JPH1091195A - Method of analyzing and synthesizing speech - Google Patents

Abstract

PROBLEM TO BE SOLVED: To permit a satisfactory voice synthesis to woman voice, moreover, a voice analysis and synthesis by means of less calculation, in a voice analysis and synthesis using LPC(linear predictive coding) system. SOLUTION: By obtaining an amplitude value and time-information at the peak point of an original wave, determining peak values having maximal amplitudes of each single oscillation, and determining a zero-amplitude point in the direction of the origin in the time-axis of the waveform having a maximal peak value in each single oscillation, information representing a period of the unit waveform is taken out starting from that point (step s1-s5). A first unit waveform selects information representing the period and n pieces of large amplitudes and sends them to synthesis side for a synthesis processing. A second unit waveform and thereafter deforms a row of pulses obtained from a previous unit waveform into a round shape, and extracts in descending order n pieces of pulses from a pulse row of the difference obtained by subtracting the deformed pulse row from residual waveform of the unit waveform to be processed. On the synthesis side, a synthesis processing is performed based on a pulse row obtained by adding the beforementioned n pieces of pulses to the pulse row which was rounddefomed one used for the previous synthesis (step s9-s14).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for obtaining a linear prediction coefficient based on a correlation between unit waveforms, particularly for a unit waveform obtained from a female voice and performing periodic repetition for each unit time. The present invention relates to a speech analysis / synthesis method for extracting information indicating a cycle of a unit waveform and other error component information from a residual waveform obtained at the time of prediction, and performing speech synthesis based on such information.

[0002]

2. Description of the Related Art As one of algorithms for synthesizing a human voice, there is an LPC method. LPC
The method can also be referred to as an algorithm that compresses input audio data to reduce the amount of data and records and reproduces the data.

The LPC method can compress voice data at a high compression rate for vowels and the like in a male voice, and can synthesize voice with high accuracy. However, there is a problem in that a high compression ratio cannot be obtained as described above, and that the synthesized sound is liable to be deteriorated with respect to the original sound.

That is, in the LPC system, input voices are several tens of
For each frame of about msec, Y1 (LPC coefficient), Y2 (cycle and intensity of voiced sound), Y3 (error component) shown below
Is converted into three elements.

(1) YI (LPC coefficient): When a voice (especially a vowel) is viewed in a short time unit such as several tens of msec, it can be understood from an example of a voice waveform of a certain vowel shown in FIG. Generally, almost the same waveform is repeated. The LPC coefficient is a linear prediction coefficient obtained from such a correlated iteration.

(2) Y2 (cycle and intensity of voiced sound): The prediction error (residual error) in the linear prediction can be divided into two components, and one of the two components. This component is a time component (referred to as Y21) and an intensity component (referred to as Y22) in the residual waveform shown in FIG. 8B, and the time component is a point (refer to FIG. In FIG. 8B, a time t between p1, p2, p3,... Is shown, and the intensity component is a component representing the amplitude h of the input speech waveform.

(3) Y3 (error component): The prediction error (residual error) in the linear prediction can be divided into two components, one of the two components. This component is a noise component that looks seemingly random among errors between the linearly predicted waveform and the actual waveform, and is a portion indicated by N in FIG. 8B.

The LPC system converts an input voice into these three components Y1, Y2, and Y3. In practice, however, a waveform such as a vowel in a male voice that has a large correlation with a waveform (correlation with a waveform such as a fricative sound). Is not included), since there is almost no error component N of Y3,
Is ignored, and the data after compression is represented by Y1, Y21, Y2.
There are only two elements and binary data indicating voice or non-voice. Therefore, the sound having a large correlation in the waveform is shown in FIG.
Since it is only necessary to send data that ignores Y3 to the synthesizing side, a high compression rate can be obtained, and high-precision speech synthesis can be performed even when speech synthesis is performed.

[0009] For example, in the case of a male voice "Ah", considering "A", the voice waveform of "A" is periodically and repeatedly generated for a while as shown in FIG. Will be. Therefore, when the same voice waveform is repeated, the voice waveform for the first cycle (here, the waveform for this one cycle is called a unit waveform, and in FIG. 8, each unit waveform is a unit waveform 10, 20, 30,. (Represented by reference numeral 40), the information (Y22) for recognizing the amplitude and the information (Y21) for recognizing the repetition cycle of the unit waveform can be used to synthesize the voice of "a". . Further, in the case of a male voice, similar waveforms often attenuate as they are within each unit waveform when viewed from each unit waveform.

[0010] As described above, regarding the male voice having a large correlation with the voice waveform, the Y1, Y2
The elements of Y1, Y22 and binary data indicating speech or non-speech are sent to the synthesizer as compressed audio data, and the synthesizer performs synthesis processing based on this data, so that humans can hear it with their ears. It is possible to synthesize a voice almost inferior to the original sound.

[0011]

However, a female voice has a residual waveform with respect to the original waveform as shown in FIG.
As shown in FIG. 8B, compared to the case of a male, FIG.
Are not clear at the points p1, p2,... Indicated by, the error component N
The distinctive feature is that it is not clear. Therefore, in the case of a female voice, the error component N of Y3 cannot be ignored. That is, since the voice of a woman also includes an important element representing the characteristics of the voice in the error component of Y3, if the voice is synthesized while ignoring this Y3, the synthesized voice will be greatly degraded from the original sound. To cope with this, data containing an error component may be sent to the synthesis side as it is, but there is a problem in terms of data amount compression.

On the other hand, there is a method called "Analysis by synthesis" as another voice analysis / synthesis method other than the above-mentioned method. This method is a method of finding a unit waveform from an original waveform without using a residual waveform, and searching for a pulse position while observing the difference by subtracting the unit waveform from the original waveform.
Data required for voice synthesis can be sent to the synthesizer at a high compression ratio even for female voices, and high-precision synthesis can be performed. When performing the process of searching, it is necessary to perform a process such as determining the position or the amplitude of the unit waveform that minimizes the difference, so that there is a disadvantage that the amount of calculation is extremely large.

According to the present invention, in the speech analysis / synthesis using the LPC method, the "Analysis by synthesi
s "as well as speech analysis and synthesis with a small amount of calculation.

[0014]

According to the speech analysis / synthesis method of the present invention, an original waveform to be processed is divided into frames at predetermined time intervals, and substantially one of the frames is divided into frames. A linear prediction coefficient based on the correlation between the unit waveforms is obtained based on a waveform (hereinafter, referred to as a unit waveform) that periodically repeats at regular intervals, and a residual waveform indicating an error component in the linear prediction is obtained. In a speech analysis / synthesis method for extracting information indicating the cycle and intensity of a unit waveform and other error component information and performing speech synthesis based on the information, planar coordinates of a time axis and an amplitude axis which are orthogonal to each other. From the original waveform of a certain frame represented above, the peak point of the waveform is examined along the time axis, the amplitude value and time information of the peak point are obtained, and the unit waveform of each unit waveform is obtained from among these peak points. Maximum amplitude The amplitude of the waveform having the maximum peak value in each unit waveform in the direction of the origin on the time axis is determined to be 0, and the unit having the amplitude of 0 is used as a reference from the residual waveform in the vicinity thereof. Extracting information indicating the cycle of the waveform and sending it to the voice synthesis side, and extracting information necessary for synthesis from error components of the residual waveform corresponding to each of the unit waveforms as pulse information,
A speech synthesis process is performed based on the pulse information and the information indicating the period together with the linear prediction coefficient.

According to this, the information indicating the period of the unit waveform in the residual waveform, such as a female voice, and the information indicating the period accurately even for a voice that cannot be distinguished from the error component, Error components necessary for synthesis can be extracted,
Data compression can be performed at a high compression rate, and a good synthesized sound can be obtained.

In the processing for extracting information indicating a cycle of a unit waveform from a residual waveform in the vicinity thereof with reference to a point having an amplitude of 0, the point having the amplitude of 0 may be set as a cycle of a unit waveform. In addition, it is also possible to extract information having a large amplitude from a residual waveform in the vicinity of the point having the amplitude 0 as a reference and obtain information indicating a cycle based on the extracted information. .

According to this, it is possible to accurately extract information indicating the cycle of each unit waveform in the residual waveform.

Also, information necessary for synthesis is extracted as pulse information from error components of a residual waveform corresponding to each of the unit waveforms, and based on the pulse information and the information indicating the period together with the linear prediction coefficient. The voice synthesis process is performed by counting the number of peaks of the waveform in at least the first unit waveform in a frame where the state changes from an unvoiced sound to a voiced sound and in a frame where the linear prediction coefficient greatly changes. Is determined to be greater than or equal to a predetermined number. If the number is equal to or greater than a predetermined number, only the minimum information necessary for synthesis is efficiently selected from error components and sent to the synthesis side as pulse information, and the peak value is determined. If the number is less than the set value, not only the minimum pulse information necessary for the above-described synthesis, but also obtain pulse information from other error components and send it to the synthesis side. It is characterized in Rukoto.

According to this, it is determined that a speech waveform whose waveform varies variously within a unit waveform has a large amount of information necessary for speech synthesis, and only the minimum necessary information is efficiently extracted from the residual waveform. By selecting well, it is possible to increase the data compression ratio and obtain a good synthesized sound. On the other hand, an audio waveform that does not have many waveforms in the unit waveform
A good synthesized sound can be obtained by judging that the amount of information necessary for speech synthesis is small and efficiently selecting a necessary error component of the residual waveform.

Further, the processing of selecting only the minimum information necessary for the synthesis from the error component is performed by selecting the error component of the residual waveform corresponding to at least the first unit waveform in the frame where the unvoiced sound state changes to the voiced sound. , A predetermined number is selected in descending order of the absolute value of the amplitude value, and the selected number is sent to the synthesis side as pulse information.

According to this, it is possible to accurately and efficiently extract the minimum information required for speech synthesis, to obtain data compressed at a high compression rate, and to obtain a good synthesized sound. .

In the process of extracting information necessary for synthesis from the error component of the residual waveform corresponding to the head unit waveform as pulse information, the amplitude direction of the original waveform corresponding to the residual waveform to be selected is determined. When going from positive to negative, the error component information for the residual waveform is negative pulse information,
When the amplitude direction of the original waveform corresponding to the residual waveform to be selected goes from negative to positive, error component information for the residual waveform is positive pulse information.

According to this, the shapes of the original waveform and the composite waveform are 1
It can be prevented from differing greatly even sometimes,
Good synthesized sound can be obtained.

A signal necessary for synthesis is extracted as pulse information from error components of a residual waveform corresponding to each of the unit waveforms, and based on the pulse information and the information indicating the period, together with the linear prediction coefficient. In the process of performing the voice synthesis process, the pulse information used to synthesize the immediately preceding unit waveform in the second and subsequent unit waveforms of the frame to be processed is used as the actual residual waveform of the unit waveform. From the residual waveform of the subtracted unit waveform, a predetermined number is selected in descending order of the absolute value of the amplitude,
Pass the selected pulse information from the analysis side to the synthesis side,
On the synthesis side, the pulse information passed from the analysis side is
The pulse information obtained by adding to the pulse information used for synthesizing the previous unit waveform is used for synthesizing the unit waveform.

According to this, for example, when trying to obtain a synthesized sound with respect to the speech waveform "Ah", the amount of information is initially small, and even if the synthesized waveform is not very similar to the original waveform, the time is short. As the amount of information accumulates over time, the synthesized waveform becomes a waveform very close to the original waveform, and a better synthesized sound can be obtained as a whole.

The processing of subtracting the pulse information used for synthesizing the immediately preceding unit waveform from the actual residual waveform of the unit waveform is used for synthesizing the immediately preceding unit waveform. Each pulse constituting pulse information is broken down into a mountain shape,
The pulse information disintegrated into the chevron is subtracted from the actual residual waveform of the unit waveform, and the pulse information passed from the analysis side is used for synthesizing the previous unit waveform. The processing of the synthesis side to add to the information breaks each pulse constituting the pulse information used for synthesizing the previous unit waveform into a mountain shape, and converts the pulse information broken into the mountain shape into the pulse information from the analysis side. Pulse information obtained by adding the passed pulse information is obtained.

As described above, by processing the pulse information by breaking it into a mountain shape, when the information obtained by adding the information of the current unit waveform to the information accumulated so far is obtained on the synthesis side,
When the pulse to be sent to the synthesis side obtained with the previous unit waveform is superimposed on the residual waveform of the current unit waveform with the position of the first pulse train aligned, and there is some deviation in the whole pulse train of the unit waveform This deviation can be absorbed and processed.

Further, in the processing for performing the speech synthesis processing based on the information indicating the cycle and intensity of the unit waveform and the error component information, processing for smoothing the envelope of the synthesized waveform is performed. Then, in the process of smoothing the envelope of the synthesized waveform, a processing target section is set for the synthesized waveform, and for each frame in the processing target section, the average of the maximum peak value of each unit waveform in the frame is calculated. Then, the peak value of each unit waveform in the entire processing target section is set to a predetermined value obtained based on the obtained average peak value.

According to this, an envelope (envelope) connecting the peak values of each frame of the processing target section in the synthetic waveform becomes a flat curve due to the perceived fluctuation of the ears. , Etc., can be prevented from becoming a gurgling sound in which a synthesized sound with respect to an original waveform of a voice in which a vowel lasts for a long time has something tangled in the throat.

[0030]

Embodiments of the present invention will be described below. As described above, in the case of the female voice, the pulse at p1, p2,... Shown in FIG. There is a feature that the distinction from the error component N is not clear. Therefore, in the case of a female voice, the error component N of Y3 cannot be ignored.

In other words, since the voice of a woman also includes an important element representing the characteristics of the voice in the error component of Y3, if the synthesis is performed ignoring Y3, the synthesized voice will be greatly deteriorated from the original sound. There is a problem that the data containing the error component can be sent to the synthesis side as it is,
There is a problem in terms of data volume compression.

Therefore, according to the present invention, the first pulse (pulse corresponding to p1, p2,... Shown in FIG. 8) in each unit waveform is efficiently extracted from the residual waveform for the voice signal of the female voice. In addition to the extraction, a component important in speech synthesis among the error components is extracted and sent to the speech synthesis side, and a pulse train convenient for synthesis is reproduced on the synthesis side. Hereinafter, embodiments of the present invention will be described.

FIG. 1 shows an example of the characteristics of a female voice waveform used in this embodiment. FIG. 1A shows one frame of a female voice such as "Ah" (Equation 1).
It is assumed that the unit waveform is composed of about four unit waveforms as in FIG. 8, but in order to simplify the drawing, only unit waveform 10 and unit waveform 20 are used. It is illustrated. The unit waveform 10 is the first unit waveform when the state changes from an unvoiced sound to a voiced sound, followed by a unit waveform 20, and although not shown here, several unit waveforms follow. For example, one frame is composed of four unit waveforms, followed by several frames to form a certain audio waveform.

FIG. 1B shows the residual waveform. As can be seen from the residual waveform, the leading pulse is a characteristic of the female voice, that is, P1, p of the male voice described in FIG.
There is almost no difference between the pulses (Y21 and Y22 components) corresponding to 2,... And other error components (Y3 component indicated by N in FIG. 8). 1C to 1F will be described later.

In such a residual waveform, a pulse (head pulse) indicating the head position of each unit waveform from the error components of the residual waveform and the head pulse are required in addition to the minimum necessary for speech synthesis. The process of extracting the error component as a pulse will be described with reference to the flowchart of FIG. 2 and FIG.

First, in the original waveform shown in FIG.
The amplitude peak is checked along the time axis, and information on the peak value and time is obtained (step s1). For example, in the example of FIG. 1 (a), when checking along the time axis, p1
, P12, p13,... Are obtained as peak portions of the amplitudes in the respective waveforms. The amplitude value of the point p11 is h1, the time t1 at that time, and the amplitude value of the point p12 is h.
2. At that time, the amplitude value of the peak point and time information of each waveform can be obtained by setting the time t2, the amplitude value of the point p13 to h3, and the time t3 at that time. FIG. 1C shows the information of each peak value obtained in this manner as coordinates of the time axis and the amplitude value. As can be seen from FIG. 1 (c), the peak value of the original waveform oscillates at a constant cycle.
Here, one unit is called a unit vibration (one unit vibration corresponds to one unit waveform as a result), and a peak point having the largest amplitude is selected from among peak values in each unit vibration (step s2). ). In this case, the points P11 and p1
4 is the largest peak value among the unit vibrations. As a result, the point P11 is the largest peak value in the unit waveform 10 and the point p14 is the largest peak value in the unit waveform 20.

Then, from the time of the peak point having the largest amplitude value selected in each unit vibration, the waveform returns to the origin o along the time axis of the original waveform, and the point where the waveform and the time axis intersect (the amplitude value Is zero) (step s3). That is, in FIG. 1 (a), returning from the time t1 of the point p11 to the origin o, a point where the waveform and the time axis intersect (a point where the amplitude is 0) is obtained, and the point is set to tx1, and similarly, the point p14 is set. Returning from the time t4 to the origin o, a point where the waveform and the time axis intersect (a point where the amplitude is 0) is obtained, and the point where the waveform and the time axis intersect (the amplitude value is 0).

When the points tx1, tx2,... Where the waveform and the time axis intersect are obtained in this manner,
Along the time axis of the residual waveform shown in (b), the Δt time is set before and after these points tx1, tx2,..., And the range of this Δt time is scanned. , An error component having the maximum amplitude value is detected (step s4). Here, the Δt time is determined based on experience based on experience, for example, based on the length of time of a unit vibration (unit waveform), such as a fraction of that time. .

Although it has been mentioned above that the distinction between the Y2 component and the Y3 component is not clear in the female voice,
The components are those that are present. Therefore, the part having the maximum amplitude within the time Δt is detected, and the information is used as the information on the start position of each unit waveform to obtain head pulses Ps1, Ps2,... As shown in FIG. s
5).

In some unit vibrations, the head position may not be extracted even after scanning for Δt. In such a case, the determination is made based on the temporal position of the leading pulse detected in the preceding and following unit vibrations. In other words, since the leading pulses are generally emitted at substantially equal intervals, the position of the leading pulse in an unclear part is determined based on the detected leading pulse interval. In addition, the magnitude of the amplitude of the leading pulse in the unclear part is the amplitude value of the leading pulse of the previous unit vibration.

As described above, the head position of each unit waveform can be specified from the original waveform and its residual waveform. As a method of extracting head position information from the residual waveform, a point (amplitude value 0) where the waveform and the time axis in each unit waveform intersect is directly used as the head point of the unit waveform, and this is used as the original waveform. Can be used as the head position information.

The head position information of each unit waveform obtained in this manner is passed to the synthesizing side, and together with the head position information, an error component (minimum necessary for synthesizing, a minimum necessary error component for synthesizing) is provided. (A typical error component) is extracted from the error component N of the residual waveform as shown in FIG.

First, the basic processing will be described. This processing is performed for each unit waveform (in FIG. 1,
The processing is performed for each of the unit waveforms 10 and 20).
It differs depending on whether the unit waveform to be processed is the first unit waveform or the second and subsequent unit waveforms.

When it is determined that the unit waveform to be processed is the first unit waveform (step s6), an error component (corresponding to the Y3 component) is sent to the synthesis side in addition to the above-described head position information. )), N pulses having a large absolute value of the amplitude are selected and sent to the synthesizing side as pulse position and pulse magnitude data (step s7). FIG. 1E shows an example of the leading pulse and the newly selected n pulses for the unit waveform 10. In the pulse shown in FIG. 1E, the pulse in the negative direction is a pulse corresponding to a large error component in the negative direction.

On the synthesizing side, the data indicating the head position of the unit waveform obtained in step s5 and the pulse position and pulse magnitude data obtained in step s7 are used to synthesize the first unit waveform. The n pulses are created on the time axis (step s8). Then, waveform synthesis is performed using the pulse train obtained as described above and the LPC coefficient described above (step s14).

On the other hand, if the unit waveform to be processed is not the first unit waveform, the process proceeds to step s9. For example, in the unit waveform 20, as in the case of the head unit waveform, the head pulse in the unit waveform and n pulses having a large amplitude among the error components are selected, and the corresponding pulse is sent to the synthesis side. Instead of using the method, here, the information obtained by the previous unit waveform and sent to the synthesizing side (the information accumulated so far) is used, and the information obtained by adding the information of the current unit waveform to the information is used. The method of obtaining on the synthesis side is used. Hereinafter, this will be described.

There is considerable correlation between pulse trains of adjacent unit waveforms. Therefore, for example, the n pulses (indicated by X1 in FIG. 1E) obtained by the unit waveform 10 in step s7 are converted into the residual pulse train of the unit waveform 20 (FIG. 1).
If the position of the first pulse train is aligned and superimposed on (b)), they should be quite consistent. However, the pulse trains of the unit waveform often have some deviation as a whole. In order to absorb this shift, the previous unit waveform (in this case,
All the pulses of the pulse train obtained in the unit waveform 10) are broken into a mountain shape (step s9), and the pulses in the unit waveform 20 are extracted using the pulses broken into the mountain shape.

The breaking of this pulse into a mountain shape is shown in FIG.
FIG. 3B shows a pulse as shown in FIG. 3A. For example, when the amplitude h of the pulse in FIG. 3A is １ ／ as shown in FIG. This is a process of distributing the remaining 分 け amplitude value to both sides, and is performed for all the pulses obtained in the unit waveform 10.

The same pulse train as that sent to the synthesizing side of the unit waveform 10 broken into a mountain shape is used as the actual residual waveform of the unit waveform 20 (the residual waveform pulse train of the unit waveform 20 is shown in FIG. X2) to obtain a pulse train of the difference (step s10), and from the pulse train of the difference, extract n error components having amplitudes with large absolute values in descending order, and extract these pulse positions and The size data is sent to the combining side (step s11). The n pulses extracted from the unit waveform 20 are indicated by X3 in FIG.

On the other hand, on the synthesizing side, a pulse train in which each pulse of the pulse train used for synthesizing the immediately preceding one (in this case, the unit waveform 10) is formed into a mountain shape (step s12), and obtained in step s11. The pulse positions and magnitudes of the n pulses (indicated by X3 in FIG. 1 (f)) are determined in step s12.
(Step s)
13). And the pulse train obtained as described above,
Waveform synthesis is performed using LPC coefficients (step s14).

As described above, for the second and subsequent unit waveforms, the pulse information up to the previous time is accumulated, and voice synthesis is performed based on the accumulated information. In other words, for example, when trying to obtain a synthesized sound for the speech waveform "Ah", the amount of information is initially small,
Even though the synthesized waveform is not very similar to the original waveform, the amount of information is accumulated over time, so the synthesized waveform becomes a waveform very close to the original waveform, and as a whole, an extremely good synthesized sound can be obtained. Can be.

Originally, if all error components of the residual waveform are sent to the synthesis side, high-accuracy speech synthesis can be performed. However, in consideration of data compression, it is necessary to reduce the data amount even a little. come. As a method for coping with this, in each unit waveform, the leading pulse and n signals having a large amplitude among error components are selected and sent to the combining side, and the combining side also sends each unit waveform. Although there is a method of synthesizing only the pulse train that has been used, since there is some correlation between the pulses to be extracted between each unit waveform, rather than extracting each pulse alone, as described above, It is better to accumulate and process the information, such as adding the current pulse information, rather than extracting the pulse information alone for each unit waveform.
This is advantageous for speech synthesis.

The above is the basic processing of the present invention for extracting the component that needs to be passed to the synthesis side from the Y3 component of the residual waveform. In performing such processing, the following processing is performed. By applying, a more favorable synthesized sound can be obtained.

First, in the first unit waveform 10, as described above, in addition to the first pulse, n pulses having a large amplitude are simply selected and sent to the synthesizing side. This may cause inconvenience.

That is, high-accuracy speech synthesis can be performed only after all the error components of the residual waveform have been sent to the synthesis side. However, a method of simply selecting n large amplitude signals and sending it to the synthesis side. In some cases, a highly accurate synthesized sound may not be obtained. This is because, when a certain unit waveform of the original waveform is, for example, as shown in FIG. 4A, a waveform after speech synthesis based on the pulse information obtained by performing the above-described processing (synthesized waveform). However, this is because the positive and negative directions of the amplitude may partially differ from the original waveform as shown in FIG. In order to cope with this, in the present invention, the following processing is performed when n large amplitudes are selected.

When the amplitude is large from the error component of the residual waveform, n
When extracting individual pieces, check the direction of the amplitude of the original waveform, select a negative pulse where the amplitude increases from positive to negative, and increase the amplitude from negative to positive. In the part, a positive pulse is selected.

More specifically, when the residual waveform for a certain unit waveform of the original waveform as shown in FIG. 5A is as shown in FIG. 5B, the amplitude of the original waveform changes from positive to negative. A residual waveform portion corresponding to the point A to be changed (this residual waveform portion is
In the drawing, it is indicated by w, and this w has a predetermined range around the point A, and the size of the range is determined in advance.)
If there is a large pulse to be extracted as an error component of the residual waveform, the pulse is extracted as a negative-direction pulse as shown in FIG. From the residual waveform portion corresponding to the point B which is about to change from positive to positive (the same as described above, this residual waveform portion is indicated by w in the figure, and w has a predetermined range around the point B). If the size of the range is determined in advance) and there is a large pulse to be extracted as an error component of the residual waveform, the pulse is shifted in the positive direction as shown in FIG. Extract as a pulse. However, the vicinity of the first pulse of the unit waveform (FIG. 5)
Only the point (in the vicinity of point C in (a)) is selected from the one with the larger absolute value regardless of the sign of the pulse.

As described above, when n large pulses are selected from the residual waveform, and the selection is performed, the direction of the amplitude of the original waveform (from positive to negative or from negative to positive) is determined. The process of extracting the pulse is performed only on the unit waveform near the beginning where the unvoiced sound state changes to the voiced sound, and in the second and subsequent unit waveforms, as described above, using the information accumulated from the previous unit waveform. Good results are obtained with the treatment.

As described above, such processing is performed only on the unit waveform near the beginning, that is, only on the first few unit waveform portions that change from the unvoiced sound state to the voiced sound. Unlike unit waveforms, there is no accumulation of information,
This takes into consideration how to obtain a synthesized waveform close to the original waveform on the synthesis side with a small amount of information. In addition, this processing may be performed for some frames immediately after the frame in which the linear prediction coefficient (LPC coefficient sequence) is compared for each frame and the value of which is greatly changed.

By performing the above processing, the relationship between the sign of the amplitude of the synthesized waveform obtained on the synthesizing side and the sign of the amplitude of the original waveform based on the pulse information extracted from the residual waveform in the first unit waveform. Does not greatly differ, and a synthesized waveform relatively close to the original waveform can be obtained even in the vicinity of the first unit waveform. However, in order to obtain a better synthesized sound, the following processing can be performed.

The number of waveform peaks is counted in the first several unit waveforms that have changed from unvoiced to voiced (unit waveform 10 in FIG. 1A) or in the unit waveform near the frame where the LPC coefficient has changed significantly. Then, the number is compared with a preset threshold value, and different processing is performed depending on whether the number is equal to or greater than the threshold value or less than the threshold value. That is, different processing is performed depending on whether the waveform in the unit waveform is a complex waveform having a large number of peaks (a waveform having a large amount of information) or a simple waveform having a small number of peaks (a waveform having a small amount of information). . This processing will be described with reference to FIG.

For example, assuming that the threshold value is 4,
As shown in FIG. 6A, the peak number of the first unit waveform is p1
If there are four (1) to (p14), it is determined that the information amount is relatively large, and if the information amount is large, to reduce the information amount,
As the error component s0 for obtaining the leading pulse and the minimum error component that needs to be sent to the synthesis side among the other error components, s1, s2, s3, s as shown in FIG.
Only the error components in the range of 4, s5 are selected, and from s0, a pulse having a large absolute value, s1, s2, s3, s4, s5
As described above, the selected pulse is sent to the synthesizing side while considering the code from the pulse train corresponding to the above, and error components in other sections are ignored.

On the other hand, when the number of peaks in the first unit waveform is only p11 and p12 and less than the threshold as shown in FIG. 6C, the information amount is originally small and wider than the above. It is determined that a pulse in the range may be selected, and an error component s0 for obtaining the leading pulse and a minimum error component of the other error components that need to be sent to the combining side are shown in FIG. Thus, the pulse selected from s0 without considering the sign, the pulse selected from the pulse train of s1, s2, and s3 considering the sign as described above, and s11 and s1
Among the pulse trains of s2, s13, and s14, n pulses having a large absolute value are obtained from the pulses selected without considering the sign, and sent to the synthesis side. That is, in the case of a unit waveform having a small amount of information, information of more sections of the residual pulse train of the unit waveform is sent to the synthesis side.

As described above, the audio waveform in which the waveform in the unit waveform changes variously has a large amount of information necessary for synthesizing the audio, and it is difficult to reproduce the waveform by simply selecting n large pulses. It cannot be said that useful information is selected efficiently. On the other hand, by efficiently selecting only the minimum necessary information from the residual waveform as described above, the data compression ratio can be increased, and a good synthesized sound can be obtained. On the other hand, a speech waveform having a small change in the waveform in the unit waveform is judged to have a small amount of information necessary for speech synthesis, and is selected from a large number of sections of the error component of the residual waveform, thereby improving accuracy. A good synthetic sound with good quality can be obtained.

By performing the above-described processing, a synthesized waveform closer to the original waveform can be obtained on the synthesis side, and an extremely good synthesized sound can be obtained.

By the way, the phenomenon that the synthesized sound obtained by the above-described processing becomes a gurgling sound such as something entangled in the throat in a portion where a vowel such as "Oi" continues for a long time. May occur. It occurs particularly frequently in female voices.

This is the envelope (envelope) of the original waveform.
As seen from the above, while the maximum value of the amplitude value does not fluctuate so much, the waveform after synthesis (synthetic waveform) obtained by the above processing has the maximum amplitude value for each unit waveform. This is considered to occur because the fluctuation is larger than the original waveform, or the maximum value of the amplitude of the unit waveform fluctuates greatly over several frames.

FIG. 7A shows an envelope 21 connecting the positive maximum amplitude values of the unit waveforms constituting each frame of several frames in the original waveform for a certain voice, and the negative maximum amplitude. Shows an envelope 22 connecting the values,
As can be seen from this figure, there is almost no unevenness and it is linear. On the other hand, as shown in FIG. 7B, a composite waveform obtained as a result of performing the above-described processing is an envelope connecting the maximum positive-side amplitude values of each unit waveform constituting each frame. The envelope 23 connecting the line 23 and the maximum amplitude value on the negative side may be a curve with large irregularities as shown in FIG. 7B. It is thought that the phenomenon of rumbling sound that something is entangled occurs. Such a phenomenon can be dealt with by performing the following processing.

First, when the continuous voice portion is longer than a predetermined time, the processing target section D excluding the rising portion and the falling portion of the voice portion in the synthesized waveform as shown in FIG. Set. In this processing target section D, the average value of the maximum amplitude value of each unit waveform in each frame (about several tens msec) is obtained. Since the positive side envelope 23 and the negative side envelope 24 of the synthesized waveform are not targets, the processing is different on the positive side and the negative side. However, since the processing method is the same, ,
Only the positive side will be described.

As described above, the processing for obtaining the average value of the maximum amplitude value in each unit waveform in one frame is performed, for example, when there are four unit waveforms in one frame. The average of each maximum amplitude value is obtained. Then, the average value of each frame is checked in units of several consecutive frames, and if the average value differs in each frame, the following processing is performed to set this to a certain constant value.

That is, the average value of the maximum amplitude values obtained as described above is obtained, and when the average value of several frames is defined as Ha, the unit waveform in each frame (the maximum amplitude of the unit waveform) is obtained. Each time the value is set to H), the portion of the waveform having a positive value is multiplied by Ha / H.

As a result, the envelope connecting the peak values in the section to be processed becomes flat like the original waveform. For example, the synthesized sound of the original waveform of the voice in which the vowel “Oi” continues for a long time is added to the throat. It can prevent the phenomenon of rumbling sound like something entangled. The same processing is performed on the negative side. However, if this process is performed, if the original waveform envelope originally has irregularities, the irregularities may be removed, but there is no problem in terms of hearing.

The processing method for making the envelope connecting the peak values of the processing target section flat like the original waveform is not limited to the method described above, but can be realized by using other methods. . Further, the above-described process of smoothing the envelope becomes more effective when the voiced sound frame is continuous for a predetermined number or more (for example, 20 frames or more) of that frame.

The program for realizing the present invention described above can be stored in a storage medium such as a floppy disk, and the present invention also includes the storage medium.

[0075]

As described above, according to the present invention,
Necessary for accurately synthesizing the information indicating the period and the intensity even for a voice that cannot be distinguished from the component Y2 indicating the cycle of the unit waveform in the residual waveform and the error component Y3, such as a female voice. Error component can be extracted, data can be compressed at a high compression rate, and a good synthesized sound can be obtained.

Also, in at least the first unit waveform in the frame where the unvoiced sound state changes to a voiced sound, the number of peaks of the waveform is counted, and it is determined whether or not the number of peaks is larger than a preset number. If you have more than
The information indicating the cycle, and only the information necessary for synthesis from the error component is efficiently selected and sent to the synthesis side, and when the number of peaks is less than a set value, the information indicating the cycle,
Since error component information is obtained not only from the minimum error component required for synthesis but also from information corresponding to other error components and sent to the synthesis side, when the amount of information necessary for voice synthesis is large By efficiently selecting only the minimum necessary information from the residual waveform, it is possible to increase the data compression ratio for obtaining the same level of synthesized speech.

Further, in the second and subsequent unit waveforms in the frame where the unvoiced sound state changes to the voiced sound, the pulse of the error component information (pulse train) accumulated by processing all the unit waveforms located before is converted into a mountain-shaped pulse. Since the voice synthesis processing is performed based on the information decomposed into the information and the information indicating the period, for example, when trying to obtain a synthesized sound for the voice waveform of “Ah”, the information amount is initially small, Even though the synthesized waveform is not very similar to the original waveform, the amount of information accumulates over time, so the synthesized waveform becomes very close to the original waveform, and as a whole, a better synthesized sound is obtained. Can be.

A predetermined number of signals having a large amplitude value are selected from at least the first unit waveform that changes from the unvoiced sound state to a voiced sound or a residual waveform corresponding to a unit waveform near a frame where the LPC coefficient greatly changes. When the direction of the amplitude of the original waveform corresponding to the signal to be selected goes from positive to negative, the error component information for the signal is set to negative error component information, and the original waveform corresponding to the signal to be selected is processed. When the amplitude direction of the signal goes from negative to positive, the error component information for the signal is set to be positive error component information, so that the sign of the pulse as the error component information to be extracted and the sign of the original waveform are significantly different. And a good synthesized sound can be obtained.

Further, according to the present invention, in a process of performing a speech synthesis process based on information indicating a cycle of a unit waveform and error component information, a peak value of an amplitude of each frame in a synthesized waveform is set to a certain value. By performing the setting process, the envelope connecting the peak values of each frame of the processing target section in the synthesized waveform becomes a flat shape similar to the original waveform, and for example, a voice in which a vowel such as “Oi” continues for a long time A phenomenon that the synthesized sound with respect to the original waveform becomes a gurgling sound with something entangled in the throat can be prevented, and an extremely good synthesized sound can be obtained.

[Brief description of the drawings]

FIG. 1 is a view for explaining processing of an embodiment of the present invention for extracting information necessary for speech synthesis from a speech waveform of a certain woman.

FIG. 2 is a flowchart illustrating basic processing according to the embodiment of the present invention.

FIG. 3 is a view for explaining a process of breaking information (pulses) necessary for speech synthesis extracted into a mountain shape in the embodiment of the present invention.

FIG. 4 is a diagram for explaining inconvenience due to differences in signs of respective amplitudes of a synthesized sound waveform with respect to an original waveform to be processed in the embodiment of the present invention.

FIG. 5 is a view for explaining a process for eliminating a problem caused by a difference in the sign of each amplitude of the synthesized sound waveform with respect to the original waveform to be processed.

FIG. 6 is a diagram illustrating a process of extracting information necessary for speech synthesis according to the number of peaks in a unit waveform according to the embodiment of the present invention.

FIG. 7 is a view for explaining processing in a case where irregularities occur in the envelope of the synthesized sound waveform as compared with the envelope of the original waveform to be processed in the embodiment of the present invention.

FIG. 8 is a view showing an example of a residual waveform for a male speech waveform and information necessary for speech synthesis obtained from the residual waveform.

FIG. 9 is a diagram showing an example of a residual waveform with respect to a female voice waveform.

[Explanation of symbols]

 10, 20,... Unit waveforms p11, p12,. B Point where the original waveform changes from negative to positive

Claims (11)

[Claims] 1. An original waveform to be processed is divided into frames at predetermined time intervals, and based on a waveform (hereinafter, referred to as a unit waveform) that is periodically repeated at substantially constant time intervals within one frame. While obtaining the linear prediction coefficient based on the correlation between each unit waveform, from the residual waveform indicating the error component in this linear prediction, information indicating the cycle and intensity of the unit waveform and other error component information are extracted, In a speech analysis / synthesis method that performs speech synthesis based on such information, an original waveform of one frame expressed on plane coordinates of a time axis and an amplitude axis which are orthogonal to each other is converted into a peak point of the waveform along the time axis. To obtain the amplitude value and time information of the peak point,
From these peak points, the peak value having the maximum amplitude in each unit waveform is obtained, and the point of the waveform having the maximum peak value in each unit waveform whose amplitude in the direction of the origin on the time axis is 0 is obtained. The information indicating the cycle of the unit waveform is extracted from the residual waveform in the vicinity of the point having the amplitude 0 as a reference and sent to the voice synthesis side, and the error component of the residual waveform corresponding to each of the unit waveforms is synthesized. A speech analysis / synthesis method, wherein necessary information is extracted as pulse information, and speech synthesis processing is performed based on the pulse information and the information indicating the period together with the linear prediction coefficient. 2. A process for extracting information indicating a cycle of a unit waveform from a residual waveform in the vicinity of the point having the amplitude 0 as a reference, wherein the point having the amplitude 0 is set as a cycle of the unit waveform. Item 4. The speech analysis / synthesis method according to Item 1. 3. A process for extracting information indicating a cycle of a unit waveform from a residual waveform in the vicinity thereof with reference to the point of amplitude 0, comprising: 2. The speech analysis / synthesis method according to claim 1, wherein information having a large absolute value is extracted, and information indicating a cycle is obtained based on the extracted information. 4. Extracting information necessary for synthesis from error components of a residual waveform corresponding to each of the unit waveforms as pulse information, and based on the pulse information and the information indicating the period together with the linear prediction coefficient. The process of performing voice synthesis processing by counting the number of peaks of the waveform in at least the first unit waveform in a frame where the state changes from unvoiced sound to voiced sound and in a frame where the linear prediction coefficient greatly changes,
Judge whether the number of peaks is larger than a preset number, and if the number is more than the preset number, efficiently select only the minimum information necessary for synthesis from error components and send it to the synthesis side as pulse information. Wherein when the number of peaks is less than a set value, not only the minimum pulse information necessary for the synthesis but also pulse information from other error components is obtained and sent to the synthesis side. The speech synthesis analysis method according to any one of claims 1 to 3. 5. A process of selecting only minimum information necessary for synthesis from the error component, comprising: selecting an error component of a residual waveform corresponding to at least a first unit waveform in a frame that changes from an unvoiced sound state to a voiced sound state. 5. The speech analysis / synthesis method according to claim 4, wherein a predetermined number is selected from the above in ascending order of the absolute value of the amplitude value, and the selected number is sent to the synthesis side as pulse information. 6. A process for extracting information necessary for synthesis as pulse information from an error component of a residual waveform corresponding to the leading unit waveform, the amplitude of the original waveform corresponding to the residual waveform to be selected being changed. When going from positive to negative, the error component information for the residual waveform is negative pulse information, and when the amplitude direction of the original waveform corresponding to the residual waveform to be selected goes from negative to positive, the residual 6. The speech analysis / synthesis method according to claim 4, wherein the error component information for the difference waveform is positive pulse information. 7. A signal necessary for synthesis is extracted as pulse information from error components of a residual waveform corresponding to each of the unit waveforms, and based on the pulse information and information indicating the period together with the linear prediction coefficient. In the process of performing the voice synthesis process, in the second and subsequent unit waveforms of the frame to be processed, the pulse information used to synthesize the immediately preceding unit waveform is used as the actual residual waveform of the unit waveform. From the residual waveform of the unit waveform thus subtracted, a predetermined number is selected in ascending order of the absolute value of the amplitude, and the selected pulse information is passed from the analysis side to the synthesis side. The pulse information passed from the
4. The speech analysis / synthesis method according to claim 1, wherein pulse information obtained by adding to the pulse information used for synthesizing the preceding unit waveform is used for synthesizing the unit waveform. 8. The process of subtracting the pulse information used for synthesizing the immediately preceding unit waveform from the actual residual waveform of the unit waveform is used for synthesizing the immediately preceding unit waveform. 8. The speech analysis / synthesis method according to claim 7, wherein each pulse constituting the pulse information is broken into a mountain shape, and the pulse information broken into the mountain shape is subtracted from an actual residual waveform of the unit waveform. 9. The pulse information passed from the analysis side,
The processing on the synthesis side for adding to the pulse information used for synthesizing the immediately preceding unit waveform is as follows: each pulse constituting the pulse information used for synthesizing the immediately preceding unit waveform is broken down into a mountain shape. 8. The speech analysis / synthesis method according to claim 7, wherein pulse information obtained by adding pulse information passed from the analysis side to the pulse information collapsed into the mountain shape is obtained. 10. A process for performing a speech synthesis process based on information indicating a cycle of a unit waveform and error component information, wherein a process for smoothing an envelope of the synthesized waveform is performed. 10. The speech analysis / synthesis method according to any one of 9 above. 11. The process of smoothing the envelope of the synthesized waveform includes setting a processing target section for the synthesized waveform, and for each frame in the processing target section, a maximum peak value for each unit waveform in the frame. 11. The speech analysis / synthesis method according to claim 10, wherein an average of the average is calculated, and a peak value of each unit waveform in the entire processing target section is set to a predetermined value obtained based on the obtained average peak value. .