JPS597400B2 - speech analysis device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech analysis device used in speech analysis and synthesis systems and speech recognition devices, and in particular to a speech analysis device for extracting partial autocorrelation coefficients, which are one of the basic parameters of speech. Regarding.

In recent years, speech analysis and synthesis systems have been developed that break down speech into sets of basic parameters and transmit them at a low bit rate, and use these basic parameters to synthesize speech on the receiving side, as well as speech analysis and synthesis systems that break down speech into sets of basic parameters, transmit them at low bit rates, and synthesize speech using these basic parameters on the receiving side. Speech recognition devices that extract speech are becoming widely used.

As one of the basic parameters of speech in this case, a speech analysis device is known that focuses on the correlation between adjacent sample values of the speech signal waveform and extracts a partial autocorrelation coefficient (Japanese Patent Publication No. 49-18007 ).

Here, the partial autocorrelation coefficient is, as already known, the difference between the correlation between two points in the audio signal waveform, predicted by the sample value between D2 points, and the actual sample value. This type of speech analysis device is expressed by the correlation of
Stages of lattice delay filters 21 to 2n are connected in cascade, and each stage of delay filters includes a delay circuit 3, multipliers 4, 5,
It is constructed as shown by an adder 6, T, and a correlator 8 which output a sum or difference signal. In this configuration, the correlator 8 calculates an analog correlation between the sampled audio signal input to the audio signal input terminal 1 and the audio signal delayed by one sample time interval by the delay circuit 3, and multiplies it by this value. Determine the coefficients of devices 4 and 5. However, in reality, the existence of the correlator 10 for determining this analog correlation increases the circuit scale, which is a major obstacle in realizing the device. Therefore, the present inventor invented a speech analysis device that does not require such a correlator 10 (Patent Application No. 55-4).
1547)0 In this invention, instead of the above correlator 10,
The coefficient generation circuit and the coefficients obtained by this circuit are added to the adder 6.
, 7 is used to modify the output signal value of at least one of them in the direction of decreasing.

With this configuration, the circuit scale of the speech analysis device can be reduced. The present invention is a further improvement of this speech analysis device.

The present invention connects multiple stages of lattice-type delay filters as described above in cascade, or repeatedly uses each means of these filters to perform signal processing at each stage, and provides sampling to the first and second input terminals of the first stage. The speech analysis device inputs a speech signal inputted to the coefficient generation means of each stage and extracts the obtained coefficients as partial autocorrelation coefficients. It consists of correlation means that correlates at least one of the input signal to the input terminal or the output signal of the delay means and the output signal of the second addition means, and attenuation means that attenuates the output of the correlation means by a predetermined amount. The attenuation amount of the attenuation means in the next stage is configured to be smaller than the attenuation amount of the attenuation means in the lower stage.

With this configuration, the present invention has the advantage that partial autocorrelation coefficients can be determined more quickly and accurately.

Examples of the present invention will be described below.

FIG. 2 shows the configuration of a speech analysis device according to an embodiment of the present invention. Even in this embodiment, it is constructed by cascading a plurality of n stages of lattice type delay filters 21 to 2n. The delay filters in each stage are constructed as follows. Reference numerals 031 and 32 are input terminals, and the input signal to the first input terminal 31 is delayed by one sample time interval T by the delay circuit 33, and then the first The input signal to the second input terminal 32 is input to the second multiplier 36 and the second adder 37. 1st, 2nd
The coefficients of the multipliers 34 and 36 are given by a coefficient generation circuit 38, and the coefficients generated by this circuit are modified by a coefficient modification circuit 39 so that the output signal value of the first adder 35 becomes smaller. . The coefficient generation circuit 38 includes an adder 40 and a delay circuit 41 that delays the output signal by one sample time interval T. Adder 40 adds the output signal of coefficient correction circuit 39 to the output signal of delay circuit 41. On the other hand, in this embodiment, the coefficient correction circuit 39 is connected to the second input terminal 3
The multiplier 42 multiplies the input signal to the first adder 2 by the output signal of the first adder 35, and the attenuator 43 quantitatively attenuates the multiplication output.The output thereof is input to the adder 40. A sampled audio signal is sent to the first and second input terminals 31 and 32 of the first-stage delay filter 21 of the audio analysis device in which the lattice-type delay filters configured in this way are connected in cascade through the audio signal input terminal 30. Input the audio signal.

The first adder 35 in each stage sends a sum or difference signal between the output signal of the delay circuit 33 and the output signal of the second multiplier 36, that is, a backward prediction error signal, to the first output terminal 44, Second
The adder 37 inputs the input signal to the second input terminal 32 and the first
The sum or difference signal with the output signal of the multiplier 34, that is, the forward prediction error signal, is sent to the second output terminal 45. A partial autocorrelation coefficient Ki (1=1 to n) is obtained from the delay circuit 41 in the coefficient generation circuit 38 of each stage, and
4, 36, and taken out to the outside for transmission.Here, as is already known, the partial autocorrelation coefficient refers to the correlation between two adjacent points in the audio signal waveform. It is expressed by the correlation of the difference between the value predicted by the sample value between points in time and the actual sample value.〇Based on the above embodiment, the coefficient correction circuit in the high-order delay filter far from the first-stage delay filter By configuring the attenuator in the filter to be smaller than the attenuator in the coefficient correction circuit of the low-order delay filter, the partial autocorrelation coefficient can be determined more quickly and accurately. This will be explained using a mathematical formula.The output signal of the delay circuit 33 of the 01st stage delay filter 21 is x1(t), and the second input terminal 32
The input signal of Y, (t). The output signal of the delay circuit 41 is k
1(t), the output signals Ell)(t) and Elf(t) of the first and second adders 35 and 37 are expressed as follows, respectively.

The multiplier 42 receives the input signal y1(t
) by the output signal Elb(t) of the first adder 35, and the attenuator 43 attenuates the output of the multiplier 42 by a predetermined amount, so the attenuator 43 multiplies the input signal by Δ. If output, this output is Δ,・E,b(t)・y
1(t).

The adder 40 adds the output of the attenuator 43 and the output k1(t) of the delay circuit 41, and the delay circuit 41 uses this added output as the next output k1(t+1). Therefore, the following equation is obtained: where, if the gain Δ1 of the attenuator 43 is appropriately selected, K converges to a first-order partial autocorrelation coefficient.

The same applies to the second-stage and subsequent lattice-type delay filters, and generally, in the i-th stage delay filter, the i-th order partial autocorrelation coefficient Ki is obtained by successive approximation as follows.

Generally, the frequency spectrum of an audio signal changes from moment to moment.

For example, the speaking speed of Japanese is usually 5 to 10 syllables/second, and the frequency spectrum of the audio signal is 100 to 200.
changes into something completely different within seconds. Furthermore, the duration of a consonant is much shorter than that of a vowel, and therefore it is thought that the characteristics of speech change at least in several tens of milliseconds. In the above speech analysis device, the partial autocorrelation coefficient is determined by successive approximation as described above.

The convergence speed of this successive approximation becomes faster as the gain Δi of the attenuator 43 becomes larger, but the residual vibration error becomes larger and the distortion of the synthesized speech increases. On the other hand, if the gain Δ1 of the attenuator 43 is made smaller, the residual vibration error will naturally become smaller, but the convergence speed will become slower. Therefore, the gain Δ1 of the attenuator 43 must be selected so that the successive approximation can sufficiently follow the change in the voice spectrum and the residual error is small. The appropriate range of the value of Δi is quite small, and according to experiments, it is about 6 dB at most. The present invention improves performance by setting the gain Δi of the attenuator 43 in each stage of the delay filter to a value close to the optimum value. That is, since the correlation of the audio signal is removed each time it passes through the delay filters at each stage, the average power of the signal gradually decreases as the number of stages it passes through increases. That is, the average power of the input signal y1(t) of the high-order (n-th stage) delay filter is smaller than the average power of the input signal Ynl(t) of the low-order (m-th stage) delay filter. However, n>m. Also, the output Enb(t) of the first adder 35, which naturally falls into the n-th stage delay filter.
The average power of is smaller than the average power of the output Errt)(t) of the first adder 35 in the m-th delay filter.
Therefore, the gains ΔN, Δ . Assuming that they are equal, the following inequality holds between the outputs of the coefficient correction circuit 39 for each filter, that is, the average values of the correction amounts. Therefore, in order to make these average values approximately equal, the following can be done.

That is, the gain of the attenuator associated with the high-order delay filter may be made greater than the gain of the attenuator associated with the lower-order delay filter.

Comparing the attenuation amounts, the attenuation amount of the water reducer that is similar to the high-order delay filter may be made smaller than that of the low-order delay filter. Here △. ≧Δ． is N
, Δ depending on the value of m. =△． In some cases, at least one of them is △. 〉△． It means that. According to experiments, the average power of the input signal Y2(f) of the second stage lattice delay filter is equal to the input signal y1(f) of the first stage.
t), and the average power of the input signal Y3(t) of the third stage lattice delay filter is approximately 1/2 to 1/2 of the average power of Y2(t). /3.

Furthermore, the average power of the lattice delay filters in the fourth and subsequent stages does not change much. Therefore, the first stage input signal y1(t)
The dynamic range of -1 to +1, and the coefficients supplied from the delay circuit 41 to the first and second multipliers 3436 are -1 to +1.
+1, attenuator 4 in each stage delay filter
For example, the gain Δl of 3 is selected as follows. Or choose as follows:

By selecting the gain of the attenuator in this manner, the correction amount input to the coefficient generation circuit 38 of each stage of delay filter can be made almost constant, and the partial autocorrelation coefficient can be obtained quickly and accurately.

In the above embodiment, the multiplier 42 is used as a correlator in the coefficient correction circuit 39 of the lattice delay filter of each stage, and the second input signal Yi(t) and the output signal Eib of the first adder 35 are
(t) multiplication output is obtained.

However, as shown in FIG.
, (t) and the output signal Ei'f(t) of the second adder 37
It may also be obtained by the multiplier 44 of . In the case of this embodiment, the equation corresponding to the above equation (5) is as follows, but the same equation (6) is obtained. Furthermore, the multipliers 42 and 44
A simple code converter can also be used instead.

For example, as shown in FIG. 4, a code converter 45 is used to change the sign of the input signal yl to the second input terminal 32 according to the signs of the output signals E and b of the first adder 35, and the output The signal may be input to an attenuator 43.
In this case, the output of the coefficient correction circuit 39 in each lattice type delay filter 2i is △i−Sgn{Elb(t)}·Y,
(t). However, therefore, the value Ki(t+1) of the output signal Ki(t) of the coefficient generation circuit 38 at time t+1
) is expressed by the following formula.

Also in this case, assuming that the gains Δ1 and Δm of the attenuator 43 in the n-th stage delay filter and the m-th stage delay filter are equal, the following inequality corresponding to the above equation (5) is obtained.

In this case as well, in order to make these average values approximately equal, the above-mentioned equation (6) should hold true.

Furthermore, although not shown, the output signal Eib(t) of the first adder 35 is input to the second input terminal 32 as an input signal Yi(
A code converter that changes the sign of t) according to the sign may be used in place of the code converter 45 of the embodiment shown in FIG. 4, as long as the above-mentioned formula (6) holds.

*Furthermore, instead of the multiplier 44 in the embodiment of the present invention in FIG. Transducers can also be used. in this case,
The inequality corresponding to the above equation (5) is as follows, but in this case as well, the above equation (6) is obtained.In addition, conversely, the second equation
It is also possible to change the sign of the output signal Xi of the delay circuit 33 depending on the sign of the output signal Eif of the adder 3T.

Furthermore, as shown in FIG. 5, the output signal Xi(t) of the delay circuit 33 and the output signal Eif(t
), the output signal Elb(t) of the first adder 35 and the input signal Yi of the second input terminal 32
(t) multiplier 4T, and these multipliers 46, 4? It is also possible to use an adder 48 that performs the addition of Bye.

Although the above embodiments are all examples in which multiple stages of lattice-type delay filters are connected in cascade, the present invention processes the calculations in the filters in a time-sharing manner and sequentially extracts partial autocorrelation coefficients. It can also be applied to analytical devices.

In this case as well, the attenuation amount of the attenuation means in the higher-order stage may be calculated to be smaller than the attenuation amount of the attenuation means in the lower-order stage.

[Brief explanation of drawings]

Fig. 1 is a circuit diagram of a conventional speech analysis device, Fig. 2 is a circuit diagram of an embodiment of the present invention, Fig. 3 is a circuit diagram of another embodiment of the invention, and Fig. 4 is a circuit diagram of another embodiment. FIG. 5 is a circuit diagram of still another embodiment. 30...Audio signal input terminal, 31...
First input terminal, 32...Second input terminal, 3
3... Delay circuit, 34... First multiplier, 35... First adder, 36...
Second multiplier, 37...Second adder, 38.
... Coefficient generation circuit, 39 ... Coefficient correction circuit, 40 ... Multiplier, 41 ... Delay circuit,
42... Multiplier, 43... Attenuator.

Claims (1)

[Claims] 1 Delay means for delaying the input signal to the first input terminal by one sample time interval, and first and second multiplication for multiplying the output signal of the delay means and the input signal to the second input terminal by coefficients, respectively. means, first addition means for outputting a sum or difference signal between the output signal of the delay means and the output signal of the second multiplication means to a first output terminal, and an input signal to the second input terminal; a second addition means for outputting the sum or difference signal with the output signal of the first multiplication means to a second output terminal;
Coefficient generation means for obtaining coefficients to be applied to the first and second multiplication means; and coefficients for modifying the coefficients obtained by this means in a direction that reduces the output signal value of at least one of the first and second addition means. By cascading multiple stages of these means or by repeatedly using these means to perform signal processing at each stage, the sampled audio signal is sent to the first and second input terminals of the first stage. In a speech analysis device that extracts the input coefficients obtained by the coefficient generation means of each stage as partial autocorrelation coefficients, the coefficient correction means inputs the output signal of the first addition means and the input to the second input terminal. It consists of a correlation means for correlating at least one of the signal or the output signal of the delay means and the output signal of the second multiplication means, and an attenuation means for attenuating the output of the correlation means by a predetermined amount. A speech analysis device characterized in that the amount of attenuation of the attenuation means is configured to be smaller than the amount of attenuation of the attenuation means in lower-order stages.