JPH10187180A - Musical sound generating device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize a musical sound generating device in which a voice- synthesized human voice sound can be musical sound-formed as a natural singing voice. SOLUTION: Partial auto correlation (PARCOR) coefficients K1-Kn generated by a PARCOR analyzing system 2 are stored in a PARCOR coefficient storage part 4, and residual waveform is stored in a residual waveform data memory 3. A control circuit reads the corresponding PARCOR coefficients K1-Kn from the PARCOR coefficient storage part 4 according to performance data, and a pitch converting means 5 pitch-converts the residual waveform read from the residual waveform data memory 3 according to the performance data. Then, a PARCOR synthesizing system 6 PARCOR-synthesizes the pitch-converted residual waveform data according to the PARCOR coefficients K1-Kn, and generates a voice with a desired pitch. Thus, the voice-synthesized human voice sound can be musical sound-formed as a natural singing voice.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice synthesizing device, and more particularly to a tone generating device capable of generating a natural singing voice.

[0002]

2. Description of the Related Art Conventionally, as a method of synthesizing a human voice based on feature parameters extracted by voice analysis, channel vocoder, linear prediction, PARCOR (Percall)
A technique called is known. These speech synthesis techniques can convert the analyzed speech into a smaller amount of information,
In other words, it focuses on analyzing speech, converting it into the form of feature parameters, and compressing the amount of information excluding redundant components not related to the meaning of words. He did not consider applying the human voice to the formation of musical sounds.

In such a situation, the channel vocoder has a simple structure and is suitable for real-time analysis and synthesis. Therefore, the channel vocoder has been applied to a tone generator which synthesizes a tone based on the power spectrum envelope of speech extracted by a filter bank. . However, in the channel vocoder, high-quality sound synthesis was not achieved due to the limitation of the number of band-pass filter stages constituting the filter bank and problems such as inability to synthesize consonants.

[0004]

On the other hand, in the conventional tone generator using the waveform memory reading method, the simplest method is to store the sampled human voice in the waveform memory and read and reproduce it at the sampling pitch. Although it is possible to generate a high-quality human voice in a simple form, if you try to read and play it at a pitch different from the pitch at the time of sampling, the formant frequency of the human voice will change according to the conversion pitch amount However, there is a problem that a natural singing voice cannot be generated.

Accordingly, the present invention has been made in view of such circumstances, and has as its object to provide a musical sound generating apparatus capable of forming a musical sound from a synthesized human voice as a natural singing voice.

[0006]

In order to achieve the above object, according to the first aspect of the present invention, an audio signal sampled in time series is subjected to Percoll analysis to generate a Percoll coefficient group and a residual waveform. Analysis means, storage means for storing a Percoll coefficient group and a residual waveform generated by the voice analysis means, coefficient reading means for reading out a corresponding Percoll coefficient group from the storage means according to performance information, Pitch conversion means for converting the stored residual waveform into a pitch in accordance with the performance information; and performing Percoll synthesis on the residual waveform pitch-converted by the pitch conversion means in accordance with a Percoll coefficient group read by the coefficient reading means. And voice synthesis means for generating voice of a desired pitch.

According to a second aspect of the present invention, the storage means adds a series of residual waveforms output from the voice analysis means for each frame in a series of frames. It is characterized in that the residual waveform is stored.

According to the third aspect of the present invention, the pitch conversion means has a plurality of read channels which operate in a time-sharing manner at regular intervals,
In each of these read channels, a residual function read out from the storage means is multiplied by a window function, and the resultant is sequentially added and synthesized to generate a residual waveform having a desired pitch.

Further, claim 4 is dependent on claim 1.
In the invention described in (1), the pitch conversion means has a plurality of read channels that operate in a time-division manner, and in each of the read channels for each cycle corresponding to a pitch specified by the performance information, The coefficient group is synchronized with a residual waveform associated with the Percoll coefficient.

In the invention according to claim 5 dependent on claim 1, the voice synthesizing means includes: a noise generating means for generating a noise signal; a noise signal generated by the noise generating means; Weighting means for weighting the pitch-converted residual waveform generated by the conversion means in accordance with one of the Percoll coefficient groups output from the coefficient reading means, and adding the weights. Are percolated in accordance with the Percoll coefficient read by the coefficient reading means.

In the present invention, the Percoll coefficient group and the residual waveform obtained by the voice analysis means by the Percoll analysis are stored in the storage means, and the coefficient reading means reads the corresponding Percoll coefficient from the storage means according to the performance information. While the group is read out, the pitch conversion means pitch-converts the residual waveform stored in the storage means in accordance with the performance information, and the voice synthesis means reads the pitch-converted residual waveform by the coefficient reading means. Percall synthesis is performed according to the coefficient group to generate a voice having a desired pitch. As a result, it is possible to form a musical tone as a natural singing voice using the synthesized voice.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A musical sound generator according to the present invention can be applied to a voice synthesizer for guiding voices using human voices, in addition to an electronic musical instrument. Hereinafter, a tone generator according to an embodiment of the present invention will be described as an example with reference to the drawings.

(1) Overall Configuration FIG. 1 is a functional block diagram for explaining the principle of one embodiment of the present invention. In this figure, reference numeral 2 denotes a PARCOR analysis system. The Percoll analysis system 2 sequentially calculates the autocorrelation of the linear prediction error between the discrete voice data xi input from the terminal IN, and generates Percoll coefficients K1 to Kn and a residual wave. The residual wave indicates whether the voice data in the analysis window is unvoiced or voiced. When the voice is unvoiced, it becomes white noise. On the other hand, when it is voiced, it becomes a pulse train forming a pitch period. , A residual waveform data memory 3 described later
Is stored.

Reference numeral 4 denotes a Percoll coefficient storage unit for sequentially storing Percoll coefficients K1 to Kn output from the Percoll analysis system 2 for each analysis frame. The analysis frame here corresponds to a speech analysis period defined by a window function described later. Reference numeral 5 denotes pitch conversion means for converting the pitch of the residual waveform data read from the residual waveform data memory 3 in accordance with pitch data designating a reproduced pitch, such as a frequency number. The details will be described later. 6
Is a Percoll synthesis system for synthesizing voice in a process reverse to that of the above-mentioned Percoll analysis system 2, and the residual waveform data output from the pitch conversion means 5 and the Percoll coefficients K 1 to K 1 read from the Percoll coefficient storage unit 4. Based on Kn, discrete audio data xi is synthesized and output from terminal OUT.

(2) Configuration of Main Part Next, referring to FIGS. 1 to 5, the configuration of each of the main parts 2 to 6 will be described. Configuration of Percoll Analysis System 2 In the Percoll analysis system 2 shown in FIG. 1, reference numeral 2a denotes a delay circuit that delays an input signal by at least one sampling period and outputs the delayed signal. Note that this delay circuit 2a
The sampling delay is not limited to one sample. When the sampling frequency Fs of the system exceeds 30 KHz, a two-sample delay is appropriate. 2b is a sample (voice data) xi (N pieces)
This is a correlator that calculates an autocorrelation value after multiplying by a window function W (n) and weighting.

As the window function W (n) weighted in the correlator 2b, a Hanning window function expressed by the following equation (1) is used.

[0017]

(Equation 1)

As shown in FIG. 3, the Hanning window function W (n) has an analysis frame having a width of about 30 ms and has
The analysis proceeds at a frame period of 0 ms. In the correlator 2b, the Percoll coefficients K1 to Kn are calculated based on the autocorrelation function shown in the following equation (2).

[0019]

(Equation 2)

In this function, the Percoll coefficients K1 to K1
Kn is normalized so as not to be affected by the amplitude of the sample. The Percoll coefficients K1 to Kn are “1” when there is a perfect correlation, “0” when there is no correlation, and “−1” when there is a completely opposite phase relationship.

Reference numeral 2c denotes a coefficient multiplier which multiplies the voice data delayed by one sample by the Percoll coefficient K1 and outputs the result. 2d denotes a coefficient multiplier which multiplies the currently sampled voice data by the Percoll coefficient K1 and outputs the result. is there. 2e
Is a coefficient multiplier 2c from the currently sampled audio data.
Is a subtractor for subtracting the output of the coefficient multiplier 2d from the audio data delayed by one sample.

The above-mentioned components 2a to 2f constitute a lattice filter 2-1. The lattice filters 2-1 to 2-n are connected in cascade by n stages to form a linear filter between samples (voice data) xi. A Percoll analysis system 2 for generating Percoll coefficients K1 to Kn representing the correlation of the prediction error is formed.

Configuration of Percoll Synthesis System 6 The Percoll synthesis system 6 shown in FIG. 1 is composed of lattice filters 6-1 to 6-n cascaded in n stages, similarly to the Percoll analysis system 2 described above. . The cascade-connected lattice filters 6-1 to 6-n each include a delay circuit 6f, coefficient multipliers 6g and 6h, an adder 6i and a subtractor 6j.
Based on Kn, speech synthesis is performed in the reverse process of the above-described Percoll analysis process.

If the amount of delay in the delay circuit 6f is made equal to the amount of sampling delay of the Percoll analysis system 2, the formant becomes the same as the analyzed audio signal. Therefore, in order to intentionally make the formant different as a special effect at the time of speech synthesis, a sampling delay amount different from that at the time of analysis may be used.

Configuration of Percoll coefficient storage unit 4 Percoll coefficient K1 output from Percoll analysis system 2
To Kn are sequentially stored in the Percoll coefficient storage unit 4 or read out from the storage unit 4 based on an instruction from a control unit (not shown). Here, FIG.
The frame storage mode of the Percoll coefficient storage unit 4 will be described with reference to FIG. In the case of the present embodiment, the analyzed Percoll coefficients K1 to Kn are arranged and stored for each analysis frame described above.

In the case of Japanese, the percall coefficients K1 to Kn hardly change in the vowel part, while the overtone changes in the consonant part are large, and accordingly, the changes in the Percoll coefficients K1 to Kn are large. Therefore, in the vowel part, as shown in FIG. 4, a stop flag “1” is set in the stop bit STB, and the analysis frame having similar Percoll coefficients K1 to Kn before and after is deleted to greatly reduce the data amount. It can be reduced.

That is, when note-on is given as performance data, the feature parameters are sequentially read out and input to the Percoll synthesis system 6 until the frame in which the stop flag is "1", and the stop flag becomes "1". When the current frame becomes a read target, the update reading of the frame is temporarily stopped. Then, when the next note-on occurs, it becomes possible to sequentially read out the characteristic parameters up to the frame in which the stop flag is set again and give it to the Percall synthesizing system 6 to perform voice synthesis. In addition, an aspect in which percall synthesis is performed according to performance data (modification)
Will be described later.

A start address STA is added to the Percoll coefficients K1 to Kn arranged and stored for each frame. The residual waveform data of the corresponding frame is read from the residual waveform data memory 3 with reference to the start address STA. In the example shown in FIG. 4, the start address STA is provided for each frame. However, the start address STA is not necessarily required for each frame, and may be added for some frames.

Configuration of Residual Waveform Data Memory 3 The residual waveform data generated in the above-mentioned Percoll analysis system 2 is obtained by multiplying a Hanning window function W (n) for each frame as shown in FIG. Is output. In the residual waveform data memory 3, the residual waveform data output from the Percoll analysis system 2 is added for each frame and stored as one continuous waveform data (see FIG. 3). By storing the residual waveform data in this manner, the residual waveform data is read out from the data memory 3 as it is, and is input to the Percoll synthesizing system 6 together with the corresponding Percoll coefficients K1 to Kn. You can get sound.

Configuration of Pitch Conversion Means 5 As a simple method for converting the pitch of the residual waveform data stored in the residual waveform data memory 3,
The interpolation readout is known in which the read address of the data is divided into a real part and a decimal part, and the residual waveform data read according to the adjacent real part addresses is interpolated by a decimal part address serving as an interpolation coefficient.

When the pitch conversion by the interpolation reading is performed, the read time changes in synchronization with the reproduction pitch, so that the Percoll coefficients K1 to K
n and the residual waveform data to be read out are shifted in time. Therefore, in the present invention, by performing pitch conversion on a plurality of time-division channels at regular intervals, residual waveform data of frames to be reproduced and Percoll coefficients K1 to Kn are reproduced.
Are synchronized with each other, and the residual waveform data read by each time-division channel is individually multiplied by a window function, and the results are successively added and synthesized to generate residual waveform data of a desired pitch.

Here, an example in which pitch conversion is performed based on two time-division read channels CH1 and CH2 will be described with reference to FIG. In this case, first,
In the read channel CH1, the corresponding residual waveform data RW1 is interpolated and read from the residual waveform data memory 3 based on the waveform read start address STA ₁ (see FIG. 2) added to the Percoll coefficients K1 to Kn of the frame to be reproduced. The read residual waveform data RW1 is multiplied by a window function.

[0033] In the next read cycle, the read channel CH2, based on the waveform reading start address STA ₂ which corresponds to the Percoll coefficient K1~Kn frame to be reproduced, the corresponding residual waveform data from the residual waveform data memory 3 RW2 is interpolated and read, multiplied by a window function, and added to the residual waveform data RW1 multiplied by the window function first. Thereafter, similarly, the residual waveform data RW3, RW4,... Alternately interpolated and read from the respective channels CH1 and CH2 are each multiplied by a window function and then sequentially added. By doing so, the Percoll coefficients K1 to K1 used for the synthesis are independent of the reproduction pitch.
Kn can be synchronized with the residual waveform data.

Next, with reference to FIG. 5, a description will be given of a functional configuration of the pitch conversion means 5 for performing pitch conversion in the above-described process. In FIG. 5, 5-1 and 5-2 interpolate and read the corresponding residual waveform data from the residual waveform data memory 3 based on the waveform read start address STA corresponding to the Percoll coefficients K1 to Kn of the frame to be reproduced. The interpolator is composed of a subtractor 5a, a coefficient multiplier 5b, and an adder 5c, and performs an interpolating operation alternately in a time-division manner at regular intervals.

That is, for example, when the residual waveform data a and b are read from the residual waveform data memory 3 in accordance with the real part address adjacent to the waveform read start address STA, the interpolator 5-1 causes the decimal part of the address STA to be read. Is used as the interpolation coefficient α to generate interpolation residual waveform data of a (1−α) + bα, while the interpolator 5-2 similarly generates a (1
−β) + bβ interpolated residual waveform data is generated. 5-3
Is a subtractor 5a, like the interpolators 5-1 and 5-2.
An arithmetic unit comprising a coefficient multiplier 5b and an adder 5c, and generates a pitch conversion output obtained by interpolating the interpolation residual waveform data output from the interpolators 5-1 and 5-2 with a window function Wn. I do.

Thus, according to the present embodiment, the residual waveform data output from the Percoll analysis system 2 is added for each frame and stored in the residual waveform data memory 3 as one continuous waveform data. By performing pitch conversion on a plurality of time division channels at regular intervals, the residual waveform data of the frame to be reproduced and the Percoll coefficients K1 to
Kn is synchronized with each other, the residual waveform data read in each time-division channel is multiplied by a window function individually, and the results are sequentially added and synthesized to generate residual waveform data of a desired pitch. It is possible to generate a corresponding pitch conversion output, thereby making it possible to form a human voice synthesized with speech as a natural singing tone.

In this embodiment, an example in which the primary interpolation is performed has been described. However, this is not limited to the primary interpolation, and a higher-order interpolation using a well-known spline function or the like is performed to realize higher quality pitch conversion. You may make it an aspect.

(3) Modification Next, with reference to FIG. 6, a description will be given of a modification in which voices of a desired pitch are percall-synthesized while synchronizing the Percoll coefficients K1 to Kn with the residual waveform data according to the performance data. In this figure, the same reference numerals are given to the same components as those in the above-described embodiment, and the description thereof will be omitted.

In the case of synchronous reproduction in accordance with the performance data, it is required that the progress of the Percoll coefficients K1 to Kn be controlled by the performance data and that the residual waveform be pitch-converted in accordance with the performance data. Therefore, in a modified example, a frame is designated in the Percoll coefficient storage unit 4 according to the performance data, and the pitch conversion means 5 performs interpolation reading using a plurality of time-division channels in each cycle according to the pitch data. A control circuit 7 for instructing is provided.

The above-mentioned stop bit STB is provided in the Percoll coefficients K1 to Kn of each frame stored in the Percoll coefficient storage unit 4, and the frame update of the Percoll coefficients K1 to Kn is temporarily performed mainly at the position of the vowel. Since it can be stopped, a frame can be designated according to the performance data. When the frames are sequentially updated, the start address STA added to the Percoll coefficients K1 to Kn is also incremented at a constant cycle. When the frame update is temporarily stopped based on the stop flag of the stop bit STB, the address increment is also stopped. .

Then, the pitch conversion means 5 repeatedly reads the time-divided plural channels as the read start address of the start address STA of the frame in which the reproduction pitch is stopped as it is. Accordingly, a pitch conversion output corresponding to a desired pitch can be generated according to the performance data while synchronizing the Percoll coefficients K1 to Kn used for the synthesis and the residual waveform data, so that a natural singing voice corresponding to the performance data can be generated. Can be obtained.

(4) Other Examples By the way, it is not necessary to strictly synchronize the Percoll coefficients K1 to Kn and the residual waveform data used for the synthesis, and the frame supply cycle of the Percoll coefficients K1 to Kn, the pitch conversion, and the window. It is not necessary to exactly match the period of the function. In general, it is appropriate to set the period of the pitch conversion and the window function later than the frame supply period of the Percoll coefficients K1 to Kn.

The reason why the two do not always have to match is based on the premise that the speech to be synthesized changes relatively smoothly and the frequency components of the residual waveforms of the preceding and succeeding frames are similar. Therefore, the pitch conversion amount is large,
Further, when the consonant whose residual waveform changes rapidly and the residual waveform becomes almost white noise, the Percoll coefficient K1
Since the mismatch between ~ Kn and the residual waveform causes sound quality deterioration, the residual waveform is estimated from the Percoll coefficients K1 to Kn,
It is conceivable that the output from the pitch conversion and the switching of the white noise source, the weighting change, etc. are used together.

The configuration of such an embodiment will be described with reference to FIG. The configuration shown in FIG. 7 is obtained by adding components 10 to 14 to the embodiment shown in FIG. In such a configuration, reference numeral 10 denotes a noise generator that generates white noise WN when synthesizing an unvoiced sound. 11
Sets the Percoll coefficient K1 to a predetermined constant (about 0.2 to 0.
This is a comparator that determines either voiced sound or unvoiced sound by comparing the magnitudes of 3) and generates multiplication coefficients W1 and W2 according to the result.

That is, when the Percoll coefficient K1 used for synthesis corresponds to "voiced sound", the comparator 11 outputs the coefficient W1 as "1" and sets the coefficient W2 as "0". . On the other hand, on the other hand, "silent sound"
Is output, the coefficient W2 is output as "1", and the coefficient W1 is set to "0".

Reference numerals 12 and 13 denote coefficient multipliers, respectively.
The coefficient multiplier 12 multiplies the pitch-converted residual waveform data by a coefficient W1 and outputs the result. The coefficient multiplier 13 multiplies the white noise WN by a coefficient W2 and outputs the result.
An adder 14 adds the outputs of the coefficient multipliers 12 and 13 and supplies the result to the Percoll synthesis system 6. Therefore, the adder 14 supplies the residual waveform data whose pitch has been converted to the Percoll synthesis system 6 when synthesizing a voiced sound (vowel), and converts the white noise WN into the Percoll synthesis system 6 when synthesizing an unvoiced sound (consonant). To supply.

According to the above arrangement, when synthesizing a consonant having a large pitch conversion amount and a drastically changing residual waveform, the white noise WN is supplied to the Percoll synthesis system 6, so that the Percoll coefficients K1 to K1 are used. Sound quality deterioration due to mismatch between Kn and the residual waveform can be prevented, and a natural human voice can be synthesized.

In the above configuration, if the white noise WN is used when synthesizing a consonant, there is no need to store the residual waveform data corresponding to the consonant in the residual waveform data memory 3. The storage capacity of the memory 3 can be reduced. Note that the above-described comparator 1
In the case of No. 1, the coefficients W1 and W2 are cross-fade according to the change of voiced sound (vowel) / unvoiced sound (consonant), and the change from voiced sound to unvoiced sound or the change from unvoiced sound to voiced sound is more natural. It can also be shaped.

[0049]

According to the present invention, the Percoll coefficient group and the residual waveform obtained by the Percoll analysis by the voice analysis means are stored in the storage means, and the coefficient reading means reads out from the storage means in accordance with the performance information. While the corresponding Percoll coefficient group is read, while the pitch conversion means pitch-converts the residual waveform stored in the storage means according to the performance information, the voice synthesizing means converts the pitch-converted residual waveform to the coefficient reading means. Therefore, a voice having a desired pitch is generated by performing a Percoll synthesis in accordance with a Percoll coefficient group read out by the user, so that a human voice synthesized with the voice can be formed as a natural singing tone.

[Brief description of the drawings]

FIG. 1 is a functional block diagram for explaining the principle of an embodiment according to the present invention.

FIG. 2 is a diagram for explaining a frame storage mode of a Percoll coefficient storage unit 4;

FIG. 3 is a diagram for explaining residual waveform data generated in a Percoll analysis system 2;

FIG. 4 is a diagram for explaining an example of pitch conversion in pitch conversion means 5;

FIG. 5 is a block diagram showing a functional configuration of a pitch conversion unit 5;

FIG. 6 is a block diagram showing a configuration of a modification.

FIG. 7 is a block diagram showing a configuration of another example.

[Explanation of symbols]

 2 Percoll analysis system (voice analysis means) 3 Residual waveform data memory (storage means) 4 Percoll coefficient storage unit (storage means) 5 Pitch conversion means (pitch conversion means) 6 Percoll synthesis system (voice synthesis means) 7 control circuit ( Coefficient reading means)

Claims (5)

[Claims] 1. A speech analysis means for performing a Percoll analysis on a speech signal sampled in a time series to generate a Percoll coefficient group and a residual waveform, and stores the Percoll coefficient group and a residual waveform generated by the speech analysis means. Storage means for reading, a coefficient reading means for reading out a corresponding Percoll coefficient group from the storage means in accordance with the performance information, and a pitch conversion means for performing pitch conversion of the residual waveform stored in the storage means in accordance with the performance information And voice synthesis means for generating a voice of a desired pitch by performing Percoll synthesis on the residual waveform pitch-converted by the pitch conversion means in accordance with a Percoll coefficient group read by the coefficient reading means. Musical sound generator. 2. The storage device according to claim 1, wherein the storage unit stores a series of residual waveforms obtained by adding the residual waveforms output from the voice analyzing unit for each frame in frame order. Musical sound generator. 3. The pitch conversion means has a plurality of read channels which operate in a time-sharing manner at regular intervals, and multiplies each of the residual waveforms read from the storage means by a window function in each of the read channels. 2. The musical tone generator according to claim 1, wherein the residual waveform is sequentially added and synthesized to generate a residual waveform having a desired pitch. 4. The pitch conversion means has a plurality of read channels operating in a time-division manner, and in each read channel for each period corresponding to a pitch specified by the performance information, a Percoll coefficient read by the coefficient read means. 2. The tone generator according to claim 1, wherein the group and the residual waveform associated with the Percoll coefficient are synchronized. 5. The speech synthesis means includes: a noise generation means for generating a noise signal; a noise signal generated by the noise generation means; and a pitch-converted residual waveform generated by the pitch conversion means.
Weighting means for weighting and adding according to any one of the Percoll coefficient groups output from the coefficient readout means, wherein an output of the weighting means is output in accordance with a Percoll coefficient read out by the coefficient readout means. The musical sound generating device according to claim 1, wherein the musical sound is synthesized.