JPH0777979A - Speech-operated acoustic modulating device - Google Patents

Abstract

PURPOSE:To adequately modulate a sound, generated by an acoustic generating device such as an electronic musical instrument, on the basis of a featured speech signal component. CONSTITUTION:A filter coefficient calculation part 302 extracts the frequencies and gains of plural peaks corresponding to a formant from the amplitude spectrum of a speech signal p(n) calculated by a spectrum calculation part 301 employing a linear predicting method and calculates plural filter coefficients. Those coefficients are held in a filter coefficient storage part 303. Then the filter coefficients held in the filter coefficient storage part 303 are used to constitute plural band-pass filters(BPF) 304, and a musical sound signal x(n) is inputted to those filters. Respective filter outputs Yj(n) (1<=j<=J) are accumulated by an accumulation part 305, whose accumulation output is outputted as a digital output musical sound signal z(n).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention can add the effect of changing the sound generated in response to the sound by modulating the sound generated from the sound generating device such as an electronic musical instrument with the sound. The present invention relates to an acoustic modulator.

[0002]

2. Description of the Related Art With the widespread use of sound generators such as electronic musical instruments, there has been a demand for a sound generator which allows a user to more easily and effectively reflect his or her will in output sound.

As one form of such an apparatus, there is a so-called vocoder capable of modulating a sound signal generated by a sound signal obtained by a user's utterance.

FIG. 14 is an explanatory view of the prior art of such a vocoder device. First, a person sings or speaks to generate a digital audio signal p (n), while a digital tone signal x (n) is generated based on a performance operation on an electronic musical instrument or the like.

Then, a DSP (Digital Signal Processo)
In the digital signal processing device such as r), the digital tone signal x (n) and the digital voice signal p (n) are processed by the plurality of band-based conversion units 1401 of # 1 to #N, which are digital signal processing. To be executed.

That is, first, the band-pass filter section (BPF section) 1403 of each band-specific conversion section 1401 band-limits the digital musical sound signal x (n) into a plurality of different frequency bands. It is divided into (n).

At the same time, the bandpass filter section (BPF section) 1402 of each band conversion section 1401
The digital voice signal p (n) is divided into a plurality of voice signal components Qj (n) band-limited within a plurality of different frequency bands. Next, the envelope extraction unit 1406 of each band conversion unit 1401 extracts the envelope signal Rj (n) of the audio signal component Qj (n).

Further, the multiplication unit 1 of each band conversion unit 1401
At 404, the envelope signal Rj (n) of the audio signal component of each frequency band is added to the tone signal component Yi (n) of each frequency band.
By multiplying by, the tone signal component Yi (n) is modulated by the envelope signal Rj (n) of the voice signal component for each frequency band.

Then, the output of the multiplication unit 1404 of each band conversion unit 1401 is accumulated in the accumulation unit 1405, and the accumulated output power is the digital output musical tone signal z (n).
Is output.

[0010]

However, in the above-mentioned prior art, the frequency characteristics of the bandpass filter sections 1402 and 1403 in each band conversion section 1401 are fixed, and the digital audio signal p (n ) Frequency characteristics of the digital tone signal x (n)
However, there is a problem in that it cannot be reflected in the frequency characteristics of.

An object of the present invention is to make it possible to properly modulate an acoustic signal based on a characteristic audio signal component.

[0012]

The present invention firstly has spectrum calculating means for calculating a frequency intensity envelope component (amplitude spectrum or power spectrum) corresponding to a vocal tract spectrum from a voice signal. This spectrum calculation means is determined based on, for example, a linear prediction analysis means for calculating a linear prediction coefficient by executing a linear prediction analysis on the audio signal, and a linear prediction coefficient calculated by the linear prediction analysis means. And an all-pole frequency intensity component calculating means for calculating an all-pole frequency intensity component as a frequency intensity envelope component. Alternatively, the spectrum calculation means performs a Fourier transform on the audio signal to calculate the frequency intensity component, and the frequency intensity envelope calculated by smoothing the frequency intensity component calculated by the Fourier transform means. And a smoothing means for calculating the component.

Next, it has a peak calculating means for calculating frequency characteristics of a plurality of effective peaks from the frequency intensity envelope component calculated by the spectrum calculating means. This effective peak corresponds, for example, to a formant in the vocal tract characteristic of speech. As the frequency characteristic of the peak, for example, the frequency of the peak and the intensity value are calculated.

Next, there is provided filter coefficient calculation means for calculating a plurality of sets of filter coefficients having transfer characteristics corresponding to the frequency characteristics of the effective peaks calculated by the peak calculation means, corresponding to a plurality of effective peaks. Have.

Further, a plurality of sets of filter coefficients calculated by the filter coefficient calculating unit are used to independently perform a plurality of filtering processes corresponding to each set of filter coefficients on the acoustic signal. Have.

Further, there is provided accumulating means for accumulating outputs of a plurality of filtering processing means executed by the filtering processing means and outputting as an output acoustic signal.

[0017]

The frequency characteristics of a plurality of effective peaks corresponding to the voice formant and the like are calculated by the peak calculating means from the frequency intensity envelope component of the audio signal calculated by the spectrum calculating means, and further by the filter coefficient calculating means, A set of filter coefficients having a transfer characteristic corresponding to the frequency characteristic of such an effective peak is calculated.

The filtering processing means independently performs, on the acoustic signal, a plurality of filtering processes that emphasize the frequency characteristics of the plurality of effective peaks, using the set of filter coefficients described above. Then, the outputs of the plurality of filtering processing means are accumulated by the accumulating means and output as an output acoustic signal.

As a result, in the acoustic signal, the frequency band corresponding to the effective peak such as the formant of the audio signal which changes from moment to moment is properly emphasized.

[0020]

Embodiments of the present invention will now be described in detail with reference to the drawings. <First Embodiment> Configuration FIG. 1 is a block diagram showing the overall configuration of a first embodiment when the present invention is applied to a keyboard instrument.

When the performer presses a key on the keyboard 109 or performs a switch operation such as tone color setting or various effect setting by the function switch 110, the performance information is sent to the CPU 101 via the bus 117.

The CPU 101 executes a program stored in the ROM 102 and processes performance information by using the RAM 103 as a work memory. The performance information processed in this way, such as note-on / off, velocity, tone color setting data, etc., is transmitted via the bus 117 to the tone generation circuit 1
It is sent to 07. This circuit generates a musical tone signal according to performance information. As the tone generation method in the tone generation circuit 107, for example, a PCM method, a modulation method, an overtone addition method, or the like is used.

Next, a digital tone signal generated by the tone generating circuit 107 (hereinafter simply referred to as tone signal) x (n)
Is input to the DSP 104 via the bus 118 dedicated to the tone signal.

On the other hand, when the performer sings or speaks into the microphone 111, the analog audio signal obtained via the microphone amplifier 112 is converted into the low-pass filter 1.
It is input to the A / D converter 114 via 13 and converted into a digital audio signal (hereinafter simply referred to as an audio signal) p (n) and input to the DSP 104. Note that the analog audio signal is not from the microphone, but the line input terminal LINE
It may be input from IN.

The DSP 104 performs a filter coefficient calculation process based on the audio signal p (n) input from the A / D converter 114, and a band pass filtering process having a characteristic determined based on the filter coefficient. , The tone signal x (n) input from the tone generation circuit 107. As a result, the tone signal x (n) is modulated by the audio signal p (n).

The digital output musical tone signal z (n) obtained by the modulation processing in the DSP 104 is sent to the D / A converter 108 via the dedicated bus 119, where it is converted into an analog output musical tone signal and is amplified. Speaker 1 through 115
Sound is emitted from 16.

Next, the configuration and function of the DSP will be described. Configuration of DSP FIG. 2 is an overall configuration diagram of the DSP 104.

The interface 201 is a bus 117 connected to the CPU 101, a bus 118 connected to the tone generation circuit 107, a bus connected to the A / D converter 114, and a bus 119 connected to the D / A converter 108. And each bus is connected to a circuit inside the DSP.

The operation ROM 202 includes the DSP 1
04 is a ROM that stores a microprogram that defines the overall operation, and the corresponding program instruction is read based on the specified address from the address counter 203. The CPU 101 of FIG. 1 sets the data in the address counter 203, and thereby the operation ROM
The address counter 203 is instructed which program is to be read from 202 and to execute the modulation processing described later.

The output of the operation ROM 202 is also given to the decoder 204, and outputs various control signals to each circuit in the DSP 104 to perform a desired operation. On the other hand, the filter coefficient RAM 105 and the work RAM 106 of FIG. 1 are connected to the internal bus of the DSP 104.

Then, the tone signal x (n), the voice signal p (n), etc. are supplied to the DSP 104 or input / output to / from the work RAM 106 via this internal bus. The register group 207 is composed of a plurality of registers for temporarily storing data being calculated, and is connected to each input / output terminal of the multiplier 205 or the adder / subtractor 206 via an internal bus. Then, in order to realize the judgment processing based on the calculation result (comparison result etc.) from the adder / subtractor 206, the address counter 203 is transferred to the address counter 203 via the flag register 208.
A flag signal indicating the judgment result is transmitted.

The address of the address counter 203 is changed according to the output of the flag register 208, and the program instruction is read from the operation ROM 202 according to the address. In this way, the judgment processing is realized. Operating Principle Next, the operating principle of the first embodiment implemented as the function of the DSP 104 will be described.

FIG. 3 shows the DSP 10 in the first embodiment.
It is a functional block diagram of FIG. In the first embodiment, first,
The spectrum calculation unit 301 based on the linear prediction method executes a speech analysis based on a method called a linear prediction method on the speech signal p (n) at a predetermined time interval (10 to 30 msec interval), so that the speech signal p Calculate the spectrum of (n).

The linear predictive analysis method utilizes the fact that there is a high correlation between the sample values of the speech waveform, and the present speech signal p (n)
Is expressed as the sum of the predicted value, which is a linear combination of the past p (about 10) samples, and the current error signal, and is expressed by the following equation.

[0035]

[Equation 1]

In the equation (1), a (i) is called a linear prediction coefficient (LPC coefficient) and e (n) is called a prediction error signal. The mean square error of this prediction error signal e (n) in a certain section is Obtaining a linear prediction coefficient under the condition of minimizing it is called Linear Prediction Coding (LPC).

Next, let us consider the meaning of the linear prediction analysis in the frequency range. First, it can be assumed that the prediction error signal e (n) is an impulse train G · u (n) with a gain G in the voiced sound section that occupies most of the speech signal. According to this assumption, Equation 1 is expressed by the following equation.

[0038]

[Equation 2]

Then, if the z-transformation of both sides of the equation 2 is taken, the following equation is obtained.

[0040]

[Equation 3]

Therefore, the following equation is obtained.

[0042]

[Equation 4]

[0043]

[Equation 5]

[0044]

[Equation 6]

[0045]

[Equation 7]

From these four equations, the impulse u
The output after inputting (n) into the transfer function S (z) is p (n), and the transfer function S (z) is a P-order equation A (A) determined by the linear prediction coefficient a (i). z) p given as the root of = 0
It shows that it becomes an all-pole model with individual poles. That is, performing the linear predictive analysis on the speech signal is equivalent to approximating the speech signal with the all-pole model.

If the speech signal is approximated by an all-pole model, that pole corresponds to the formant of the speech spectrum. Then, the amplitude spectrum | S (z) | at an arbitrary frequency f based on the all-pole model determined based on the linear prediction coefficient a (i) is obtained by converting the z transform into a complex frequency in the formula (6). The following equation 8 can be calculated by substituting the following equation 8 with the equation 7 into the following equation 9 modified from the equation 5.

[0048]

[Equation 8]

[0049]

[Equation 9]

Here, j is a complex number unit, π is a circular constant, and f
max is the Nyquist frequency. In the amplitude spectrum | S (z) | calculated in this way, the influence of the harmonic structure attributed to the pitch of the voiced sound source of the voice signal p (n) is removed, and the peak due to the formant of the vocal tract spectrum is accurate. Is approximated by.

The audio signal p calculated as described above
An example of the amplitude spectrum | S (z) | of (n) is shown in FIG.
From this figure, the order P of the linear prediction coefficient a (i) is 10 to 20.
It can be seen that the peak due to the formant (indicated by an arrow in the figure), which is a characteristic of the audio signal p (n), is accurately represented in the case of the degree.

Therefore, in the first embodiment, the filter coefficient calculation unit 302 of FIG. 3 uses the linear prediction method to calculate the amplitude spectrum of the audio signal p (n) at each predetermined time interval calculated by the spectrum calculation unit 301. From | S (z) |, the frequency Fj and the gain (amplitude value) Gj (1 ≦ j ≦) of J sets of peaks, which are also shown as arrows in FIG.
J) are extracted, and the filter coefficients for the J sets of band-pass filters (BPF) # 1 to #J each having the characteristics shown in FIG. 5 are calculated at the above-mentioned predetermined time intervals.

The thus calculated filter coefficients for the J sets of bandpass filters of # 1 to #J are held in the filter coefficient storage unit 303. Then, using the filter coefficients held in the filter coefficient storage unit 303, J band-pass filters (BPF) 304 of # 1 to #J each having the characteristics shown in FIG. 5 are configured. The musical tone signal x (n) is input to.

Finally, J band pass filters 304
Output Yj (n) (1.ltoreq.j.ltoreq.J) is accumulated in the accumulator 305, and the accumulated calculation force is output as a digital output musical tone signal z (n).

With the above processing, the tone signal x (n) is properly emphasized in the frequency band corresponding to the formant of the voice signal p (n) which changes from moment to moment. Specific Operation First, the operation of the musical sound generating circuit 107 to generate the musical sound signal x (n) based on the control operation executed by the CPU 101 is not particularly related to the present invention, and therefore its details are omitted.

FIG. 6 is an operation flowchart relating to the vocoder process executed by the DSP 104 (FIGS. 1 and 2) in accordance with the microprogram stored in the operation ROM 202. The spectrum calculation process, the filter coefficient calculation process, and the bandpass filtering. Including processing.

By the processing according to this operation flow chart, the spectrum calculation unit 301 and the filter coefficient calculation unit 302 by the linear prediction method of FIG. 3 are sampled for each sampling period common to the tone signal x (n) and the voice signal p (n). , J of # 1 to #J
By executing the processing corresponding to the band pass filter (BPF) 304 and the accumulating unit 305 by the time division processing, the digital output musical tone signal z (n) is obtained for each sampling period, and the processing shown in FIG. It is output to the D / A converter 108.

First, in step S601, the contents of the filter coefficient RAM 105, the work RAM 106 of FIG. 1 and the register group 207 in the DSP 104 of FIG. 2 are initialized.

Next, in step S602, the A / D of FIG.
In the converter 114, it is monitored whether the A / D conversion executed in each cycle corresponding to the sampling frequency fs is completed.

When the A / D conversion is completed and the determination in step S602 is YES, in step S603, the A / D converted audio signal p (n) is transferred to the interface 201 (FIG. 2).
Variable p (n) in the work RAM 106
Are sequentially stored in. Along with this, the tone generation circuit 107
The tone signal x (n) input from (FIG. 1) is also fetched from the interface 201 and sequentially stored in the variable x (n) in the work RAM 106. The work RAM 106
In the three variables x (n), x (n−1), and x (n−)
In each of 2), continuous tone signals of the present sample and the past two samples are stored.

Next, in step S604, a spectrum calculation process is executed. A detailed operation flowchart of this processing is shown in FIG. By this processing, the function of the spectrum calculation unit 301 according to the linear prediction method of FIG. 3 described above is realized.

Note that this spectrum calculation processing is shown to be executed each time a sample of the audio signal p (n) is input in the operation flowchart of FIG. 6, but in practice, in particular, Based on a time control process (not shown), it is executed once every predetermined time (10 to 30 msec).

The processing of steps S701 to S703 in FIG. 7 is processing for calculating the linear prediction coefficient a (i) described in the operating principle of the first embodiment. for that reason,
As described above in the explanation of the operating principle of the first embodiment, first, it is necessary to calculate the autocorrelation function r (i) of the audio signal p (n) at every predetermined time interval (10 to 30 msec interval). .

Therefore, in step S701, the audio signal p (n) for N samples of a predetermined time length (10 to 30 msec) input in the past, which is held in the work RAM 106, is cut out, and the Hamming window is applied thereto. Is multiplied by the window function. This is to change the amplitude of the audio signal p (n) so that the amplitudes of the cut-out parts at the beginning and the end gradually approach 0, thereby eliminating the adverse effects of the signal components of those parts on the spectrum analysis. Processing.

The number of samples N is determined depending on the sampling frequency fs and the predetermined time length, but is generally about N = 256 to 1024 samples. Next, in step S702, an autocorrelation function r (i) (0 ≦ i ≦ P, P is a linear prediction is defined by the following equation for the N samples of the audio signal p (n) multiplied by the window function. Coefficient order)
Is calculated.

[0066]

[Equation 10]

Then, in step S703, a linear prediction coefficient a (i) is obtained by solving a P-element simultaneous linear equation called a normal equation shown by the following equation using the above-mentioned autocorrelation function r (i). (1 ≦ i ≦ P) is calculated.

[0068]

[Equation 11]

This normal equation can be solved by a general numerical calculation method of simultaneous equations such as the Gauss-Seidel method and also by a method called Levinson's algorithm. Since these methods are known methods, detailed description thereof will be omitted. An example of an algorithm for obtaining a linear prediction coefficient is 68 in the document "<< Electronic Science Series >> 86 Computer Speech Processing" (Take Yasuiin / Masayuki Nakajima) (Industry Press).
Pp. 70-70 and the like.

Next, in steps S704 and S705, the amplitude spectrum | S (z) is calculated using the linear prediction coefficient a (i) based on the equations (5), (7) and (8) described above.
| Is calculated.

First, in the amplitude spectrum characteristic | S (z) |, the change width Δf of the frequency f to be calculated is defined by the following equation.

[0072]

[Equation 12]

Here, the Nyquist frequency fmax and the sampling frequency fs in the equation (8) have the relationship shown by the following equation.

[0074]

[Equation 13]

From the equations (13) and (12), the following equation is established.

[0076]

[Equation 14]

Therefore, the arbitrary frequency f is expressed by the following equation.

[0078]

[Equation 15]

From the equations 13 and 14, the above equation 8 is obtained.
The formula is represented by the following formula.

[0080]

[Equation 16]

That is, the amplitude spectrum | S (z) | can be calculated by using the linear prediction coefficient a (i) based on the above-described equations (7), (9) and (16). .
Here, the value of Expression 16 at each frequency m · Δf (0 ≦ m ≦ N / 2) can be expressed by the following expression.

[0082]

[Equation 17]

These values are the linear prediction coefficients a (i) (1 ≦
i ≦ P) and the value 0, the real part data re (m) represented by the following formula 18 and the imaginary part data i in which all values are 0
For m (m) (1 ≦ m ≦ N), the fast Fourier transform (F
FT) and the resulting real part output data RE (m) and imaginary part output data IM (m) (1 ≦ m ≦
N / 2) can be calculated by calculating the following equation (19). Since the FFT operation is a well-known operation technique, its details are omitted.

[0084]

[Equation 18]

[0085]

[Formula 19]

As a result, from the DC component to the Nyquist frequency f
Amplitude spectrum value consisting of N / 2 samples for each change width Δf up to max | S (z) | = S (m) (1 ≦ m ≦ N / 2)
Can be calculated based on the equations (7) and (9) and the equation (19).

When the above-mentioned calculation processing is associated with the actual processing of FIG. 7, it becomes as follows. First, step S703
Linear prediction coefficient a (i) (1 ≦ i ≦ P) calculated by
Real part data re expressed by the following equation 18 consisting of
(m), and imaginary part data im (m) (1 ≦
For m ≦ N), a fast Fourier transform (FFT) operation is performed in step S704.

Next, in Step 705, Step S70
1 for the real part output data RE (m) and the imaginary part output data IM (m) (1 ≦ m ≦ N / 2) calculated in FIG.
Equation 9 is calculated. Further, the autocorrelation function r (i) (0 ≦ i ≦ P) calculated in step S702 and step S7
The gain G of Expression 7 is calculated based on the linear prediction coefficient a (i) (1 ≦ i ≦ P) calculated in 03. And the number 1
Using the calculation results of Equation 9 and Equation 7, the amplitude spectrum value S consisting of N / 2 samples for each change width Δf from the DC component to the Nyquist frequency fmax based on Equation 9 described above | S
(z) | = S (m) (1 ≦ m ≦ N / 2) is calculated.

The amplitude spectrum value S (m) (1 ≦ m ≦ N / 2) obtained as a result of the above spectrum calculation processing is held in the work RAM 106. Then, step S of FIG.
At 605, a filter coefficient calculation process is executed.

A detailed operation flowchart of this processing is shown in FIG. By this processing, the function of the filter coefficient calculation unit 302 of FIG. 3 described above is realized. Note that this filter coefficient calculation processing is shown to be executed each time a sample of the audio signal p (n) is input in the operation flowchart of FIG. 6, but in reality it is synchronized with the spectrum calculation processing. Then, it is executed once every predetermined time (10 to 30 msec).

In this process, the value of the register m in the register group 207 (FIG. 2) in the DSP 104 is changed to the value in step S8.
After being initialized to 2 in 02, +1 in step S810
Incremented by N, step S811 returns N
The peak search process of step S803 is executed until it is determined that the value exceeds / 2.

That is, in step S803, the register m
Amplitude spectrum value S at discrete frequency position m corresponding to the value of
It is determined whether or not (m) is larger than the amplitude spectrum values S (m-1) and S (m + 1) of discrete frequency positions m-1 and m + 1 before and after that, respectively. The peak frequency position indicated as, etc. is searched. When the peak frequency position m is extracted because the determination in step S803 is YES, the following steps S804-
By executing the process of S809, the filter coefficient for the bandpass filtering process corresponding to the j-th peak is calculated. The value of the register j in the register group 207 (FIG. 2) corresponding to the number j is set to 1 in step S801, and then the steps S804 to S8.
Every time the processing of 09 is executed, +1 in step S809.
To be done.

First, in step S804, step S8
The peak frequency position m when the judgment of 03 is YES
The actual peak frequency Fj [Hz] corresponding to is calculated. Now, the actual frequency f = corresponding to the discrete frequency position m =
Fj [Hz] can be calculated by the following equation from the equations 12 and 15 described above.

[0094]

[Equation 20]

Next, in step S805, step S8
The peak frequency position m when the judgment of 03 is YES
The amplitude spectrum value S (m) corresponding to is the peak gain G
Calculated as j.

In steps S806 to S808, the band pass having the frequency characteristic shown in FIG. 9 and the functional configuration shown in FIG. 10 is based on the peak frequency Fj and the peak gain Gj calculated in steps S804 and S805. The filter coefficient for the filtering process is calculated. The bandwidth Q is a fixed value for easy calculation.

First, in FIG. 10, 1001, 100
The portion indicated by Z ^{-1 in} 2 represents a delay unit that delays one sampling clock cycle. 1003, 10
Reference numeral 04 is an adder. In addition, reference numerals 1005 to 1009 denote multiplying units, which respectively filter coefficients A0j and A1.
Multiplies j, A2j, B1j, and B2j.

As shown in FIG. 9, the bandpass filtering process having the functional configuration of FIG. 10 has a center frequency equal to the peak frequency Fj, a gain equal to the peak gain Gj, and a bandwidth with a fixed value Q. If it has a certain frequency characteristic, the filter coefficients A0j, A1j,
A2j, B1j, and B2j are expressed by the following formula 21 and formula 22.
Using the coefficient Kj and the coefficient TXj calculated by the formula, the calculation is performed based on the following formulas 23 to 27.

[0099]

[Equation 21]

[0100]

[Equation 22]

[0101]

[Equation 23]

[0102]

[Equation 24]

[0103]

[Equation 25]

[0104]

[Equation 26]

[0105]

[Equation 27]

Then, based on the peak frequency Fj and the peak gain Gj calculated in steps S804 and S805, it is steps S806 and S807 that calculate the above-mentioned formulas 21 and 22, and the results thereof are shown. Based on the equations (23) to (27), the filter coefficients A0j, A
In step S808, 1j, A2j, B1j, and B2j are calculated. These calculated filter coefficients are held in the filter coefficient RAM 105.

After the above processing, the value of the register j indicating the peak number is incremented by 1 in step S809. Then, through steps S810 and S811, step S
At 803, the next peak frequency position is retrieved.

As a result of repeating the above processing, step S8.
In 11, it is determined that the value of the register m exceeds N / 2,
When the determination is YES, the total peak number J is calculated as (current value of register j-1) in step S812, and the result is stored in the register J in the register group 207 (FIG. 2). Then, the filter coefficient calculation process of step S605 of FIG. 6 shown in the operation flowchart of FIG. 8 ends.

In steps S606 to S611 of FIG.
The above-mentioned J stored in the filter coefficient RAM 105 is calculated once every predetermined time (10 to 30 msec).
Each sample x of the tone signal using a set of filter coefficients
For (n), J times of band-pass filtering processes # 1 to #J having the characteristics shown in FIG. 5 are sequentially executed.

Now, the bandpass filtering process having the functional configuration of FIG.

[0111]

[Equation 28]

It can be realized by the discrete calculation processing. Therefore, in FIG. 6, the value of the register j indicating the number j of the bandpass filtering process is the value in step S60.
After being initialized to 1 in 6 in step S610, j is held in the filter coefficient RAM 105 in step S607 until it is determined in step S609 that the total number J indicated by the register J is exceeded while incrementing by 1 in step S610. Filter coefficients A0j, A1j,
Using A2j, B1j, and B2j, the above equation 28
The bandpass filtering process represented by the formula is executed.

Then, in step S608, the output Yj (n) of the j-th bandpass filtering process is accumulated in the digital output musical tone signal z (n) held in the work RAM 106.

As a result of the above processing, if the determination in step S609 is YES, the output Yj (n) (1.ltoreq.j.ltoreq.J) of J times of band pass filtering processing is performed in step S611.
The digital output musical tone signal z (n) obtained by accumulating
It is output from the work RAM 106 to the D / A converter 108 (FIG. 1) via the interface 201 (FIG. 2) and the bus 119 (FIG. 1) at a timing synchronized with the sampling clock.

After that, the process returns to step S602 again, and the above-described series of processes for sampling the next voice signal p (n) and tone signal x (n) is repeated. As described above, in the tone signal x (n), the frequency band corresponding to the formant of the voice signal p (n), which changes moment by moment, is properly emphasized. <Second Embodiment> Next, a second embodiment of the present invention will be described. Configuration The configuration of the second embodiment is exactly the same as the configuration of FIGS. 1 and 2 in the first embodiment. Operating Principle In the above-described first embodiment, as described above, the amplitude spectrum calculated based on the linear prediction coefficient a (i) can accurately approximate the peak due to the formant, which is a feature of the voice signal (FIG. 4). Utilizing this, the frequencies and gains of a plurality of peaks are extracted from the amplitude spectrum, a plurality of sets of filter coefficients having characteristics as shown in FIG. 5 are calculated, and a plurality of sets of filter coefficients using those filter coefficients are calculated. Bandpass filtering processing has been executed.

On the other hand, in the second embodiment described below, as shown in the functional block diagram of the DSP 104 of FIG. 11, first, the spectrum calculation unit 1101 by Fourier transform is operated for a predetermined time (10 to 30 msec). By performing Fourier transform by the FFT method or the like on the minute voice signal p (n), an amplitude spectrum including a harmonic structure attributable to the pitch of the voiced sound source of the voice signal p (n) is calculated. Then, the Fourier transform spectrum calculation unit 1101 removes the influence of the harmonic structure by smoothing the amplitude spectrum. The amplitude spectrum thus calculated can almost express the peak due to the formant, which is a feature of the audio signal p (n), as shown in FIG.

The subsequent calculation of the filter coefficient and the construction of the bandpass filter using the same are the same as in the case of the first embodiment. By the above processing, the first
The filter coefficient can be calculated with a smaller amount of calculation than in the case of the embodiment. Specific operation A specific operation of the second embodiment based on the above-described operation principle will be described below.

First, the overall function of the vocoder processing executed by the DSP 104 is the same as that shown in the operation flowchart of FIG. 6 in the first embodiment. In FIG. 6, the second embodiment differs from the first embodiment only in the spectrum calculation processing in step S604.

A detailed operation flowchart of this processing is shown in FIG. By this processing, the function of the spectrum calculation unit 1101 by the Fourier transform of FIG. 11 described above is realized.

Note that this spectrum calculation processing is shown to be executed each time a sample of the audio signal p (n) is input in the operation flowchart of FIG. 6, but in practice, in particular, Based on a time control process (not shown), it is executed once every predetermined time (10 to 30 msec).

First, in step S1301, for the same purpose as in step S701 of FIG. 7 in the above-described first embodiment, for example, a predetermined time length (10 to 30 mse) input in the past which is held in the work RAM 106.
The audio signal p (n) for N samples in c) is cut out and multiplied by a window function such as a Hamming window.

Next, in step S1302, the real part data re (m) in which the audio signal p (n) for N samples multiplied by the window function is set, and the imaginary part data im in which all values are 0. A fast Fourier transform (FFT) is calculated for (m) (1 ≦ m ≦ N).

Subsequently, in step S1303, the real part output data RE (m) and the imaginary part output data IM (m) of the first half obtained by the FFT operation of step S1302 (1≤
For m ≦ N / 2), the following formula 29 is calculated to calculate the amplitude spectrum value S (m) ′ (1) consisting of N / 2 samples for each change width Δf from the DC component to the Nyquist frequency fmax. ≦ m ≦ N / 2) is calculated.

[0124]

[Equation 29]

Then, in step S1304, with respect to the amplitude spectrum value S (m) '(1≤m≤N / 2),
The smoothing process is executed. This smoothing process is, for example, a process of calculating a moving average in the frequency direction of about several tens of samples. As a result, the amplitude spectrum value S (m) (1 ≦ m where the influence of the harmonic structure due to the pitch of voiced sound is reduced
≦ N / 2) is calculated.

As the processing described above, step S in FIG.
The spectrum calculation processing of 604 is subsequently performed. After that, in the filter coefficient calculation process of step S605 of FIG. 6, the above-described amplitude spectrum value S (m) (1 ≦ m ≦ N / 2)
On the other hand, the same processing as that of FIG. 8 described in the first embodiment is executed, and as a result, a predetermined time (10 to 30 ms) is reached.
ec) once in the filter coefficient RAM 105.
A set of filter coefficients is obtained.

Then, similarly to the case of the first embodiment, in steps S606 to S611 of FIG. 6, each sample x () of the tone signal is generated by using the above-described J sets of filter coefficients obtained in the filter coefficient RAM 105. For n), the bandpass filtering processes # 1 to #J having the characteristics shown in FIG. 5 are performed J times in sequence. <Other Embodiments> In the above-described first or second embodiment, when the filter coefficient is calculated from the amplitude spectrum value of the audio signal p (n), the bandwidth Q of each peak is set to a fixed value. If the processing time allows, these may be calculated individually from the amplitude spectrum. In this case, the bandwidth Q is calculated as the interval between two frequencies before and after the peak frequency whose amplitude value is equal to 1/2 of each peak value.

Since the amplitude spectrum value of the audio signal p (n) becomes smaller as the frequency becomes higher, in order to sufficiently modulate the tone signal x (n) at a high frequency,
The high-frequency component of the audio signal p (n) may be emphasized in advance by a high-frequency emphasis filter or the like, and then the amplitude spectrum may be calculated by the linear prediction method or the Fourier transform.

Further, the amplitude spectrum can be calculated by various methods other than the linear prediction method and the Fourier transform method. For example, the amplitude envelope spectrum in which the influence of the harmonic structure due to the pitch of voiced sound is removed by the cepstrum method. It can also be calculated.

[0130]

According to the present invention, a plurality of filtering processes for emphasizing frequency characteristics of a plurality of effective peaks corresponding to a voice formant calculated from a frequency intensity envelope component of a voice signal are applied to an acoustic signal. Can be performed against. Therefore, in the acoustic signal, it is possible to properly emphasize the frequency band corresponding to the effective peak such as the formant of the audio signal, which changes from moment to moment.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

FIG. 2 is an overall configuration diagram of a DSP.

FIG. 3 is a functional block diagram of a DSP according to the first embodiment.

FIG. 4 is a diagram showing an example of a spectrum calculated based on a linear prediction method.

FIG. 5 is a characteristic diagram of a bandpass filter.

FIG. 6 is an operation flowchart of vocoder processing.

FIG. 7 is an operation flowchart of spectrum calculation processing in the first embodiment.

FIG. 8 is an operation flowchart of a filter coefficient calculation process.

FIG. 9 is a frequency characteristic diagram of bandpass filtering processing.

FIG. 10 is a functional configuration diagram of a bandpass filtering process.

FIG. 11 is a functional block diagram of a DSP according to the second embodiment.

FIG. 12 is a diagram showing an example of a spectrum calculated based on the Fourier transform method.

FIG. 13 is an operation flowchart of a spectrum calculation process in the second embodiment.

FIG. 14 is an explanatory diagram of a conventional technique.

[Explanation of symbols] [Explanation of symbols]

 101 CPU 102 ROM 103 RAM 104 DSP 105 Filter Coefficient RAM 106 Work RAM 107 Musical Sound Generation Circuit 108 D / A Converter 109 Keyboard 110 Function Switch 111 Microphone 112 Microphone Amplifier 113 Low Pass Filter 114 A / D Converter 115 Amplifier 116 Speaker 117, 118, 119 Bus 201 Interface 202 Operation ROM 203 Address counter 204 Decoder 205 Multiplier 206 Adder-subtractor 207 Register group 208 Flag register 301 Linear prediction method spectrum calculation unit 302 Filter coefficient calculation unit 303 Filter coefficient storage unit 304 Bandpass filter (BPF) ) 305 Accumulation unit 1101 Spectrum calculation unit by Fourier transform

Claims (3)

[Claims] 1. A spectrum calculating means for calculating a frequency intensity envelope component corresponding to a vocal tract spectrum from a voice signal, and frequency characteristics of a plurality of effective peaks from the frequency intensity envelope component calculated by the spectrum calculating means. A filter for calculating a plurality of sets of peak calculating means for calculating and a set of filter coefficients having transfer characteristics corresponding to the frequency characteristics of the effective peaks calculated by the peak calculating means, corresponding to the plurality of effective peaks. Using the coefficient calculation means and the plurality of sets of filter coefficients calculated by the filter coefficient calculation means, a plurality of filtering processes corresponding to the respective sets of filter coefficients are independently executed on the acoustic signal. Filtering processing means, and the plurality of filtering processing executed by the filtering processing means Voice control acoustic modulation apparatus characterized by having a accumulating means for outputting as an output sound signal accumulating the output of the stage, the. 2. The spectrum calculation means performs linear prediction analysis on the speech signal to calculate a linear prediction coefficient, and the linear prediction analysis means calculates the linear prediction coefficient. The voice control acoustic modulation according to claim 1, further comprising: an all-pole frequency intensity component calculating unit that calculates an all-pole frequency intensity component determined based on a coefficient as the frequency intensity envelope component. apparatus. 3. The spectrum calculating means executes a Fourier transform on the audio signal to calculate a frequency intensity component, and a smoothing of the frequency intensity component calculated by the Fourier transforming means. The sound control acoustic modulator according to claim 1, further comprising: a smoothing unit that calculates the frequency intensity envelope component by converting the frequency intensity envelope component.