JP3422716B2 - Speech rate conversion method and apparatus, and recording medium storing speech rate conversion program - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech speed conversion method and apparatus for converting only a speech speed while maintaining the quality of recorded voice and the voice quality of a speaker, and a recording medium storing a program.

[0002]

2. Description of the Related Art In a recorded voice database for accumulating word voices used in various automatic voice response devices and the like, it is desired that the voice speeds of the respective voice data contained therein be well matched. However, when the database becomes large, it is difficult for even a well-trained speaker to speak so that all speech data have the same speech speed, so only the speech speed is converted without degrading the sound quality of recorded speech data. Technology is required.

On the other hand, various voice speed conversion methods have been devised in order to listen to recorded voice data such as television broadcast in a short time and to play back slowly for elderly people.
Among them, there is a method of changing the reproduction time by appropriately inserting / deleting a voice waveform section as a method that is relatively simple and does not cause deterioration of the quality of the original sound and a change in the sound quality. Such a method can be roughly summarized as follows.

1. Distinction based on whether or not to distinguish silent sections, consonant sections, and vowel sections

Human voice can be roughly divided into the above three sections. Generally, when a person tries to speak fast / slow, the time length is adjusted in the order of a silent section, a vowel section, and a consonant section. Focusing on this, the speech to be processed is first divided into these three sections, and the time length is adjusted according to each section, or the time length adjustment is performed by the same method without distinguishing between them. .

2. Distinguishing by inserting / deleting waveform length

There are cases where the length of the section to be inserted / deleted is always set to a constant length, and where the length of the waveform of the basic period (hereinafter referred to as the pitch section waveform) or an integral multiple thereof. is there.

3. Distinguishing by insert / delete method

When performing insertion / deletion, a waveform in a certain section may be inserted / deleted as it is, or a plurality of cut-out waveforms may be multiplied by a window function and added together so that the section lengths are adjusted. .

[0010]

In all of the methods that have been devised in the past, the first purpose is to obtain a desired speech rate conversion rate, and not the method that places the highest priority on sound quality from the beginning.
Further, each of the above conventional techniques has the following problems.

1. Distinguishing between silent sections, consonant sections, and vowel sections

The method of adjusting the time length according to each section is similar to the human speech speed adjusting method, and a natural output voice can be obtained. However, in the conventional technique, the time length of each section is adjusted by a completely different method. For this reason, it is necessary to divide the voice waveform into each section first, and then perform the processing according to the section.

2,3. About the length of the waveform to be inserted / deleted and the insertion / deletion method

The basic cycle of human voice is 2.
It has a very wide range of 5 to 7 msec and 5 to 20 msec for men. Considering this, it is considered that smoother output speech can be obtained by changing the length of the waveform to be inserted / deleted according to the length of the pitch section waveform.

However, it is very difficult to accurately determine the fundamental period of the voice waveform. Therefore, in a simple period extraction method, the connected portion is often discontinuous, and in order to avoid this, a method of applying a window function to a plurality of cut waveforms and adding them so that they overlap each other adjusts the section length. It is often done. However, in this method, since multiplication, addition, etc. are performed over the entire waveform, the voice quality of the original sound may be impaired. On the other hand, if the voice waveform is analyzed in the frequency domain by Fourier transform, etc., or if the low-pass filter is applied to extract the basic period with high accuracy, the discontinuity of the connection part can be reduced even if the waveform is inserted or deleted as it is. However, since the amount of calculation is large, it lacks real-time performance and the device becomes large-scale.

The present invention has been made in view of the above problems, and it is an object of the present invention to perform speech speed conversion by a simple method while maintaining the original sound quality. Also, by changing the parameters to be given, it is possible to specify how much importance is attached to the sound quality, and it is possible to perform flexible speech speed conversion processing according to the application.

[0017]

In order to solve the above-mentioned problems, the invention according to claim 1 inserts / deletes a waveform (hereinafter referred to as a pitch section waveform) having a basic period of a voice waveform as appropriate. In the speech speed conversion method that expands or shortens only the stationary section of the voice waveform to convert the speech speed of the recorded voice,
The waveform similarity lower limit value is set as a parameter, the input speech waveform is sequentially cut out from the beginning in short intervals (hereinafter, the cut out waveform is referred to as a frame interval waveform), the waveform similarity of the frame interval waveform is calculated, and When the waveform similarity is larger than the waveform similarity lower limit value given as the parameter, it is considered that the frame section waveform is stationary,
Insert a pitch section waveform to the frame section waveform
It is characterized in that the pitch section wave is deleted.
When determining the point to insert or delete a shape,
For each time point in the interval, two laps adjacent to each other with the time point in between
Squared average value or sum of squared difference of pitch section waveform for period
Or calculate the average of absolute values or the sum of absolute values,
Insert / delete the pitch interval waveform at the point where
It is characterized by making it a point.

According to the second aspect of the invention, the speech waveform
Waveform of basic period (hereinafter referred to as pitch section waveform)
By inserting / deleting,
Speech speed conversion that expands and shortens only the speed to convert the speech speed of recorded voice
In the equipment, set the waveform similarity lower limit as a parameter.
And the input voice waveform in a short section from the beginning.
Cutout (Hereinafter, the cutout waveform is referred to as the frame section waveform.
Call), calculate the waveform similarity of the frame section waveform,
The waveform similarity is the waveform type given as the parameter.
If it is larger than the lower limit of similarity, the frame section waveform is
Considering that it is stationary, insert / delete the pitch section waveform.
And means for performing insertion / deletion of the pitch section waveform
When deciding which point to
Then, the pitch section for two cycles that are adjacent to each other across the time point.
Of the mean square of the difference, the sum of squares, or the absolute value of the waveform
The point at which the sum of averages or absolute values is calculated and which is the minimum
Be the point to insert / delete the pitch section waveform
It is a characteristic speech speed conversion device.

The invention according to claim 3 is the same as that of claim 1.
To execute the described speech speed conversion method using a computer
A computer-readable note that stores the speech speed conversion program.
It is a recording medium.

The speech rate conversion method according to the present invention is applied to the vowel section.
Inserting / deleting the pitch interval waveform only for the interval length
Adjust the overall length by stretching more.

However, as in the prior art, first the vowel section
The extraction process is not necessary, and the waveform is
By using the fact that it is stationary,
Calculate the autocorrelation coefficient to determine if the frame section is a vowel section.
It is determined whether or not the basic cycle is obtained at the same time.

Since only the stationary portion of the waveform is processed, there are few errors in the extraction of the basic period, and the connected portion is less likely to be discontinuous. Therefore, noise is unlikely to be mixed in without performing processing such as applying a window function.

[0023]

1 is a block diagram showing an example of an embodiment of a speech speed conversion apparatus according to the present invention. The speech speed converter 1 according to the present invention comprises a voice waveform input unit 2 and a threshold value input unit 3.
It comprises a speech rate conversion rate input unit 4, a frame cutout unit 5, an autocorrelation coefficient calculation unit 6, a processing judgment unit 7, a waveform insertion / deletion unit 8 and a voice waveform output unit 9.

The voice input unit 2 acquires a voice whose voice speed is desired to be converted. The threshold value input unit 3 acquires the lower limit of the autocorrelation coefficient indicating the similarity of the waveform, and sends it to the processing determination unit 7. The speech rate conversion rate input unit 4 acquires a speech rate conversion rate representing how much the speech rate is to be increased / decreased, and sends it to the waveform insertion / deletion unit 8.

The frame cut-out unit 5 is responsive to the frame start point sent from the waveform insertion / deletion unit 8 to the voice input unit 2
A part (hereinafter referred to as a frame section waveform) of the speech waveform obtained in step S1 is cut out. If the frame segment waveform to be cut out is the processing target, it is sent to the autocorrelation coefficient calculation unit 6 and the waveform insertion / deletion unit 8, and if it is not the processing target, it is sent to the speech waveform processing output unit 9.

The autocorrelation coefficient calculation unit 6 calculates the autocorrelation coefficient of the frame section waveform sent from the frame cutout unit 5 and gives the maximum value of the autocorrelation coefficient to the processing judgment unit 7. The shift of the frame section waveform is sent to the waveform insertion / deletion unit 8 as the length of the basic cycle. Whether the processing determination unit 7 compares the maximum value of the autocorrelation coefficient sent from the autocorrelation coefficient calculation unit 6 with the lower limit of the correlation coefficient sent from the threshold value input unit 3 to insert / delete the waveform. Is determined and the result of the determination is sent to the waveform insertion / deletion unit 8.

The waveform insertion / deletion unit 8 receives the determination result sent from the processing determination unit 7 as to whether or not to perform the insertion / deletion, and if necessary, with respect to the frame section waveform sent from the frame cutout unit 5. Then, the waveform of the basic cycle (hereinafter referred to as the pitch section waveform) is inserted / deleted and sent to the voice waveform output unit 9. The waveform insertion / deletion unit 8 also uses the speech rate conversion rate sent from the speech rate conversion rate input unit 4 and the basic period length sent from the autocorrelation coefficient calculation unit 6 to start the next frame. Is determined and sent to the frame cutout unit 5.

The voice waveform output section 9 is a frame cutting section 5
Alternatively, the frame section waveforms sent from the waveform inserting / deleting unit 8 are sequentially stored or accumulated until the processing of the input voice is completed, and then output.

FIG. 2 is a flow chart for explaining the operation of the speech speed conversion apparatus 1 according to the present invention shown in FIG. Hereinafter, an operation example of the speech speed conversion device when the speech speed of voice is specifically multiplied by α will be described with reference to the block diagram of FIG. 1 and the flowchart of FIG.

The voice input unit 2 acquires a voice whose voice speed is desired to be converted, and if necessary, a digital signal x (n); n = 0,
1, 2, ... (Step S1). Here, n is an index corresponding to time. The threshold value input unit 3 acquires the lower limit value β of the autocorrelation coefficient representing the degree of similarity of the waveform and sends it to the processing determination unit 7 (step S2).

In order to prevent the deterioration of the sound quality such as the inclusion of noise, it is desirable that the insertion / deletion of the waveform is performed in the stationary portion of the waveform. Generally, when the waveform is stationary, the autocorrelation coefficient is close to 1. . The lower limit value β of the autocorrelation coefficient acquired by the threshold value input unit 3 is the insertion / insertion of the waveform of the basic cycle (hereinafter referred to as the pitch section waveform) when the waveform to be processed is stationary.
This is to specify whether to delete. If β is made small, insertion / deletion is performed even in a portion where the waveform is not stationary, so noise is likely to be mixed, but an output close to the target speech speed can be obtained. On the contrary, when β is close to 1, the risk of noise mixing can be suppressed. By setting the value of the lower limit value β of the autocorrelation coefficient according to the purpose of the voice speed conversion, it is possible to perform flexible processing such as "changing the voice speed within a range that does not deteriorate the quality". For example, β is set to 0.7 here.

The voice speed conversion rate input unit 4 acquires a voice speed conversion rate α indicating how fast or slow the voice speed is, and sends it to the waveform inserting / deleting unit 8 (step S3). For example, here, considering the case of increasing the speech speed to 1.2 times, the speech speed conversion rate α =
Set to 1.2.

The frame cutout unit 5 cuts out a part of the voice waveform (hereinafter referred to as a frame section waveform) acquired by the voice input unit 2 according to the frame start point n _bi sent from the waveform insertion / deletion unit 8. The speech speed conversion apparatus sequentially processes the speech waveform from the beginning for each frame section. If the sent frame start point n _bi is larger than the end point n _{e (i-1) of} the frame segment waveform cut out last time, first

[0034]

[Equation 1]

Is cut out and sent directly to the speech waveform output unit 9 (step S6), and then a fixed length M starting from n _bi
Sample waveform

[0036]

[Equation 2]

The signal is cut out and sent to the autocorrelation coefficient calculation unit 6 and the waveform insertion / deletion unit 8 (step S7). Note that n _bi is i at the index n of the digital signal x (n).
N _{b (i-1)} is the start point of the (i−1) th frame, n _ei is the end point of the i th frame,
_{ne (i-1)} represents the end point of the (i-1) th frame.

Frame start point n sent_biBut last time
End point n of the generated frame section waveform _{e (i-1)}If
Immediately n_biWaveform of M samples starting with x
(N_bi) ~ X (n_ei) Is cut out to calculate the autocorrelation coefficient
6 and the waveform insertion / deletion unit 8 (step S7).

FIG. 3 shows an example of a frame section waveform having a length of M samples. In this example, the sampling frequency is 16
This is a cut-out of a part of the audio data of kHz, and M
= 240, in this case 15 msec.

The autocorrelation coefficient calculation unit 6 receives the frame section waveforms x (n _bi ) to x (n) sent from the frame cutout unit 5.
_ei ) autocorrelation coefficient y (m) is calculated (step S
8). y (m) is the calculation of only the frame section waveform here,

[0041]

[Equation 3] It is defined as

FIG. 4A shows an image of calculating the autocorrelation coefficient, and FIG. 4B shows the autocorrelation coefficient y (m) of the frame section waveform of FIG. The autocorrelation coefficient indicates how similar a certain waveform is to a waveform obtained by shifting the waveform by m. When the frame section waveform is periodic, the two waveforms are similar when they are shifted by an integer multiple of the basic period T, and y
(M) has a large value. On the contrary, if the waveform is not very periodic, there is no similar part when the waveform is shifted, so y
The value of (m) is small. As an example in which the waveform is not periodic, FIGS. 5 and 6 show another frame section waveform and its autocorrelation coefficient.

After calculating y (m), the autocorrelation coefficient calculation unit searches for the maximum value y (T) within the frame section, and the maximum value y (T) of the autocorrelation coefficient is sent to the processing determination unit 7. , The shift of the frame section waveform giving the maximum value, that is, the length T of the basic period is sent to the waveform inserting / deleting unit 8 (step S9). In the example of FIG. 3, y (T) = 0.75 and T = 55, and in the example of FIG. 5, y (T) = 0.56 and T = 53.

The processing determination section 7 is a threshold value input section 3 and the maximum value y (T) of the autocorrelation coefficient sent from the autocorrelation coefficient calculation section 6.
The lower limit value β of the autocorrelation coefficient sent from is compared (step S10). If the value of y (T) is large, the waveform is stationary, so that noise is unlikely to be mixed even if waveform processing is performed. In the example of FIG. 3, since y (T) = 0.75> 0.7 = β, it is determined to insert / delete the pitch section waveform. In the example of FIG. 5, since y (T) = 0.56 <0.7 = β, it is determined not to insert / delete the pitch section waveform. For example, a signal indicating the determination result is sent to the waveform insertion / deletion unit 8 as 1 when the insertion / deletion is performed and 0 when the insertion / deletion is not performed.

The waveform insertion / deletion unit 8 first receives the determination signal sent from the processing determination unit 7 as to whether or not to perform the insertion / deletion, and operates differently. When the insertion / deletion is not performed, the next frame start point n _{b (i + 1)} is determined as the next point of this frame, that is, n _ei +1 and sent to the frame cutout unit 5 (step S11), and the frame section The waveform is sent to the audio waveform output unit 9 without any processing (step S23). In the case of inserting / deleting, first, using the length T of the basic period sent from the autocorrelation coefficient calculation unit 6, a point P where the pitch section waveform is inserted / deleted in the frame section is searched (step S12). ). The point P is a point between the two most similar periods in the frame section. For example, m
= N _bi + T to n _ei -T

[0046]

[Equation 4]

And m such that z (m) is minimized
Be P. The calculation of z (m) is not the above definition, that is, the calculation of the sum of squares of the difference between the pitch section waveforms for two cycles that are adjacent to each other with the time point sandwiched, | x (n) −x (n +
It is also possible to calculate the average value, the root mean square value or the sum of absolute values of T) |. FIG. 7 shows how P is searched in the frame section waveform shown in FIG.

Next, it is determined whether the pitch section waveform is to be inserted or deleted depending on whether the speech rate conversion rate α sent from the speech rate conversion rate input unit 4 is larger or smaller than 1 (step S13). If α> 1, the speech speed is increased, so x
(P + 1) to x (P + T) are deleted from the frame section waveform (step S14). When α <1, the speech speed is slowed down. Therefore, x (P + 1) to x (P + T) are once stored in the buffer, and they are stored in the frame section waveforms x (P) and x (P).
It is inserted between (+1) (step S15).

Next, the frame movement amount M'is calculated from α and T (step S16). (M '± T) / M' = 1 /
From α,

[0050]

[Equation 5]

It is sufficient to calculate Next frame start point n
_{b (i + 1)} is basically n _bi + M '+ 1 (step S
18). However, when the next frame start point is before the point where the waveform is inserted / deleted, that is, n _bi +
If M '<P + T, P + T + 1 is set as the next frame start point _{nb (i + 1)} (steps S17 and S19). The n _{b (i + 1)} thus determined is sent to the frame cutout unit 5.

Next, of the processed frame section waveforms,
Frame start point n_biTo the next frame start point n_{b (i + 1)}−
Outputs the frame section waveform that processed the waveforms up to 1 as a voice waveform
It is sent to the force unit 9 (steps S21 and S22). However, next
If the frame start point of is outside the frame section,
Side n_{b (i + 1)}-1> n_eiFrame interval x
(N _bi) ~ X (n_ei) Is output (steps S20, S
23). In this case, the remaining x (n_ei+1) to x (n
_{b (i + 1)}) Is as it is before cutting out the next frame section
It is output (S6).

Examples of how to proceed the frame and output range are shown in FIG.
As shown in FIG. For example, in the frame section of T = 55 and M = 240, when α = 1.2 (FIG. 8),

[0054]

[Equation 6]

Since P + T <n _ei <n _bi + M ', n _bi +331 is sent to the frame cutout unit 5 as the next frame start point n _{b (i + 1)} , and x (n _bi ) to x ( n _ei )
The waveform from which 55 samples are deleted during one cycle is sent to the voice waveform output unit 9 (steps S17 → S18 → S20 → S2).
3).

When α = 1.35 (FIG. 9),

[0057]

[Equation 7]

Since P + T <n _bi + M '<n _ei , n _bi +213 is sent to the frame cutout unit 5 as the next frame start point n _{b (i + 1)} , and x (n _bi ) to x ( n _bi +
The waveform obtained by deleting 55 samples for one cycle in the middle of M ′) is sent to the speech waveform output unit 9 (steps S17 → S18 → S20).
→ S21).

When α = 1.5 (FIG. 10),

[0060]

[Equation 8]

Since n _bi + M '<P + T <n _ei , P + 55 + 1 is sent to the frame cutout unit 5 as the next frame start point n _{b (i + 1)} , and x (n _bi ) to x (P +
The processed part of T), in this case x (P + 1)-
Since x (P + T) is deleted, x (n _bi ) to x (P)
To the voice waveform output unit 9 (steps S17 → S19 →
S22).

In this example, a point P at which a waveform is inserted / deleted
When searching for, the entire frame section was searched, but if the range of this search is set to one cycle of T, then the next frame start point n _{b (i + 1)} will be before P + T. The possibility is reduced, and a conversion rate close to the desired speech rate conversion rate is obtained. The voice waveform output unit 9 sequentially outputs the frame section waveforms sent from the frame cutting unit 5 or the waveform inserting / deleting unit 8. It is determined whether all the input voices have been output (step S24). If the output has not been completed, the process returns to step S4 to start the next frame processing.

When the voice waveform output section 9 sequentially outputs the transmitted frame section waveforms, the input voice is processed one after another for each frame section, so that it is possible to output in real time to the input. The present invention can be applied to, for example, audio output when changing the playback speed of a VTR (video tape recorder). In applications where real-time processing is not required, such as conversion of the voice speed of each voice data in the voice database, the voice waveform output unit 9 accumulates the frame section waveform until all the processing of the input voice is completed, and then, if necessary. The connection portion may be smoothed and output.

As described above, the speech speed conversion apparatus according to the present invention acquires two parameters, that is, the speech speed conversion rate α and the lower limit value β of the autocorrelation coefficient, and sets the speech speed as high as possible with the sound quality indicated by β. I try to get closer to α times. By these methods of giving, it becomes possible to perform flexible processing according to the application area. For example β
If the value is set to a value close to 1 such as 0.8, the processing will be performed within the range in which the voice speed can be converted while maintaining the voice quality, so that the high voice quality such as the voice speed uniformity of the recorded voice database is required. Suitable for various uses.

Further, without giving α, it is instructed only to increase or decrease the speech speed, and steps S16, S17 and S1 are performed.
8, the next frame start point n determined in S20
_{If b (i + 1)} is always set to P + 1, it is possible to perform processing of making the sound quality at the level designated by β as fast / slow as possible.

Moreover, it is only necessary to understand the recorded contents,
In applications where sound quality is not so important, if an extremely small value such as 0.1 is given, insertion / deletion processing will be performed if the waveform is a periodic frame section, and an output close to the given α can be obtained. .

The embodiment of the present invention is not limited to the configuration shown in FIG. 1 and FIG.
One or both of the speech rate conversion rate α and the lower limit value β of the autocorrelation coefficient can be set in advance, or can be changed by selecting from preset values. is there. Further, the speech speed conversion device of the present invention can be realized not only by hardware using a combinational logic circuit or the like, but also by using a computer and software executed thereby. Further, this software can be distributed via a computer-readable recording medium or a communication line.

[0068]

As is apparent from the above description, according to the voice speed conversion method of the present invention, the voice speed conversion can be performed while maintaining the sound quality of the original sound. Despite the simple method, there is little noise mixing. Further, how much importance is attached to the sound quality can be designated by giving the lower limit value β of the autocorrelation coefficient, and the flexible speech speed conversion processing according to the application can be performed.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of an embodiment of a speech speed conversion apparatus according to the present invention.

FIG. 2 is a flowchart illustrating the operation of the speech speed conversion device according to the present invention.

FIG. 3 is a diagram showing an example of a frame section waveform.

FIG. 4 is a diagram for explaining an example (a) of a method of calculating an autocorrelation coefficient (b) for the frame section waveform shown in FIG.

FIG. 5 is a diagram showing another example of a frame section waveform.

6 is a diagram showing an autocorrelation coefficient of the frame section waveform shown in FIG.

7 is a diagram illustrating a method of determining a point P at which a waveform is inserted / deleted in the frame section waveform shown in FIG.

FIG. 8 is a diagram showing an example of a range of a waveform to be output and a start point of a next frame (when α = 1.2).

FIG. 9 is a diagram showing an example of a waveform range to be output and a start point of a next frame (when α = 1.35).

FIG. 10 is a diagram showing an example of a range of a waveform to be output and a start point of a next frame (when α = 1.5).

[Explanation of symbols]

1 Speech rate converter 2 Voice waveform input section 3 Threshold input section 4 Speech rate conversion rate input section 5 frame cutout 6 Autocorrelation coefficient calculator 7 Processing judgment section 8 Waveform insertion / deletion section 9 Voice waveform output section

─────────────────────────────────────────────────── ─── Continuation of the front page (56) Reference JP-A-8-292789 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 21/04

Claims (3)

(57) [Claims] 1. By appropriately inserting and deleting a waveform of a fundamental period of a speech waveform (hereinafter referred to as a pitch section waveform),
In the speech rate conversion method that expands and shortens only the stationary section of the speech waveform to convert the speech rate of the recorded speech, set the waveform similarity lower limit value as a parameter, and cut the input speech waveform sequentially from the beginning into shorter sections. However, (hereinafter, the cut-out waveform is referred to as a frame section waveform), the waveform similarity of the frame section waveform is calculated, and when the waveform similarity is higher than the waveform similarity lower limit value given as the parameter, the frame regarded as the interval waveform is constant, and depending on whether speech rate conversion rate to be input is less than or greater 1 characterized by performing the insertion and deletion of pitch period waveform with respect to the frame sections waveform, the pitch
When determining the point at which the
Each time point in the ram section is adjacent to the other time point.
The mean square value of the difference between the pitch period waveforms for two cycles
Is the sum of squares or the average of absolute values or the sum of absolute values
Insert the pitch section waveform at the point where this is the minimum.
A speech speed conversion method characterized in that points are deleted . 2. A waveform of a fundamental period of a voice waveform (hereinafter referred to as "pit
(Called H section waveform)
Recorded sound by expanding and shortening only the stationary section of the sound waveform
In the speech speed conversion device for converting the speech speed of, the means for setting the waveform similarity lower limit value as a parameter and the input speech waveform are sequentially cut out in short sections from the beginning ( below.
The waveform cut out below is called the frame section waveform),
Calculate the waveform similarity of the frame section waveform and
Is the lower limit of the waveform similarity given as the parameter.
The frame section waveform is stationary when it is larger than
And the input speech rate conversion rate is less than 1
A method to insert / delete the pitch section waveform depending on whether it is large or not.
The pitch interval waveform is inserted.
When determining the point to enter / delete,
For each time point, two adjacent periods with the time point in between
Root mean squared difference or sum of squares of pitch interval waveform or
Calculates the average of absolute values or the sum of absolute values, and this is the minimum
Is the point where the pitch section waveform is inserted or deleted.
A speech speed conversion device characterized in that 3. The speech speed conversion method according to claim 1 is calculated.
Stores a speech speed conversion program for execution using a
Computer readable recording medium.