JPH0219478B2 - - 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 speech recognition devices include bandpass film analysis, cepstrum analysis, and modified cepstrum analysis.

A speech wave can be considered to be a radiation output from the vocal tract excited by the vibration of the vocal cords, and the speech 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 (1) When the formula is Fourier transformed, G _f (w) = R _f (w) × S _f (w) ... (2) becomes. The sound source characteristic S _f (w) is a periodic line spectrum, and the vocal tract characteristic R _f (w) is an envelope of the voice spectrum G _f (w). As a method for obtaining this envelope, there is a bandpass filter analysis in which a plurality of bandpass filters having a bandwidth of a certain value or more are arranged in the audio band. By having a bandwidth above a certain value, the influence of the line spectrum, which is the sound source characteristic, is weakened, and by arranging multiple lines, 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 greater. 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
_Then , the cepstrum _is obtained by inverse _Fourier transformation.

G _c (q) = R _c (q) + S _c (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 S _c (q) of the periodic line spectrum S _f (w) appears only in the vicinity of the period T _p of the sound source. On the other hand, the vocal tract spectrum R _f (w) appears as an envelope of G _f (w), and its cepstrum
R _c (q) appears in the low quefrency region. 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
As described in Publication No. 185098, by using an extraction section that extracts only the frequency components within the band from the audio spectrum and shifts them to zero frequency, the influence of characteristics outside the band of the transmission path can be removed. Can be done. In addition, scale conversion of the frequency axis using a mapping function that compresses the high frequency part, for example,
By using a scale conversion section 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.
This is to rearrange the data only within the band of the transmission path into log scale or Mel scale using S _n =M (S _e ). That is, the S _eth spectrum is
_S is the nth 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 transformation, a small-sized pseudo bandpass filter can be realized as described in Japanese Patent Application 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 purpose of the present invention is to provide a compact device for approximating the cepstrum by replacing both the Fourier transform and inverse Fourier transform in cepstrum analysis with the Walsh transform. Our objective is to provide analytical equipment.

The speech analysis device according to the present invention includes a first Walsh transformer that performs a Walsh transform on an input signal, a log transformer that obtains a Walsh power spectrum from the output of the Walsh transformer and performs a log transform, and a Walsh alternating number axis a scale conversion unit that maps the output of the log conversion unit onto a modified alternating number axis using a mapping function S _n =M (S _e ) that performs scale conversion; and a second scale conversion unit that performs a Walsh conversion of the output of the scale conversion unit. It has a conversion section.

Next, the Walsh transform used in the present invention will be explained. The Walsh transform is defined by equation (5), W=H _i・X ...(5) where W is the Walsh spectrum ₁ H ₁ = ( ^{1 1} _1-1 ).

(○× is Kronetska product) The fast Walsh transform (abbreviated as FWT), which is a method for quickly calculating the above equation (5), was developed in December 1975.
It is described in "Study on Fourier transform and Walsh transform" from pages 1134 to 1135 of the Journal of the Institute of Electronics and Communication Engineers Vol. 59-A. No. 12.

The vector of input time series arranged in reverse binary order is X.
If the Walsh spectrum is W and the transformation matrix is C, FWT can be expressed as the product of W=C・X =G _o・G _2-1 ...G ₁・X ...(6) and n times of matrices. . Here each G _i
is determined from equations (7), (8), and (9).

However, I _i is an identity matrix with 2 ⁱ rows and 2 ⁱ columns, and diag( )
is a diagonal matrix with diagonal elements in parentheses. Here, 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. If n = 3, then X =
(X ₀ X ₄ X ₂ X ₆ X ₁ X ₅ X ₃ X ₇ ). Furthermore, each G _i is becomes.

Each row of the matrix G _i has only two non-zero elements, and represents 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 ₀ W ₂ W ₄ W ₆ W ₇ W ₅ W ₃ W ₁ ).

The speech analysis device of the present invention has the advantage that it can be configured with a simple arithmetic unit using an adder/subtractor by replacing the Fourier transform and inverse Fourier transform in cepstrum analysis. 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 first embodiment of the present invention is as shown in FIG.
first buffer memory unit first Walsh calculation unit 2;
The first Walsh conversion control section 3, the log conversion section 4, the scale conversion section 5, the second buffer memory section 6, the second
It is composed of a Walsh calculation section 7 and a second 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 calculation is advanced from the first stage to the nth stage by the control signal of the first Walsh transformation control unit 3 according to the calculation flowchart for n=4 shown in FIG. 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 Y _b =X _a −X _b ...(10) or Y _a =X _a −X _b Y _b =X _a +X _b ...(11), and Fig. It is determined by the first Walsh calculation unit 2 shown in FIG. In the butterfly operation, first X _a ,
X _b is read from the first buffer memory section 1,
The data are stored in registers 201 and 202, respectively.
The adder 211 and the subtracter 212 calculate the addition and subtraction of equation (10) or (11), and the results Y _a and Y _b are stored in the first buffer memory section 1 where X _a and X _b were stored. will be written to. That is, when formula (10) is selected, the output of the adder 211 is written to the location where X _a of the first buffer memory section 1 was stored, and the output of the subtracter 212 is written to the location of X _b . When formula (11) is selected, the output of the adder 211 is written to the location X _b , and the output of the subtracter 212 is written to the location X _a . This selection of (10) or (11) is performed by the first Walsh conversion control section 3. When the processing is completed up to the n-th stage, which is the final stage, a Walsh spectrum is obtained in the first buffer memory 1.

After the Walsh transformation is completed, the result is shown in Figure 5.
A log power Walsh spectrum is obtained by the log conversion unit 4 and the scale conversion unit 5, and is mapped onto the modified alternating number axis using a mapping function S _n =M (S _l ) that performs scale conversion. That is, the scale conversion control section 51 issues a control signal according to the time chart shown _in FIG _{. 1} are sequentially read out, squared by the multiplier 41 of the log conversion unit 4, and the power wash spectrum (P _i = W ² _2i + W _2i+ ² ₁ ) is obtained using the adder 42 and the accumulator 43, and then log conversion is performed. The mapping function value M(i) which is log-converted in the unit 4 and designated using the signal a2 as an address is read out from the mapping function table memory unit 52, and its output MM(i)
The log power wash spectrum (logP _i ) is stored at address M(i) of the second buffer memory section 6 with the address signal a3 of the second buffer memory section 6 being set as the address signal a3 of the second buffer memory section 6.

The second Walsh transformation operates in the same manner as the first Walsh transformation, and is executed by the second buffer memory section 6, the second Walsh transformation section 7, and the second Walsh transformation control section 8. In addition, in the Walsh transformation, the forward transformation matrix C and the inverse transformation matrix C ^-1 have the same form (ignoring the difference in constant multiplication), so
Even if the second Walsh transform unit 7 performs the inverse Walsh transform, the final result will be the same. by the way,
Normally, in speech recognition, only the low-order terms of the cepstrum are used, so the second Walsh transform only needs to calculate the low-order terms. Therefore, the second Walsh transformation requires only the low-order terms of the transformation matrix in equation (5), that is, (12)
For k where the number of alternations in the equation is small, W _k = _N-1 〓 ^l=0 H _kl X _l ...(12) However, only N=2 ⁿ needs to be calculated.

The second embodiment of the present invention is an apparatus that calculates the second Walsh transformation using equation (12), and the second Walsh transformation control section 8 and the second Walsh calculation section 7 in the first embodiment are replaced by a seventh The configuration has been changed to the one shown in the figure. The second Walsh conversion control section 7 issues a control signal according to the time chart shown in FIG. 8, clears the accumulator 72 with the signal Cl7,
According to the signal b1, the modified log power Walsh spectrum X _l =logP _M(i) is read out from the second buffer memory section 6, and it is +1 according to the Walsh transformation matrix H _kl .
Alternatively, the adder/subtracter 71 performs addition or subtraction with the accumulator 72 in response to the -1 signal _b2 . In other words, if signal b2 is +1, ACC+X _l →ACC
and if signal b2 is -1, ACC→X _l →
Do ACC. When the signal b1 becomes N-1, a Walsh transform value _Wk , that is, a pseudo cepstrum is obtained in the accumulator 72.

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 .
Let G _o ^T be the time series X′ in the normal order as shown in equation (13),
It is clear that it is also possible to take the products sequentially up to G ₁ ^T and obtain the Walsh spectrum W' in reverse binary order as a result.

W′=G ₁ ^T・G ₂ ^T …G _o ^T・X′ …(13) Also, the power spectrum is obtained as P _i =W ² _2i , W ² _2i+1 , but it requires a multiplier. Because there is 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. 3 is a block diagram of the first wallet calculation section 2, and FIG. FIG. 5 is a block diagram of the log conversion section 4 and scale conversion section 5, FIG. 6 is a time chart of the scale conversion, and FIG. 7 is a flowchart of the calculation of the first Walsh transformation. 8 is a block diagram of the second Walsh calculation unit 7 in the example, and FIG. 8 is a time chart of the second Walsh transformation. In the figure, 1 is a first buffer memory section, 2 is a first Walsh calculation section, 3 is a first Walsh conversion control section, 4 is a log conversion section, 5 is a scale conversion section,
6 is a second buffer memory unit, 7 is a second wallet calculation unit, 8 is a second wallet conversion control unit 201,
202 is a register, 211 is an adder, and 212 is a subtracter. 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/subtractor, and 72 is an accumulator.

Claims (1)

[Claims] 1. A first Walsh transform unit that performs a Walsh transform of an input signal, a log transform unit that obtains a Walsh power spectrum from the output of the Walsh transform unit and performs a log transform thereof, and a mapping function S that performs a scale transform of a Walsh alternating number axis. By _n = M(S _e ), the above
A speech analysis device comprising: a scale converter that maps the output of the log converter onto a modified alternating number axis; and a second Walsh converter that performs Walsh transform of the output of the scale converter.