JPS616698A - Pseudo bandpass filter 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] <Industrial application field> The present invention relates to an improvement in a simplified speech analysis device.

<Prior art and its problems> Normally, a speech recognition device analyzes a speech waveform, performs discrimination calculation between the time series of feature parameters that are the output of the analysis, and pre-stored bangs, and calculates the recognition result. I'm looking for. Conventionally, a bandpass filter analysis using Fourier transform has been used as a speech analysis method used in this speech recognition device. In this analysis, the waveform is Fourier-transformed to obtain a Fourier spectrum, and a certain band of the power spectrum is summarized as the output of the bandpass filter. Here, it is determined by Fourier transform, and can be determined at high speed by the well-known FFT algorithm. However, exp (j 2
πki/N), and it was necessary to provide a multiplier.

On the other hand, Walsh transform is a simplified conversion method of the Fourier transform, and an analysis method using this Walsh transform is described in ``Speech Recognition Apparatus'' in Japanese Patent Laid-Open No. 57-700. This Walsh transform can be considered as quantizing the element exp (j2πk i /N) in the Fourier transform to ±1, and an approximation of the Fourier transform can be obtained.

Since the elements in the Walsh transform are ±1, calculations can be made using only adders and subtracters, and since no multipliers are required, a compact analysis device can be realized.

However, in the bandpass filter analysis using the Walsh transform method, the quantization for the Fourier transform is rough, and the approximation is poor.This is due to the fact that a sinusoidal wave is approximated by a square wave. ), which had the disadvantage of poor analytical accuracy.

<Object of the Invention> The object of the present invention is to provide a small-sized speech analysis device that obtains a better approximation of band-pass filter analysis using Fourier transform by replacing the Fourier transform used for band-pass filter analysis with multivalued Walsh transform. That is, the object of the present invention is to provide a small-sized speech analysis device that can obtain a highly accurate spectrum.

<Configuration of the Invention> A pseudo bandpass filter analysis device according to the present invention includes a multi-value Walsh transform section that performs a multi-value Walsh transform on an input signal, and a power calculation section that calculates a pseudo power spectrum based on the output of the multi-value Walsh transform section. , has an in-band power calculation unit that calculates n in-band powers from the output of the power calculation unit.

<Principle of the present invention> Next, the multivalued Walsh transform used in the present invention will be explained. As already mentioned, the Walsh transform replaces the trigonometric function, which is an orthogonal function system in the Fourier transform, with a Walsh function device, which is a binary function of ±1. is obtained. However, since the trigonometric function was approximated to a binary function of ±1, the degree of approximation was poor. On the other hand, a multi-value Walsh transform device that can obtain a better approximation of Fourier transform by simple calculation by converting the Walsh function into multi-value and complex numbers is disclosed in Japanese Patent Application No. 1986-063186 filed by the same applicant. ) "Multi-level Walsh transform device". Here, we will discuss the principle of multivalued Walsh transform.

The column vector of input time series arranged in reverse binary order is X1 The multi-valued Walshy spectrum is W1 The transformation matrix is C, then W=C,X =G -G...G1・X(6) n
I]-1 It can be expressed as a product of n matrices. Here, each Gi is (7
), (8), and (9).

G, =-=] JiXI, 1i (
7) Ta/koshi■ is the Kronecker product LH-dlag(1, Ca1L(aiL"'+(al
:] ('') However, ■1 is a unit matrix with 21 rows and 21 columns, and diag() is a diagonal matrix whose diagonal elements are in parentheses.

Here, [:ai] is determined by the number of multivalues, and in the case of 8 values, a i = exp (jπ/21), ak = exp (j
θ) and exp(jθ))=1, O≦θ〈3, then 1=t+J, −≦θ〈hot=j, −≦θ〈〒, then = −1+j1−≦θ〈π shall be. In addition, in the case of 16 values, θ (exp(J)]-1o≦θ〈For Jin = 1-F-T1
, when 100≦θ×7=1+” -≦θ〈■, 2=hand j When 100≦θ〈■=j −≦θ〈I, then chant 1-7+j ,,≦θ〈7 When 1-x+J-≦θ〈■%4 Knee1+-7J, , When ■≦θ〈π.

In addition, reverse binary ordering means expressing a natural number in binary, thinking of a number with its digits reversed, and arranging it in the order of that number, n
In the case of = 3, X = (X(,, X4 , X2 , X6 , xl, X5
, X3, X7). Furthermore, in the case of 8-level Walshy transformation, each 01 becomes. In each row of G1, there are only two noble elements that are not zero, indicating that they are obtained by an operation isomorphic to the butterfly operation used in Rashi fast Fourier transform. This non-cell element is C±1. ±j), so it can be executed only by adding and subtracting complex numbers. 16 more
Non-zero element 1 for value Walsh transform
1.

is (±1. ±j), (±-2±J), (±1. ±-
J), it can be executed using only shift operations and addition/subtraction of complex numbers, and does not require a multiplication circuit. Also, Wi and WN 1 (N=2n) of this special multivalued Walsh vector
is a conjugate complex number.

In the pseudo bandpass filter analysis device according to the present invention, a pseudo spectrum of an input signal is obtained by the above-described multilevel Walsh transform, and the predetermined in-band power is obtained by combining the pseudo spectra.

<Example 1> The misfire H will be explained in more detail based on an example.

As shown in FIG. 1, the first embodiment of the present invention includes a multi-level Walsh transform section 1, a power calculation section 2. It is composed of an in-band power calculation section 3.

The multivalued Walsh transform unit 1 is, for example, disclosed in Japanese Patent Application No. 58-0.6.
Figure 5 or 10 described in specification No. 3186
It can take the configuration shown in the figure. An input signal is input to a multi-value Walsh transform unit 1, subjected to multi-value Walsh transform, and a pseudo spectrum W1 is output.

The power calculation section 2 is a section that calculates the power of the pseudo spectrum W1, and is configured as shown in FIG. 2. In the power calculation unit 2, the achimulator 23 is first cleared, and then the real part W1R and the imaginary part W, ■ of the pseudo spectrum W1 obtained in the multi-level Walsh transform unit 1 are sequentially input and squared in the multiplier 21. Then, the sum is obtained by an achimulator composed of an adder 22 and a register 23. That is, pseudo power spectrum P, = W
, R+W, l (10) are calculated.

The in-band power calculation unit 3 is a part that calculates the sum of the pseudo power spectra in the K divided bands, and is configured as shown in FIG. The pseudo power spectrum P obtained by the power calculation section 2 is stored in the buffer memory section 31.
The memory of this is temporarily burned.

The buffer memory section 31 is controlled by the signal l and signal k issued from the control section 36 according to the time chart shown in FIG.
Pl is read from the weighting coefficient memory unit 32, and \V is read from the weighting coefficient memory section 32, respectively, and input to the multiplier 33. The output of the multiplier 33 is sequentially summed by an accumulator constituted by an adder 34 and a register 35. That is, the kIp-th in-band power is divided. Here, Wk is a coefficient representing the characteristics of the bandpass filter, and generally takes a value as shown in FIG. Subsequently, the control unit 36 increases k by one to obtain K in-band powers, that is, outputs of K pseudo bandpass filters.

<Embodiment 2> A second embodiment of the present invention is a case in which a power spectrum is calculated approximately as a sum of absolute values without using a multiplier, and the power calculation unit 2 in the first embodiment is replaced by a 6
The configuration has been changed to the one shown in the figure. First the accumulator 24 is cleared and then the pseudo spectrum WI
The real part W1R and the imaginary part W, □ are sequentially input, the absolute value is determined by the absolute value circuit 24, and the sum thereof is determined by the adder 23 and the register 23. In other words, pseudo-no;
War spectrum P = < w, ) + (VV', 1) ・
...(12)+R can be divided.

<Embodiment 3> The third embodiment of the present invention is a case in which the weighting coefficient representing the physical properties of the bandpass filter is set to 1 or 0 at the beginning of the in-band power meter, so that the weighting coefficient is determined without using a multiplier. The in-band power calculation section 3 in the first embodiment is now changed to the configuration shown in FIG. 7. From the control section 36 to the eighth
It is issued according to the time chart shown in the figure. The signal i reads out the power spectra from the S-th to the e-th, and the sum is added to the adder 3.
4 and the accumulator 35.゛That is,
First, the pronounced in-band power is calculated.

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, the total yγ of the in-band power Bk is (
Although it is calculated using equation (11) or equation (13), it is clear that the method of performing I no 0 (,x transformation shown in equation (14)) or the method of taking the square root shown in equation (15) can also be adopted.

<Effects of the Present Invention> The pseudo-pand-pass filter analysis device fft,B of the present invention replaces the Hou IJ transform in band-pass filter analysis with a multivalued Walsh transform, and is configured with simple arithmetic units such as adders and subtracters. has the advantage of being able to Furthermore, it has the advantage that a more accurate spectrum can be obtained compared to a pseudo bandpass filter analyzer using Walsh transform.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the first embodiment of the present invention, FIG. 2 is a block diagram of the power calculation section 2, FIG. 3 is a block diagram of the in-band power calculation section 3, and FIG. 4 is a block diagram of the band-pass filter. A diagram showing weighting coefficients representing characteristics, FIG. 5 is a time chart of the in-band power calculation section 3, FIG. 6 is a block diagram of the power calculation section 2 of the second embodiment, and FIG. 7 is a diagram of the third embodiment. FIG. 8 is a block diagram of the in-band power calculating section 3 of the third embodiment, and FIG. 8 is a time chart of the in-band power calculating section 3 of the third embodiment. In the figure, 1...multivalued Walsh transform unit, 2...
Power calculation unit, 3... In-band power calculation unit, 21... Multiplier, 22... Adder, 23...
- Register, 24... Absolute value circuit, 31... Buffer memory section, 32... Weighting coefficient memory section, 33... Multiplier, 34... Adder, 35.
. . . register; 36 . . . control unit. 71'1 Figure 71-2 Figure O 3 Figure O 4 Figure I %i;+ 5 Figure 76 Figure 71-7 Figure 71-8 Figure + 2 - ・e(+1 8(2) ・・e(2)
5(K)--<1(K)R

Claims (1)

[Claims] a multi-value Walsh transform unit that performs a multi-value Walsh transform on an input signal; a power calculation unit that calculates a pseudo power spectrum from the output of the multi-value Walsh transform unit; and a band that calculates n in-band powers from the output of the power calculation unit. A pseudo bandpass filter analysis device characterized by having an internal power calculation section.