be considered to be the radiation output from the vocal tract excited by the vibration of the vocal cords of the voice wave, and the voice signal G(t) is expressed as the impulse response R(t) of the vocal tract as shown in equation (1). It is expressed by convolution of the sound source waveform S(t). G(t)=R(t)*S(t) ……(1) * is the convolution operation When formula (1) is Fourier transformed, Gf(w)=Rf(w)×Sf(w) ……(2) becomes. The sound source characteristic Sf(w) is a periodic line spectrum, and the vocal tract characteristic Rf(w) is an envelope of the voice spectrum Gf(w). As a method for obtaining this envelope, there is a band-pass filter analysis in which a plurality of band-pass filters having a bandwidth of a certain value or more are arranged in the audio band. Having a bandwidth above a certain value weakens the influence of the spectrum, which is a sound source characteristic, and by arranging multiple envelopes, it was possible to obtain the overall characteristics of the envelope, that is, the vocal tract characteristics. By the way, in order to obtain more precise vocal tract characteristics, it is necessary to narrow the bandwidth of the bandpass filter, but when it is narrowed, the influence of the line spectrum, which is the sound source spectrum, becomes more pronounced. For this reason, the bandwidth of the bandpass filter cannot be narrowed beyond a certain level, and more precise vocal tract characteristics cannot be obtained by bandpass filter analysis.
On the other hand, cepstral analysis is a method for separating vocal tract characteristics and sound source characteristics and obtaining more precise vocal tract characteristics. In cepstral analysis, equation (2) is further transformed into log
Transform, log|Gf(w)|=log|Rf(w)|+log|Sf(w)...(3) Next, obtain the cepstrum by inverse Fourier transformation. Gc (q) = Rc (q) + Sc (q) ... (4) As shown in equation (4), the product in the spectral domain becomes the sum in the cepstrum domain. sound source spectrum
The cepstrum Sc(q) of the periodic line spectrum Sf(w) appears only in the vicinity of the period Tp of the sound source. On the other hand, the vocal tract spectrum Rf(w) appears as an envelope of Gf(w), and its cepstrum
RC(q) appears in the low que frency part. That is, by performing cepstral analysis on the audio signal, it is possible to obtain vocal tract characteristics that are separated from the sound source characteristics into components with low quefrency. Furthermore, the specification of Japanese Patent Application No. 56-069031 (JP-A-57-069031)
As described in Publication No. 185098), by using an extraction section that extracts only the frequency components within the band from the voice spectrum and wrinkles them to zero frequency, the influence of the characteristics outside the band of the transmission path is removed. be able to. In addition, scale conversion of the frequency axis using a mapping function that compresses the high frequency part, for example,
By providing a scale conversion unit that performs log scale conversion, Mel scale conversion, etc., a modified cepstrum can be obtained in which weight is placed more on low frequencies than on high frequencies, that is, having characteristics close to human auditory characteristics. The scale conversion is performed using a mapping function as shown in Figure 1.
By using Sm=M(Sl), data only within the band of the transmission path is rearranged into log scale or Mel scale. The spectrum of Sl and
This is to be the Smth deformed spectrum. However, the above-mentioned bandpass filter analysis, cepstrum analysis, and modified cepstrum analysis
It is based on Fourier transform and requires multiplication with a trigonometric function system, which has the disadvantage of increasing the size of the device. 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 only by addition and subtraction. It is known that by using this Walsh transform, a small pseudo bandpass filter can be realized as described in Japanese Patent Laid-Open No. 57-700. However, the pseudo bandpass filter had the disadvantage that finer and more precise vocal tract characteristics could not be obtained. The object of the invention is in cepstral analysis.
To provide a compact device for obtaining an approximate value of the cepstrum by replacing the Fourier transform and inverse Fourier transform with a multivalued Walsh transform, that is, a compact speech analysis device that can obtain finer and more precise vocal tract characteristics. be. The speech analysis device according to the present invention includes a first multi-value Walsh transform unit that performs a multi-value Walsh transform of an input signal, and a log transform that obtains a Walsh power spectrum from the output of the first multi-value Walsh transform unit and performs a log transform on the output of the first multi-value Walsh transform unit. and the mapping function Sm=M(Sl) that performs scale conversion of the Walsh alternating number axis.
It has a scale conversion section that maps the output of the log conversion section onto a modified alternating number axis, and a second multi-value Walsh conversion section that performs multi-value Walsh conversion of the output of the scale conversion section. Next, the multivalued Walsh transform used in the present invention will be explained. The Walsh transform replaces the trigonometric function, which is an orthogonal function system in the Fourier transform, with the Walsh function, which is a binary function of ±1, and an approximate value of the Fourier transform can be obtained only by addition and subtraction. However, since the trigonometric functions were approximated to binary functions of ±1, the degree of approximation was poor. On the other hand, the multivalued Walsh function, which can obtain a better approximation of the Fourier transform by simple calculations, was developed by the same applicant in the Showa era.
This method is described in Japanese Patent Application No. 1986-63186 entitled "Multiple Walsh Transform Device" filed on April 11, 1958. Here, we will discuss the principle of multivalued Walsh transform. As already mentioned, the Walsh function is
Since it is a trigonometric function quantized to ±1, by introducing a multivalued Walsh function with finer quantization, it can be made closer to the Fourier spectrum. For example, it is shown in FIG. A multivalued Walsh function with eight elements is considered. But with this method

Since it has elements such as [expression], multiplication is required for its conversion. In the specification of Japanese Patent Application No. 58-63186, in the case of 8-value Walsh transform, as shown in FIG.
-j, 1-j) are used. 8 due to this function system
Since the value Walsh conversion is an operation between the products of ±1 and ±j, it can be executed only by addition and subtraction. Also, based on the same idea, the 16-value Walsh transform is as shown in Fig. 12, -1, -1- 1/2j, -1-j, -1/2-j, -j, 1/2-j
, 1-j, 1-1/2j) are used. The 16-value Walsh transform using this function system is a multiplication operation with ±1, ±1/2, ±j, ±1/2j, so it can be performed only by halving using a shifter and addition/subtraction. , virtually no multiplication is required. If the column vector of the input time series arranged in reverse binary order is X, the multilevel Walsh spectrum is W, and the transformation matrix is C, then W=C・X = G _o・G _o-1 ...G ₁・X ...(6) It can be expressed as a product of n matrices. Here each G _i is
Determined by equations (7), (8), and (9). G _i = E _i I _oi where is the Kronetzker product L _i = diag (1.[ai].[a ² _i ].….[a ^2i-1 ₁ ])…
…(9) where I _i is an identity matrix with 2 ⁱ rows and 2 ⁱ columns, and diag( )
is a diagonal matrix whose diagonal elements are inside the parentheses. Here, [a _i ] is determined by the number of multilevel conversions,
In the case of 8 values, a _i = e x p (jπ/2i), a ^k _i = e x p (jθ), [e x p (jθ)] = 1, when 0θ < π/4 = 1 + j, π When /4θ<π/2 =j, when π/2θ<3π/4 =1+j, and when 3π/4θ<π. In addition, in the case of 16 values, [e×P(jθ)] = 1, when 0θ<π/8 = 1+1/2i, when π/8θ<π/4 = 1+j, when π/4θ<3π/8 When = 1/2 + j, when 3π/8θ < π/2 = j, when π/2θ < 5π/8 = -1/2 + j, when 5π/8θ < 3π/4 = 1 + j, 3π/4θ < 7π /8=−1+1/2j, 7π/8θ<π. In addition, reverse binary order means to express a natural number in binary, consider a number with its digits reversed, and arrange it in the order of that number. In the case of n=3, X=(X ₀ X ₄ X ₂ X ₆ X ₁ X ₅ X ₃ X ₇ ). Furthermore, in the case of 8-level Walsh transform, each G _i is becomes. The number of non-zero elements in each row of G _i is 2
This shows that it can be obtained by an operation isomorphic to the butterfly operation used in fast Fourier transform. Since this non-zero element is (±1.±j), it can be executed only by adding and subtracting complex numbers. Furthermore, in the case of 16-value Walsh transform, the non-zero elements are (±1/±1/2, ±j/±1/2j), so it can be executed only by shift operations and addition/subtraction of complex numbers. Also,
W _i of the multilevel Walsh spectrum obtained at this time
and W _o/2-i (N=2 ⁿ ) are conjugate complex numbers. The speech analysis device of the present invention has the advantage that it can be configured with a simple arithmetic unit such as an adder/subtractor by replacing the Fourier transform and inverse Fourier transform in cepstral analysis with a multivalued Walsh transform. Furthermore, compared to pseudo-bandpass filter analyzers that use Walsh transform, it has the advantage of providing more detailed and precise vocal tract characteristics. Next, the specific configuration of the apparatus of the present invention will be explained with reference to the drawings. The embodiment of the present invention, as shown in FIG. It is composed of a 2-buffer memory 6, a second multi-valued Walsh calculation section 7, and a second multi-valued Walsh conversion control section 8. First, input time series data is input to the first buffer memory section 1 and temporarily stored. After being stored, the first
The calculation proceeds from stage to nth stage. The processing at the i-th stage is to execute the 2 ^n-1 butterfly operations of the i-th stage shown in FIG. 4, and means to multiply by the matrix of G _i in equation (7). The butterfly operation is Y _a =X _a +X _b ·a _k Y _b =X _a −X _b ·a _k (10), and is obtained by the first multivalued Walsh arithmetic unit 2 shown in FIG. At the beginning of butterfly operation
X _a . X _b is read from the first buffer memory section 1, and the real part and imaginary part of X _a are stored in registers 201 and 2.
02, the real part and imaginary part of X _b are stored in the register 203,
204, respectively. Complex number multiplication of X _a · a _k is performed by the following four types of addition and subtraction in the case of 8-value Walsh transform. (z _R + jz _I ) = (X _bR + jX _bI ) · a _k , when a ₀ = 1 z _R = X _bR z _I = X _bI when a ₂ = 1 + j z _R = X _bR −X _bI z _I ＝X _bR ＋X _bI a ₄ When = j z _R = −X _bI z _I ＝X _bR a ₀ When = 1 + j z _R = −X _bR −X _bI z _I ＝X _bR −X _bI ...(11) In the calculation flowchart of FIG. 4, a ₁ .a ₃ .a ₅ .a ₇ are the same values as a ₀ .a ₂ .a ₄ .a ₆ , respectively. The calculation of equation (11) is the first
It is determined by the switch 211 and the adders/subtractors 221 and 222 under the control signal of the multilevel Walsh conversion control section 3. In other words, the switch 211 selects the inputs of the adders/subtractors 221 and 222 from X _bR , X _bI , or zero, and the adders/subtractors 221 and 222 perform addition, subtraction, or addition sign inversion to perform the calculation of equation (11) above. . Subsequently, addition and subtraction in equation (10) are performed by adders 231 and 232 and subtracters 233 and 234, dividing into real and imaginary parts. The obtained result Y _a ,
Y _b is written to the location in the first buffer memory section 1 where X _a and X _b were stored. When the processing is completed up to the n-th stage, which is the final stage, the first buffer memory 1
A multivalued Walsh spectrum is obtained. After the multivalued Walsh transform is completed, the log conversion unit 4 and scale conversion unit 5 shown in FIG.
A power multilevel Walsh spectrum is obtained, and is mapped onto a modified alternating number axis using a mapping function S _n =M (S _e ) that performs scale conversion. That is, the scale conversion control section 51 issues _a control signal according to the time chart shown in _FIG . Read W _2i+1 sequentially,
Squared by the multiplier 41 of the log conversion unit 4 and added by the adder 42
The power multilevel Walsh spectrum (Pi=W ² _2i +W ² _2i+1 ) is obtained using the and accumulator 43,
Subsequently, the mapping function value M(i), which is log-converted in the log conversion section 4 and designated using the signal a2 as an address, is read out from the mapping function table memory section 52, and the output M(i) is stored in the second buffer memory section 6. The log power multilevel Walsh spectrum (logP _i ) is stored in the second buffer memory section 6 as the address signal a3.
(i) Store at address. The second multi-value Walsh conversion operates in the same manner as the first multi-value Walsh conversion, and is executed by the second buffer memory section 6, the second multi-value Walsh conversion section 7, and the second Walsh conversion control section 8. Note that the above explanation assumes that the multivalued Walsh transform is performed again after scale conversion, but after this scale conversion (more precisely, immediately after converting to an absolute value), the signal becomes a real function that is symmetrical about the vertical axis. It is becoming. Generally, the Fourier transform and inverse Fourier transform of a real function symmetric about the vertical axis give the same result. The same thing holds true in this multivalued Walsh transformation. Therefore, in the embodiments already described, even if the second multi-value Walsh transform section 7 performs the multi-value inverse Walsh transform, the result will be the same. By the way, in normal speech recognition, only the low-order terms of the cepstrum are used, so the second multi-level Walsh transform only needs to calculate the low-order terms. Therefore, if we calculate only the low-order terms of the transformation matrix of the second multivalued Walsh transformation equation (6), that is, only the small k of W _k = _N-1 〓 ^l=0 H _kl・X _l ...(12) good. Furthermore, since Xl is an even function, it is the real part of H _kl .
Just use H′ _kl . W _k = _N-1 〓 ^l=0 H′ _kl・X _l ...(13) The second embodiment of the present invention is a device for calculating the second Walsh transform using equation (13). In this embodiment, the second multi-valued Walsh conversion control section 8 and the second multi-valued Walsh calculation section 7 in the embodiment are changed to the configuration shown in FIG. The second multi-level Walsh conversion control section 7 issues a control signal according to the time chart shown in FIG. The adder/subtractor 71 reads out the multi-valued Walsh spectrum X _l =logPM(i) and performs addition or subtraction with the accumulator 72 using the +1 or -1 signal b2 according to the real part H′ _kl of the multi-valued Walsh transform matrix. I do. That is, when the signal b2 is +1, ACC+X _l →ACC is performed, and when the signal b2 is -1, ACC-X _l →ACC is performed. When the signal b1 becomes N-1, a Walsh transform value _Wk , that is, a pseudo cepstrum is obtained in the accumulator 72. Next, a third embodiment of the present invention is an apparatus in which a 16-value Walsh transform is adopted as the multi-value Walsh transform, and the first multi-value Walsh calculation section in the first embodiment is changed to the configuration shown in FIG. This has been changed. The calculation proceeds in the same manner as in the first embodiment.
The difference from the first embodiment is that there are eight types of values for the multiplication element a _k in the butterfly operation. Complex number multiplication in equation (10) is performed using the following eight operations. (z _R + jz _I ) = (X _bR + jX _bI )・a _k When a ₀ = 1 z _R = X _bR z _I = X _bI When a ₁ = 1 + 1/2j z _R = X _bR - 1/2X _bI z _I =1/2X _bR +X _bI a ₂ =1+j z _R =X _bR −X _bI z _I =X _bR +X _bI a ₃ =1/2+j z _R =1/2X _bR −X _bI z _I = X _{bR +} 1/2X _bI a ₄ When ₌ j z _{R =} -X _bI z _I = X _bR a ₅ = -1/2 + j When z _R = _-1/2 -1/2X _bI a ₆ When =-1+j z _R =-X _bR -X _bI z _I =X _bR -X _bI a ₇ When =-1+1/2j z _R =-X _bR -1/2X _bI z _I = 1 _{/ 2x bR -
XbR} ． _XbI ． 1/2X _bR . 1/2X _bI . One of the zeros is selected, and adder/subtractor 221 and 222 perform addition or subtraction and sign inversion to execute complex multiplication of equation (11). Although the present invention has been described above based on Examples, these descriptions do not limit the scope of the present invention. In particular, in the embodiment of the present invention, as an FWT algorithm, the input time series is arranged in reverse binary order as shown in equation (6), and the products are sequentially obtained from G ₁ to G _o .
(13) The forward-ordered time series X′ and G _o ^T as shown in equation (13),
Take the products sequentially up to G ₁ ^T , and as a result, the inverse 2
It is obvious that the method of obtaining the progressive Walsh spectrum W′ can also be adopted. W′=G _1T・G _2T・…G _o ^T・X′……(15) Also, the power spectrum is obtained as P _i =W ² _2i +W ² _2i +1, but since a multiplier is required, P _i =
It is obvious that a method of approximately obtaining the power spectrum as a sum of absolute values such as |W _2i |+|W _2i +1| can also be adopted.

[Brief explanation of drawings]

FIG. 1 is a diagram showing scale conversion, FIG. 2 is a block diagram of the first embodiment of the present invention, FIG. The figure shows the calculation flow of the first multi-level Walsh transform, FIG. 5 is a block diagram of the log converter 4 and the scale converter 5, FIG. The figure is a block diagram of the second multi-level wallet calculation unit 2 in the second embodiment of the present invention, FIG. 8 is a time chart of the second multi-level wallet conversion, and FIG. A block diagram of the first multivalued Walsh calculation unit 2 in the embodiment of FIG. 10, FIG. 11,
FIG. 12 is a diagram for explaining the multivalued Walsh transform used in the present invention. In the figure, 1 is a first buffer memory section, 2 is a first multi-value Walsh calculation section, 3 is a first multi-value Walsh conversion control section, 4 is a log conversion section, 6 is a second buffer memory section, and 7 is a second multi-value Walsh arithmetic unit,
8 is a second multi-value Walsh conversion control unit 201, 20
2, 203, 204 are registers, 211, 212
is a switch, 221, 222 is an adder/subtractor, 23
1,232 is an adder, 233,234 is a subtracter,
241 and 242 are shifters. In FIG. 5, 41 is a multiplier, 42 is an adder, 43 is an accumulator, 44 is a log converter, 51 is a scale conversion control section, 52 is a mapping function table memory, 71 is an adder, and 72 is an accumulator.

Claims (1)

[Claims] 1. A first multi-value Walsh transform unit that performs a multi-value Walsh transform of an input signal; a Walsh power spectrum is obtained from the output of the first multi-value Walsh transform unit; Mapping function Sm = M (Sl) that performs scale conversion
A speech analysis device comprising: a scale conversion unit that maps the output of the log conversion unit onto a modified alternating number axis; and a second multi-value Walsh conversion unit that performs multi-value Walsh conversion of the output of the scale conversion unit. .