JPH10282991A - Speech rate converting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To facilitate calculation of a correlative function by presuming pitch lengths of a head segment and the next segment the same ratio, overlapping both segments, shifting so that waveforms are matched with each other and evaluating the correlation function in the shifted state. SOLUTION: A first correlater 41 calculates the correlative function between the data in a first memory 38 and a second memory 39, and obtains a first correlative delay amount obtained by shifting relative positions so that the large positions of the correlative function of two segments coincide with each other. On the other hand, a waveform deviation value calculator 42 obtains pitch information from a pitch buffer 34, and presumes that the waveforms of the head segment and the next segment have the same pitch, and only different phases, and obtains a deviation value shifting so as to be matched with each other by shifting. A second correlater 43 calculates the correlative function based on the data of the first memory 38 and the second memory 39 and the deviation value from the waveform deviation value calculator 42, and obtains a second correlative delay amount obtained by shifting the relative positions so that the large positions of the correlative function of two segments coincide with each other.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice speed conversion device for changing the speed of a coded voice without changing its pitch.

[0002]

2. Description of the Related Art In a message message of an answering machine or a video recording a television picture, a user wants to play at a normal speed what he wants to hear or watch, and to play it at a high speed in places that are not important or in which he is not interested. . When playing back meeting recordings, it is important to play back important parts at a slow speed. Techniques have been developed to change only the speed of voice for such purpose without changing the pitch (frequency) of the sound.

[0003] Japanese Patent Application Laid-Open No. 4-104200 discloses such an audio speed converter. In this method, voice is expressed in units of segments, one block is composed of a plurality of segments, and the number of segments in this block is increased or decreased to make the speed low or high. In order to minimize the unnaturalness of the sound due to the speed change, when the speed is increased, the first and second segments are weighted and added to form one segment, and the third and subsequent segments are successively added thereto. One block segment is reduced by one to increase the speed. When the speed is to be reduced, the first and second segments are weighted and added to form one segment, the second and subsequent segments are continued, and one segment following this is added after the last segment of the block. By adding, the number of segments in one block is increased by one to reduce the speed.

FIG. 17 is a block diagram showing a voice speed converter disclosed in Japanese Patent Laid-Open No. 4-104200. The voice speed conversion is performed based on the speed ratio α. α = (reproduction time length after speed conversion) / (reproduction time length before speed reproduction) (1) The input signal is converted into a digital signal by the A / D converter 11, and
The data is stored in the data buffer 12. As shown in FIG. 2A, audio data often has a continuous waveform having a pitch L. For this reason, as shown in (B), a segment in which data within the pitch L is set is set, and a plurality of these segments are grouped into one block, and the number of segments in this block is increased or decreased to increase or decrease the voice speed. .

The speed control circuit 13 controls the divider 14 to store the data of the first segment of the block in the first memory 15.
And the data of the next segment is stored in the second memory 16. The correlator 17 includes a first memory 15 and a second memory 16
The calculation of the correlation function of the data is performed to calculate the correlation delay amount (rk). The correlation delay amount rk is rk + or rk- shown in FIG. 3B or 3C. (A) shows a speech waveform in each segment. a or c based on the waveform of b
Rk is the amount by which the waveform of b is deviated from the waveform of b, and (+) is the case where the waveform is shifted to the right (later in time) with respect to b.
A case shifted to the left side (time forward) is defined as (-).

The window function generator 20 includes a first memory 15 and a second memory 15.
A weight coefficient for the data in the memory 16 is generated. The weight coefficient is based on the speed ratio α and the correlation function generated by the correlator 17.
A window function to increase or decrease is generated, and the sum of the respective weighting factors is set to 1.0 for each corresponding data in the first memory 15 and the second memory 16.

The first multiplier 18 multiplies the data in the first memory 15 by the weighting coefficient, and the second multiplier 19 multiplies the data in the second memory 16.
Is multiplied by the weight coefficient, and the result is output to the adder 21. The adder 21 adds one to the position having a high correlation by shifting it by rk based on the correlation delay amount (rk) of the correlator 17 as shown in FIG. I do. The speed control circuit 13 outputs the signal from the divider 14 to the second and subsequent segments or the third
The segments 22 and subsequent segments are output, and the combiner 22 outputs this value to the D / A converter 23 as one block of data following the added value. The D / A converter 23 converts the digital signal into an analog signal and outputs the analog signal.

FIG. 18 shows the result of voice speed conversion when the speed ratio α is 3/4 and 3/2. (A) to (d) show the speed ratio α
Shows the case where 3/4, and (e) to (h) show the case where α = 3/2. (A) shows original audio data in which one block contains four segments. (B) is the correlation delay amount rk
Is 0, that is, the waveforms of segment 1 and segment 2 are the same. At this time, the added value of the first segment and the second segment is the same as the length of the original segment, and the length of one block is ／. The oblique line of the first segment indicates a weighting factor, and the weighting factor for the original first segment gradually decreases from 1 to 0, and the weighting factor for the original second segment gradually increases from 0 to 1. As for the length of the segment, the length of the segment generated by the addition is increased or decreased, but the others are fixed at ts.

FIG. 3C shows the case where the correlation delay amount rk is (+), which is the case of FIG. 3B. In this case, the segment obtained by adding the first segment and the second segment is longer than the original segment length ts by rk. (D)
Is the case where the correlation delay amount rk is (-), and the segment obtained by adding the first segment and the second segment is shorter than the original segment length ts by rk as shown in FIG. 3C. Therefore, in the cases (c) and (d), the speed ratio α is 3
It does not become / 4. The data pointers P1 and P2 indicate the start addresses of the data input to the first memory 15 and the second memory 16, and P1 indicates the next position of the length of 3 segments (3 ts) as shown in FIG. Indicates the start position of the first segment of the next block, and is the same position in (b) to (d). P2 indicates the start position of the second segment of the next block, and in this case, indicates the start address of segment 6.

(E) to (h) show the case where the speed ratio α is 3/2, and (e) shows that one block consists of two segments,
It is shown that a segment 3 is added to this to make the speed ratio α 3/2. (F) is a case where the correlation delay amount rk is 0, the first segment and the second segment are added to have the same length as the original segment length ts, and the second segment and the segment 3 of the next block are added. Thus, the length of one block is set to the length of three segments (3 ts). (G)
Is the case where rk is (+) and the length of the added segment is 1
It is larger than the segment length ts. (H) is rk
Is (-), the length of the added segment is shorter than the length ts of one segment. As a result, one block becomes longer than 3 ts in (g), and becomes shorter in (h). The pointer P1 is after the segment 2, and the pointer P2 indicates the start position of the start segment of the next block after the segment 3 in (f).

[0011]

The amount of calculation of the correlator for calculating the correlation function is enormous, and a large-capacity calculator is required.
Also, as shown in FIGS. 18C and 18G, the correlation delay amount r
If k + is large, the pointers P1 and P2 are separated,
When the similarity between the data in the first memory and the data in the second memory is impaired, the accuracy of the correlation delay amount output from the correlator is lost, and the quality of the reproduced sound is degraded. In addition, since the segment length ts is fixed except for the segments generated by addition, if the pitch of the audio data is longer than the segment length ts, the similarity between the first segment and the next segment of the block cannot be compared with high accuracy. There is also a problem.

The present invention has been made in view of the above problems, and has as its object to facilitate calculation of a correlation function of a correlator. Further, the position of the pointer is set to indicate the start position of the first segment and the second segment of the next block,
Make sure that the two pointers are not too far apart. It is another object of the present invention to make the length of the segment variable when the pitch length of the audio data is longer than the segment.

[0013]

In the voice speed converter of the present invention, the pitch length of the first segment of a block and the next segment are assumed to have the same ratio, and both segments are overlapped and shifted so that their waveforms match. The shift amount is obtained, and then the correlation function of both segments is calculated in a state shifted by the correlator.

According to the present invention, the calculation of the correlation function between two adjacent segments is facilitated.

[0015]

According to the first aspect of the present invention, there is provided a decoder for decoding coded audio data, and a temporal information in which pitch information of audio data obtained from the decoder and data within one pitch are adjacent to each other. A pitch storage unit for storing pitch polarity information indicating whether the data is similar to the data of the subsequent pitch or the data of the previous pitch, and indicates whether the pitch of the audio data obtained from the decoder is clear or unknown. And a flag storage unit for storing the generated flag, and divides the data from the decoder into segments of a predetermined length, and sets a predetermined number of continuous segments into one block, and divides and outputs the segments of each block in the order of input. A divider, a first memory for storing a first segment of each block output from the divider, and a second memory for storing a second segment of each block output from the divider. Calculated a second memory, the correlation function between the contents of the content and the second memory of the first memory, the first
A first correlator for calculating a correlation delay amount, and a waveform shift for calculating a shift amount by shifting the waveform of the first segment from the data of the pitch storage unit so that the waveform of the next segment substantially coincides with the waveform of the next segment. An amount calculator, a second correlator that calculates a correlation function between the contents of the first memory and the contents of the second memory based on the output of the waveform shift amount calculator and calculates a second correlation delay amount; A window function generator for generating a weighting factor for the data of the segment; a first multiplier for multiplying the content of the first memory by the weighting factor;
A second multiplier for multiplying the content of the memory by the weighting factor;
An adder for adding the output of the first multiplier and the output of the second multiplier at a position where the value of the correlation function is large based on the output of the first correlator or the output of the second correlator; A switch for switching and outputting the outputs of the adder and the second correlator in accordance with the flag in the flag storage unit, and the output of the adder and, subsequently, when the speed is reduced from the divider, From the second segment onwards, when speeding up, the third
A combiner that outputs a through section segment composed of segments subsequent to the segment, and a speed control unit that controls the divider and the combiner based on a speed ratio that represents a conversion ratio of the speed of audio.

[0016] Decoding the encoded data,
The pitch information, the pitch polarity information, and the flag indicating whether the pitch is clearly shown are obtained. The audio data is divided into segments, and a predetermined number of segments are made into one block, and audio speed conversion is performed in units of this block. The speed conversion is performed based on the speed ratio α obtained by dividing the converted playback time length by the playback time length before conversion, and the number of segments in one block is predetermined according to the speed ratio α. The speed control unit controls the divider to output the first segment of the block to the first memory, outputs the second segment to the second memory, and outputs the second and subsequent segments or the third segment according to the speed reduction or the speed increase. The segment and subsequent segments are output to the combiner. The first correlator calculates a correlation function between the first segment and the second segment, and calculates a first correlation delay amount indicating a distance between highly correlated positions of both segments.

The waveform shift amount calculator calculates the shift amount by shifting the waveform of the first segment and the waveform of the second segment so as to substantially coincide with each other based on the pitch information. If the waveform and the phase of both segments are the same, the shift amount is zero. The second correlator calculates the correlation function between the first segment of the first memory and the second segment of the second memory in such a state that the waveforms of both segments are shifted so as to substantially coincide with each other, so that the correlation is high. The search range for finding the position is greatly narrowed, and the amount of calculation is greatly reduced. This makes it possible to easily calculate the second correlation delay amount.

The window function generator generates a weight coefficient for each of the data of the first segment and the data of the second segment by a predetermined window function. The weighting factor is such that, when both segments are added to form a new added segment, the data size of the added segment is the same as the data of the original segment, so that the addition does not make the sound unnatural. It is provided in order to. Therefore, the sum of one weight coefficient and the other weight coefficient of the data to be added in both segments is 1.0. The weight coefficient of one segment increases gradually from the head of the segment to the rear end, and the weight coefficient of the other segment decreases gradually from the head of the segment to the rear end.

The first multiplier multiplies the data in the first memory by the weighting factor, and the second multiplier multiplies the data in the second memory by the weighting factor. The changeover switch outputs the first correlation delay amount or the second correlation delay amount according to the flag in the flag storage unit.
The adder outputs the output of the first and second multipliers to the first or second output.
Based on the correlation delay amount, one is shifted to a position where the correlation function is large, and the addition is performed. The combiner continuously outputs the through segment input from the divider after the addition segment obtained by the addition as one block of data.
Thus, when the second correlator is used, the calculation of the correlation delay amount can be performed quickly.

According to the second aspect of the present invention, the speed control section includes:
The length of the segment is changed based on the pitch information.

The pitch of voice data differs between a female voice and a male voice. If the length of a segment is fixed, the correlation value between segments deteriorates in the case of data having a pitch exceeding the length of the segment. Therefore, by making the length of the segment variable according to the pitch, the correlation value is improved. That is, the sound quality is improved.

According to the third aspect of the present invention, the length of the segment is set to an initial segment length based on a female voice, and when the pitch of the pitch information exceeds the initial segment length, the length of the segment To the initial segment 2
Double it.

The length of the initial (default) segment is set to a length based on the female voice, and if the pitch exceeds this initial segment, the length is set to a segment twice as long as the initial segment. Accordingly, the length of the block is also doubled. The number of segments in one block does not change. As a result, data of one pitch or more always exists in one segment, and the correlation function of two segments can be calculated.

According to a fourth aspect of the present invention, in the block in which the length of the segment is switched, the switch outputs the output of the first correlator to the adder.

In the block in which the length of the segment is switched, the pitch is a part that is slightly unstable, so that the waveform shift amount calculator becomes inaccurate. Therefore, the correlation delay amount of the first correlator is used. Prevent audio quality degradation.

In the invention according to claim 5, the changeover switch outputs the output of the second correlator to the adder when the pitch of the audio data in the flag storage section is clear.

When the pitch is unclear, the calculation of the shift amount becomes inaccurate in the waveform shift amount calculator, so that the addition is performed using the output of the first correlator. When the pitch is clear, the amount of deviation can be calculated quite accurately, so that the second correlation delay amount is quickly calculated using the output of the second correlator.

According to the invention of claim 6, when the data of the first segment and the next segment of the block are similar based on the pitch polarity information, the changeover switch outputs the output of the second correlator to the adder. Output.

When the data of the first segment of the block and the data of the next segment are similar, the waveform shift amount calculator can accurately shift the waveforms of both segments so that they coincide with each other. I get it. As a result, the second correlation delay amount of the second correlator can be accurately and quickly obtained, and the adder performs an addition process based on the second correlation delay amount.

In the present invention, twice the initial segment is set as a new segment, the first new segment is composed of a first initial segment and a second initial segment in the order of data transmission, and the next new segment is a second segment in the order of data transmission. The first initial segment is similar to the second initial segment, the second initial segment is similar to the third initial segment, and the fourth initial segment is similar to the third initial segment. If so, the changeover switch outputs the output of the second correlator to the adder.

The first initial segment is similar to the second initial segment, the second initial segment is similar to the third initial segment, and the third initial segment is similar to the fourth initial segment. Will be similar. As a result, the waveforms of the two new segments can be accurately shifted by the waveform shift amount calculator so that the waveforms of the two new segments match, and an accurate shift amount is obtained. Since the second correlation delay amount can be accurately and quickly obtained by the second correlator using the shift amount, the adder adds the second correlation delay amount.

In the present invention, the head of data input to the first memory and the second memory indicates a head address of the segment.

Conventionally, as described with reference to FIG.
Is less than or equal to 1, the pointer P1 is located at a position that is an integral multiple of the segment length ts. On the other hand, the pointer P2 is the second block of the next block.
As shown in (c), when the distance between P1 and P2 increases, the similarity between the segments decreases, and the sum of data starting with P1 and data starting with P2, and The continuity of subsequent data deteriorates, resulting in an unnatural reproduction sound as voice. By setting P1 as the head address of the head segment of the next block and P2 as the head address of the next segment, the length of the head segment is always between P1 and P2, and it is possible to prevent the similarity from being reduced due to being too far apart. .

According to a ninth aspect of the present invention, a correlation delay amount integrating means for integrating the first correlation delay amount and the second correlation delay amount, and a delay flag is generated when the integrated value substantially reaches the segment length. A delay flag generator, wherein the speed control unit adds or deletes a segment of a through section of the block based on the delay flag.

When the correlation delay amount rk is positive, the length of one block is longer by rk than the block length at which rk is 0 and the speed ratio α has a predetermined value. ((C) of FIG. 18)
In the case of (1), the rk of each block is integrated, and when the rk reaches the +1 default segment length or more, one segment in the through section is deleted. When rk is negative, the speed ratio α is shorter by rk than the length of the block where the speed ratio α has a predetermined value. (In the case of (d) in FIG. 18) The rk of each block is integrated, and when the length becomes equal to or smaller than the −1 default segment length, one segment of the through section is added. This is one by one
Even if one block cannot maintain the speed ratio α, the speed ratio α can be maintained considering the number of blocks in a certain range.

According to the tenth aspect of the present invention, based on the pitch polarity information, the speed control unit may determine that the first segment of the current block is similar to the next segment, and that the last segment of the current block is the first segment of the next block. If it is similar to the segment, the last segment of the current block is removed to shorten the through section of the current block.

If the first segment of the current block is similar to the next segment and the last segment is similar to the first segment of the next block, the last segment is calculated by adding the first segment to the next block. However, a natural reproduced sound can be obtained. Therefore, the last segment is removed from the current block. This reduces the segment by one and shortens the block.

According to the eleventh aspect of the present invention, based on the pitch polarity information, the speed control section determines that the first segment of the current block is similar to the last segment of the immediately preceding block, If the segment is similar to the next segment of the block, the last segment of the previous block is taken into the current block as the first segment, the old first segment of the current block is added as the second segment, and the current segment is added. Is increased by one segment.

If the first segment of the current block is similar to the last segment of the previous block, and the first segment of the next block is similar to the next segment, the last segment of the previous block is Is obtained and added to the head segment of the current block, a natural reproduced sound can be obtained. This increases the segment of the current block by one and makes the block longer.

In the twelfth aspect of the present invention, the speed control unit may determine that the first segment of the current block is similar to the last segment of the immediately preceding block based on the pitch polarity information, and the last segment of the current block. Is similar to the first segment of the next block, the current block is set to the segment of the previous block as the first segment, the old first segment of the current block is added as the second segment, and the last segment is set to the next segment. Is added as the first segment of the block.

If the current block is similar to the last segment of the previous block and the last segment of the current block is similar to the first segment of the next block, the last segment of the previous block is It adds at the first segment of the current block, and adds the last segment of the current block to the first segment of the next block. As a result, a natural reproduced sound can be obtained.

According to the thirteenth aspect of the present invention, a decoder for decoding coded audio data, a pitch information of the audio data obtained from the decoder, and a pitch within a pitch which is adjacent to the temporally succeeding pitch information. A pitch storage unit for storing pitch polarity information indicating whether the data is similar to the data of the previous pitch or similar to the data of the previous pitch; and a predetermined number of continuous segments obtained by dividing the data from the decoder into segments of a predetermined length. As one block, a divider for dividing and outputting the segments of each block in the order of input, a first memory for storing the first segment of each block output by the divider, and a The second memory for storing the second segment of the block, and the waveform of the first segment and the waveform of the next segment are approximately determined from the data of the pitch storage unit. A waveform shift amount calculator for calculating the shift amount by shifting so as to coincide with each other, and calculating a correlation function between the content of the first memory and the content of the second memory based on the output of the waveform shift amount calculator. A correlator for calculating the amount of correlation delay, a window function generator for generating a weighting factor for the data of the segment, a first multiplier for multiplying the content of the first memory by the weighting factor, A second multiplier for multiplying the content by the weighting coefficient, and an adder for adding the output of the first multiplier and the output of the second multiplier at a position where the value of the correlation function is large based on the output of the correlator Then, an output of the adder and, subsequently, a through section segment composed of the second and subsequent segments when the speed is reduced, and the third and subsequent segments when the speed is increased, input from the divider are output. With the coupler Comprising a speed control unit for controlling the divider and the combiner based on the speed ratio representing the conversion ratio of the sound speed, the.

When the encoded data is decoded,
Pitch information and pitch polarity information of audio data can be obtained.
The audio data is divided into segments, and a predetermined number of segments are made into one block, and audio speed conversion is performed in units of this block. The speed conversion is performed based on the speed ratio α obtained by dividing the converted playback time length by the playback time length before conversion, and the number of segments in one block is predetermined according to the speed ratio α.
The speed controller controls the divider to output the first segment of the block to the first memory, and the second segment to the second memory.
The data is output to the memory, and the second and subsequent segments or the third and subsequent segments are output to the coupler according to the speed reduction or the speed increase.

The waveform shift amount calculator calculates the shift amount by shifting the waveform of the first segment and the waveform of the second segment so as to substantially coincide with each other based on the pitch information. If the waveform and the phase of both segments are the same, the shift amount is zero. With the waveforms of both segments shifted so as to substantially coincide with each other, the correlator calculates the correlation function between the first segment of the first memory and the second segment of the second memory to find a position having a high correlation. The search range for finding is greatly narrowed, and the amount of calculation is greatly reduced. Thereby, the calculation of the correlation delay amount can be easily performed.

The window function generator generates a weighting coefficient for the data of the first segment and the data of the second segment according to a predetermined pattern. The weighting factor is calculated by adding both segments to create a new added segment.
It is provided so that the size of the data of the added segment is the same as that of the data of the original segment, and the addition does not make the sound unnatural.
Therefore, the sum of one weight coefficient and the other weight coefficient of the data to be added in both segments is 1.0. The weight coefficient of one segment increases gradually from the head of the segment to the rear end, and the weight coefficient of the other segment decreases gradually from the head of the segment to the rear end.

The first multiplier multiplies the data in the first memory by the weighting factor, and the second multiplier multiplies the data in the second memory by the weighting factor. The adder adds the outputs of the first and second multipliers by shifting one of them to a position having a large correlation function based on the correlation delay amount. The combiner continuously outputs the through segment input from the divider after the addition segment obtained by the addition as one block of data.

According to a fourteenth aspect of the present invention, the audio data is input, divided into segments of a predetermined length, and a predetermined number of continuous segments are formed into one block, and the segments of each block are divided and output in the order of input. A first memory for storing the first segment of each block output by the divider; and a second memory for each block output by the divider.
A second memory for storing a third segment; a correlator for calculating a correlation function between the contents of the first memory and the contents of the second memory to calculate a correlation delay amount; and generating a weight coefficient for the data of the segment. A window function generator;
A first multiplier for multiplying the content of the memory by the weight coefficient;
A second multiplier for multiplying the content of the second memory by the weighting coefficient, and a position where the value of the correlation function is large between the output of the first multiplier and the output of the second multiplier based on the output of the correlator. , And an output of the adder followed by a through segment composed of the second and subsequent segments input from the divider when the speed is low, and the third and subsequent segments when the speed is high. A combiner that outputs a section segment, and a speed control unit that controls the divider and the combiner based on a speed ratio that represents a conversion ratio of a voice speed, and the speed control unit controls the divider. To make the length of the segment variable.

The input audio data is divided into segments, and a predetermined number of segments are made into one block, and audio speed conversion is performed for each block. The speed conversion is performed based on the speed ratio α obtained by dividing the converted playback time length by the playback time length before conversion, and the number of segments in one block is predetermined according to the speed ratio α. The speed controller controls the divider to output the first segment of the block to the first memory,
The second segment is output to the second memory, and the second and subsequent segments or the third and subsequent segments are output to the coupler according to the speed reduction or the speed increase. The correlator calculates a correlation function between the first segment and the second segment, and calculates a correlation delay amount indicating a distance between highly correlated positions of both segments.

The window function generator generates a weight coefficient for each of the data of the first segment and the data of the second segment according to a predetermined pattern. The weighting factor is calculated by adding both segments to create a new added segment.
It is provided so that the size of the data of the added segment is the same as that of the data of the original segment, and the addition does not make the sound unnatural.
Therefore, the sum of one weight coefficient and the other weight coefficient of the data to be added in both segments is 1.0. The weight coefficient of one segment increases gradually from the head of the segment to the rear end, and the weight coefficient of the other segment decreases gradually from the head of the segment to the rear end.

The first multiplier multiplies the data of the first memory by the weighting factor, and the second multiplier multiplies the data of the second memory by the weighting factor. The adder adds the outputs of the first and second multipliers by shifting one of them to a position having a large correlation function based on the correlation delay amount. The combiner continuously outputs the through segment input from the divider after the addition segment obtained by the addition as one block of data.

The speed controller controls the divider to change the length of the segment. This makes it possible to appropriately reproduce data with different voices, such as a man or a woman.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of an audio speed conversion device that realizes each embodiment of the present invention. The A / D converter 31 converts an input signal (audio data) from an analog signal to a digital signal. In the case of the present embodiment, the audio encoder / decoder 32 decodes the encoded input signal.
The waveform obtained by the audio encoder / decoder 32 is written into the buffer 33, and the pitch information and the pitch polarity information obtained by the audio encoder / decoder 32 are simultaneously stored in a pitch buffer (pitch storage unit) 34. Is voiced when the pitch is clearly expressed, and unvoiced when the pitch is unclear, and voiced / unvoiced discrimination flag information, which is indicated by a flag, is written to the flag buffer (flag storage unit) 35.

In the speed control circuit 36, a speed ratio α is set in advance, and an initial segment length and a double segment length twice as long as the initial segment length are set. As described with reference to FIG. 18 showing the speed conversion, the number of segments of one block is determined according to the speed ratio α, and the first segment and the next segment are multiplied by a weighting factor and added to form one segment.
To increase the speed, the third segment and subsequent segments (hereinafter referred to as “through-segment segments”) are added to this added segment to make the block one segment less.
When the speed is to be reduced, the second and subsequent segments are added after the addition segment to make the block one segment more. An arbitrary speed ratio α can be obtained by changing the number of segments in one block.

The speed control circuit 36 receives pitch information and pitch polarity information from the pitch buffer 34 and
7 and a coupler 49 described later. The divider 37 reads out the audio data from the data buffer 33, divides it into the size of the initial segment, determines the number of segments in one block, outputs the first segment to the first memory 38, and stores the next segment in the second memory 39. When the speed is reduced, the segments after the second segment and the first segment of the next block are combined, and when the speed is increased, the segments in the block after the third segment (these segments are referred to as through-segment segments) are combined. Output to

The first correlator 41 calculates the correlation function between the data in the first memory 38 and the data in the second memory 39, and shifts the relative position so that the large positions of the correlation functions of the two segments coincide with each other. The correlation delay rk is obtained. On the other hand, the waveform shift amount (rk) calculator 42 obtains the pitch information from the pitch buffer 34, and assumes that the waveforms of the first segment and the next segment have the same pitch and differ only in the phase.
A shift amount (that is, a phase difference) shifted so as to coincide with each other is obtained. This shift amount can be easily obtained based on the above assumption.

The second correlator 43 calculates a correlation function based on the data in the first memory 38 and the second memory 39 and the shift amount from the waveform shift (rk) calculator 42, and calculates the correlation function of the two segments. The second correlation delay amount rk obtained by shifting the relative position so that the position having the larger value of. Since the shift amount is known, the data search range when the second correlator 43 calculates the correlation function is significantly smaller than that in the case of the first correlator 41, and an accurate correlation function can be obtained in a short time.

The window function generator 45 includes a first memory 38 and a second
The weight coefficients for the data in the memory 39 are output separately. This weighting factor is determined by the speed ratio α, the first correlator 41,
On the basis of the correlation function generated by the second correlator 43, a window function that gradually increases or decreases as the data goes from the head to the tail of the segment is generated. This window function is stored in the first memory 38.
And the sum of the respective weighting factors for the corresponding data in the second memory 39 is set so as to be 1.0. I try not to change.

The first multiplier 47 multiplies the data in the first memory 38 by the weight coefficient from the window function generator 45, and the second multiplier 48 calculates the weight coefficient from the second memory 39 and the window function generator 45. The result is multiplied and output to the adder 44. Changeover switch 4
0 is the first correlation delay rk of the first correlator 41 or the second correlation delay rk of the second correlator 43 according to the pitch information of the pitch buffer 34, the pitch polarity information and the flag of the flag buffer 35.
The correlation delay amount rk is output to the adder 44.

The adder 44 compares the output of the first multiplier 47 and the output of the second multiplier 48 with the first correlation delay rk or the second
Based on the correlation delay amount rk, one of the values is shifted and added to a large value of the correlation function to generate an added segment.
Output to Under the control of the speed control circuit 36, the combiner 49 sends the segment after the second segment and the first segment of the next block sent from the divider 37 after the addition segment from the adder 44 at the time of speed reduction.
The data is output as block data, and at the time of high-speed operation, the segments in the block after the third segment are output as one block. The D / A converter 50 converts the data from the combiner 49 from a digital signal to an analog signal and outputs it. This enables sound reproduction.

The adder 46 integrates the first correlation delay rk and the second correlation delay rk generated by the first correlator 41 and the second correlator 43 in the order of generation, and stores the result in the rk buffer 51. When the integrated value of the rk buffer 51 reaches the length of the +1 default segment or becomes shorter than the length of the -1 default segment, a flag is set from the delay flag generator 52, and this flag information is stored in the speed control circuit 36. Is output to

When the pitch becomes larger than the initial segment length due to the pitch information, the speed control circuit 36 switches to a double-length segment which is twice the initial segment, and doubles the length of one block. Segment length shift register 5
Reference numeral 3 indicates that the segment length has been switched, and the switch 40 switches the output of the first correlator 41 and the output of the second correlator 43 based on this switching information.

FIG. 2 is a diagram for explaining audio data and segments. As shown in (A), the audio data has one cycle of data having one large amplitude and a plurality of small amplitudes, and is often repeated. The length between large amplitudes is referred to as a pitch L. This pitch L is, for example, about 10 ms. If L = 7.5 ms is set as the default length as a statistical approximation of the female voice, one large amplitude will enter each segment. (B) shows the segment set in this way. This segment is called an initial (default) segment.

On the other hand, the pitch of a male voice or the like is long and often exceeds L = 7.5 ms. (C) shows such audio data. Since the pitch L1 is longer than the segment of (B), there may be a case where a large amplitude does not appear in one segment. Therefore, as shown in (D), a double length segment which is twice the segment length is adopted. If the number of segments in one block is fixed, the length of one block is also 2
Double.

(E) is a diagram for explaining pitch polarity information. When the data of two adjacent segments are similar, whether the data is similar to the segment on the right side (time progress direction) or similar to the segment on the left side is called pitch polarity information. The figure shows a case where the second segment and the third segment are similar. L (2) when the second segment is centered and similar to the right segment
> 0. L (3) <0 when the third segment is the center and is similar to the left segment. In addition,
If the segments are similar on both sides as shown in (A), they are represented by L (2)> 0 and L (2) <0.

FIG. 3 is a diagram for explaining the correlation delay amount rk. (A) shows the waveforms a,
b and c are shown. ts indicates the initial segment length. (B)
Indicates a state in which the waveform a is superimposed on the waveform b so that the maximum amplitude coincides. In this case, when the waveform a is shifted by rk to the right, the waveforms a and b of both segments substantially coincide. The shift amount rk at this time is rk +. (C) shows a state where the waveform c is shifted rk to the left with respect to the waveform b and overlapped so that the maximum amplitudes match. At this time, the shift amount is rk-.
These rk +, rk- are set as the correlation delay amount rk.

Next, a first embodiment will be described. FIG. 4 shows the segment variable processing when the speed ratio α = 3/4. In this processing, the correlation delay amount rk is shown in a standard form with 0 for simplicity. 3A, the audio data in the data buffer 33 is divided into segments, and pitch length information L (x) synchronized with the segment x is obtained from the pitch buffer. One block is composed of four segments, and the pitch information L (x) indicates the first segment of each block.

(B) shows a control procedure of the divider 37 when L (1) <default segment (initial segment) is established based on the data shown in (a). In this case, the length of the segment is the initial segment length, and the pointer P1 (a) shown in FIG.
8. The data is input to the second memory 39, and the subsequent third and fourth segments are output to the combiner 49 as they are.
The data in the first memory 38 and the data in the second memory 39 are weighted and added by the window function and output to the combiner 49. The output of the combiner 49 is as shown in the figure.

(C) shows the control means of the divider 37 when the pitch length L (5) is larger than the initial segment length, and when the segment length is equal to the initial segment length × 2. With the pointers P1 (b) and P2 (b) shown in (b) as the start address, the segment length is output to the first memory 38 and the second memory 39, respectively. In the subsequent data, the original segments 8 and 9 enter the third double-length segment, and the original segments 10 and 11 enter the fourth double-length segment and are output to the combiner 49. First memory 3
8. The data in the second memory 39 is weighted and added by the window function and output to the combiner 49, which outputs a data string as shown in the figure.

(D) is the same as (c), and the pitch L
Since (13) is larger than the initial segment length, a double-length segment is used. (E) shows a case where the pitch L (13) is shorter than the initial segment length and is the same as (b).
When switching from the double-length segment to the initial segment in this way, since the pointer P2 (c) is determined to be followed by the double-length segment, the segments to be added are 12 and 14, and the segment 13 is missing. Playback sound is slightly worse.

FIG. 5 shows a process in which the segment length is variable when the speed ratio α = 3/2. The case where the correlation delay amount rk is set to 0 for simplicity. (A) to (e) are almost the same as (a) to (e) in FIG. The difference is that the diagonal line of the addition segment indicating the weight of the window function is opposite to that in the case of speeding up in FIG. This is because the addition segment is followed by the second segment, but the end of the addition segment is the first
This is because, when the segment is close to the segment, the segment is well connected to the next second segment, and the reproduced sound becomes natural. By switching the segment length to two steps according to the pitch L as described above, female and male sounds can be accurately reproduced, and versatile processing can be performed.

FIG. 6 shows detailed processing of the waveform shift amount (rk) calculator 42 and the second correlator 43. (A) shows the audio data in the first segment and the second segment, and the vertical line in the segment shows the large amplitude data described in FIG.
L indicates the pitch of this amplitude. a indicates the length from the beginning of the first segment to the first large amplitude, and b indicates the length of the first large amplitude from the beginning of the second segment. When the second correlator 43 is used, it is assumed that the pitch L is substantially the same for both the first and second segments.

(B) shows a state in which the first segment and the second segment are overlapped with the beginning of each segment.
rk- indicates the amount of movement when the second segment is moved to the left (in the direction opposite to the progress of the time axis) to match the position of large amplitude, and rk + is to the right (in the direction of progress of the time axis). This shows the amount of movement when the positions of large amplitudes are matched. The diagram on the left side of (C) shows that the correlation function is calculated by the second correlator 43 in a state where the second segment is moved rk- with respect to the first segment and the waveforms are almost matched. If the pitch L is the same in both segments as assumed in (A), rk− or rk + is the correlation delay rk
However, the pitch L is often different between the two segments. However, since the difference between the pitches L is not so large, when calculating the correlation function in a state where both segments are shifted by rk-, the search range may be reduced. For example, if the initial segment length is set to 7.5 ms, 8 kHz
In the case of z sampling discrete data, there are 60 samples of default segment data. Usually, multiplication and addition are performed on data of about 40 samples in order to obtain a correlation function. For example, a correlation function is obtained. The figure on the right side of (C) shows that the second segment is rk +
The second correlator 43 is moved in a state where the waveforms are substantially matched.
Shows the calculation of the correlation function. By determining rk from both left and right calculations, an accurate correlation delay amount r
k is obtained. Thus, the amount of calculation can be significantly reduced as compared with the case where the calculation is performed by the first correlator 41.

Next, a second embodiment will be described. This embodiment relates to switching between a first correlator 41 and a second correlator 43 by a changeover switch 40. According to the voiced / unvoiced discrimination flag of the flag buffer 35, the output of the second correlator 43 is used in the case of voiced data having a clear pitch, and is added by the output of the first correlator 41 in the case of unvoiced data having an unclear pitch. The unit 44 performs addition. Further, the second correlator 43 is selected based on the pitch polarity information.

FIG. 7 shows the switching of the changeover switch when deriving the correlation delay amount rk when the speed ratio α = 3/4.
(A) shows an initial segment sequence containing the original voice.
(B) is a case where the segment is the initial segment and the flag of the flag buffer 35 is voiced. Note that L (x)> 0 or L (x) <0 in the drawing indicates the pitch polarity information described with reference to FIG. That is, if L (1)> 0, the first segment is similar to the second segment on the right, and L (1)> 0
(2) <0 indicates that the second segment is similar to the first segment on the left. In the case of (1), when L (1)> 0 and L (2) <0, the first segment and the second segment are similar to each other.
The correlation delay amount rk has sufficient accuracy. Thereby, the second correlator 43 is selected, and the others select the first correlator 41. The same applies to the case of (1), and the segment sequence is as shown in (b).

(C) is a case where the segment length is twice as long as the initial segment length, and the flag is voiced. Are L (1)> 0, L (2)> 0,
Second correlator 43 only when L (3) <0 and L (4) <0
Switch to output. In this case, the first segment is similar to the second segment, and the second segment is similar to the third segment. The third segment is similar to the second segment,
The fourth segment is similar to the third segment. This is the case when the first and second segments are similar, the third and fourth segments are similar, and the second and third segments are similar. The same applies to the case of.

FIG. 8 shows switching of the changeover switch when deriving the correlation delay amount rk when the speed ratio α = 3/2. (A) shows a segment string storing the original audio,
(B) shows a case where the flag has an initial segment length and the flag of the flag buffer 35 indicates voiced. Is L
Second correlator 4 only when (1)> 0 and L (2) <0
3 and the output of the first correlator 41 is used for the others. , Is the same. (C) shows a case where a double-length segment is used as a segment and the segment is voiced.
When L (1)> 0, L (2)> 0, L (3) <0,
The output of the second correlator 43 is selected only when L (4) <0, and the output of the first correlator 41 is selected otherwise.
The same applies to the case of. By the processing shown in FIGS. 7 and 8,
When the waveforms are weighted by the window function and then added, the reliability of the correlation delay amount rk can be improved by selecting the output of the second correlator 43 only when the waveforms are more similar. Further, by selecting the output of the second correlator 43, the amount of calculation can be significantly reduced.

Next, a third embodiment will be described. In the present embodiment, the positions of pointers P1 and P2 indicating the start addresses of the first segment and the second segment to be added are shown, and a comparison is made between the conventional example and the present invention. FIG.
Shows pointers P1 and P2 when the speed ratio α = 3/4 in comparison with the conventional method and the present invention. (A) is a process when the correlation delay amount rk> 0, and (b) is a process when rk <0. In the case of (a), the pointer P
1 is forcibly advanced to the right (in the direction of progress of the time axis) by rk, moved to the head position of the head segment of the next block, and processing of the next block is started. In the case of (b), the pointer P1 is forcibly retracted to the left (reverse direction on the time axis) by rk as compared with the conventional case, and the processing of the next block is started. It should be noted that P2 has conventionally shown the head position of the second segment of the next block, and therefore does not need to be changed. One block before the speed conversion is composed of four segments.

FIG. 10 shows the positions of the pointers P1 and P2 when the speed ratio α = 3/2, in comparison between the conventional method and the present invention. (A) is a case where the correlation delay amount rk <0,
The pointer P2 is forcibly advanced to the right by rk as compared with the related art, and the processing of the next block is started. In the case of (b), the pointer P2 is forcibly retracted to the left by rk as compared with the conventional case, and the processing of the next block is started.

The processes shown in FIGS. 9 and 10 cause P1 and P2 to be too far apart as shown in FIGS. 9 (a) and 10 (b), thereby reducing the reliability of the correlation delay rk. Can be prevented. Thereby, natural sound can be reproduced.

Next, a fourth embodiment will be described. In the present embodiment, if the correlation delay amount rk is accumulated, the speed ratio α cannot be maintained. Therefore, each time the accumulated amount becomes a fixed amount, for example, a segment length, one segment is added to the processing block at that time (in the case of rk−). , Or one segment is deleted (in the case of rk +). Thus, the speed ratio α can be maintained.

FIG. 11 shows that when the accumulated value of rk is equal to or longer than the length of one initial segment when the speed ratio α = 3/4, one segment in the through section (the segment following the addition segment) is deleted or added. Indicates that (A) is r
The standard form is shown when k = 0. (B) In the case of rk +, when the accumulated value of rk + exceeds the initial segment length ts, one initial segment is deleted from the block being processed at that time. In (c), when the integrated value of rk- exceeds one initial segment, one initial segment is added to the through section of the block being processed at that time. This gives the speed ratio α
= 3/4 can be maintained. That is, it can be in the same state as the standard form of (a) as a whole.

FIG. 12 shows processing when the accumulated amount of the absolute value of rk exceeds one initial segment length when the speed ratio α = 3/2. (A) shows the standard form when rk = 0,
(B) is the case of rk-, where the value of rk has the value of (-). The integrated value of the absolute value of rk is the initial segment length ts
Is exceeded, one segment is removed from the through section of the block currently being processed. (C) is the case of rk +, where rk has a positive value. When the integrated value of rk exceeds the initial segment length ts, 1 is added to the through section of the block being processed.
Add segments. As a result, the speed ratio α = 3 /
2 can be maintained. That is, it can be in the same state as the standard form of (a) as a whole.

Next, a fifth embodiment will be described. In the present embodiment, pitch polarity information L (x)> 0 or L (x) <0
Select the segments to be added based on
Since the segments in the block may change, the number of segments in the block is also adjusted.

FIG. 13 shows a case where the speed ratio α = 3/4.
(A) shows a standard form in which the number of segments in a through section is not increased or decreased according to pitch polarity information. In (b), the pitch polarity information used in the processing of the next block is read in advance and compared with the pitch polarity information used in the current processing block. Is L
(1)> 0 and L (5) <0 when the pitch polarity information of adjacent blocks face each other. In this case, the first segment and the second segment of the first block are similar, and the fourth segment of the first block and the first segment (segment 5) of the second block are similar. At this time, in the first block, the segments 1 and 2 are set as addition segments, and in the second block, the segment 4 of the first block is fetched, and the segment 5 and the addition segment are generated. Thus, a natural reproduced sound can be obtained by using segments similar to each other as addition segments. However, the first block is therefore shortened by one initial segment.

Is the case where L (5) <0 and L (9) <0. In this case, the second block takes in the segment 4 from the first block and generates the segment 5 and the addition segment. In addition, the segment 8 is passed to the third block. As a result, the second block takes in the segment 4 from the first block and passes the segment 8 to the third block, so that the number of segments in the block does not change. However, since the added segments are similar to each other, a natural reproduced sound can be obtained. It is the same as the case.

Is that L (13) <0 and L (17)> 0
In this case, the pitch polarity information on both sides of the fourth block repel each other. That is, the first segment 13
Is similar to segment 12 of the third block. Fifth
In the block, the first segment 17 and the next segment 18
Are similar to each other and are not relevant to the fourth block. The fourth block takes in the segment 12 of the third block, and generates the leading segment 13 and the addition segment.
As a result, the fourth block fetches one initial segment from the third block, thereby increasing the number of segments by one.

FIG. 14 shows a case where the speed ratio α = 3/2.
(A) shows a standard form in which the number of segments in a block is not increased or decreased according to pitch polarity information. (B) is the same as (b) of FIG. 13, prefetches the pitch polarity information used in the next block and compares it with the pitch polarity information used in the current processing block. If they are facing each other, the through section is shortened by one initial segment length, and if they repel each other, as in, the through section is extended by one initial segment length and shifted to the time on the standard time axis. Should not occur. By this processing, the reproduced voice can be maintained at the specified speed ratio α, and in the case of weighting addition by the window function, it is easier to take more cases where the pitch polarity information faces each other, and the correlation delay amount rk A natural sound can be reproduced by maintaining and improving the reliability of.

FIG. 15 shows a sixth embodiment. This embodiment is a simplification of the first embodiment shown in FIG. 1. The first correlator 41 in FIG. 1 is abolished and only the second correlator 43 is used, which is referred to as a correlator 43a. 35,
The switch 40, the adder 46, the rk buffer 51, the delay flag generator 52, and the segment length shift register 53 are abolished, and the functions of the remaining devices related to the abolished devices are also abolished.

FIG. 16 shows a seventh embodiment. This embodiment is an improvement of the conventional example shown in FIG. 17, in which the length of a segment is variable so that female voice data and male voice data can be appropriately processed.

[0090]

As described above, the present invention has the following effects. Using the pitch information, the pitch polarity information, and the voiced / unvoiced discrimination flag information, the correlation delay amount rk, which was conventionally derived over a long period of time using only a correlator, can be reduced by using a waveform shift amount (rk) calculator and a correlator. Can be derived in time. By controlling the divider according to the magnitude of the pitch information and switching the segment length to two stages, versatile audio reproduction processing including male and female voices can be performed. When weighting and adding by the window function to the data in the first memory and the second memory, pitch polarity information depending on the correlation before and after the waveform of the audio data is used, and only the waveforms having higher correlation are used to calculate rk from the pitch information. By calculating the second correlation delay amount rk by the calculator and the second correlator, the reliability of the correlation delay amount rk can be improved, the amount of calculation can be reduced, and the quality of reproduced sound can be improved.

The distance between P1, which indicates the head address of the data input to the first memory, and P2, which indicates the head address of the data input to the second memory, is too far depending on the magnitude of the correlation delay rk. Correlation delay rk
In order to prevent loss of reliability, the interval between P1 and P2 is kept constant to improve the reliability of the correlation delay amount rk, thereby improving the quality of reproduced sound. By accumulating the correlation delay amount rk output by the first correlator and the second correlator, and adding or deleting one segment when the length becomes equal to or more than one segment length, the reproduced voice can be maintained at the specified speed ratio α. it can. In a series of processing including weighting addition and through section control, by pre-reading the pitch polarity information used in the next processing, and controlling the length of the through section using the result,
By maintaining the specified speed ratio α and making the weighted addition process possible when pitch polarity information is opposed as much as possible, the reliability of the correlation delay amount rk can be improved, and a natural reproduced voice can be obtained.

By using only a correlator that calculates a correlation value from the amount of waveform deviation, the apparatus can be simplified. By adding a segment variable function to a conventional device, audio data can be appropriately processed.

[Brief description of the drawings]

FIG. 1 is a block diagram of an audio speed conversion device according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating audio data and segments.

FIG. 3 is a diagram for explaining a correlation delay amount rk.

FIG. 4 is an explanatory diagram of a segment variable process when the speed ratio α = 3/4.

FIG. 5 is an explanatory diagram of segment variable processing when the speed ratio α = 3/2.

FIG. 6 is an explanatory diagram of the processing of the rk calculator and the second correlator

FIG. 7 is an explanatory diagram of a changeover switch process when deriving a correlation delay amount at a speed ratio α = 3/4.

FIG. 8 is an explanatory diagram of a changeover switch process when deriving a correlation delay amount at a speed ratio α = 3/2.

FIG. 9 shows pointers P1 and P2 at a speed ratio α = 3/4.
Illustration

FIG. 10 shows pointers P1 and P at a speed ratio α = 3/2.
2 explanatory diagram

FIG. 11 is a view for explaining correction processing accompanying rk successive addition processing at a speed ratio α = 3/4.

FIG. 12 is a view for explaining correction processing accompanying rk successive addition processing at a speed ratio α = 3/2.

FIG. 13 is an explanatory diagram of through section adjustment processing based on pitch polarity information at a speed ratio α = 3/4.

FIG. 14 is an explanatory diagram of through section adjustment processing based on pitch polarity information at a speed ratio α = 3/2.

FIG. 15 is a block diagram of a sixth embodiment of the present invention.

FIG. 16 is a block diagram of a seventh embodiment of the present invention.

FIG. 17 is a block diagram of a conventional voice speed converter.

FIG. 18 is a schematic diagram of a conventional speed conversion.

[Explanation of symbols]

 Reference Signs List 31 A / D converter 32 Audio encoder / decoder 33 Data buffer 34 Pitch buffer 35 Flag buffer 36 Speed control circuit 37 Divider 38 First memory 39 Second memory 40 Changeover switch 41 First correlator 42 Waveform shift amount ( rk) calculator 43 second correlator 43a correlator 44 adder 45 window function generator 46 adder 47,48 multiplier 49 combiner 50 D / A converter 51 rk buffer 52 delay flag generator 53 segment length shift register

Claims (14)

[Claims] 1. A decoder for decoding encoded audio data, and pitch information of audio data obtained from the decoder and data in one pitch are similar to data of an adjacent temporally later pitch. A pitch storage unit for storing pitch polarity information indicating whether the data is similar to the data of the previous pitch,
A flag storage unit for storing a flag indicating whether the pitch of the audio data obtained from the decoder is clear or unknown, and dividing the data from the decoder into segments of a predetermined length and storing a predetermined number of continuous segments as 1 A divider that divides and outputs the segments of each block in the order of input, a first memory that stores the first segment of each block that is output by the divider, and a first memory that stores the first segment of each block that is output by the divider. A second memory for storing a second segment, a first correlator for calculating a correlation function between the contents of the first memory and the contents of the second memory, and calculating a first correlation delay amount; The waveform of the first segment and the waveform of the next segment are superimposed so that the waveforms of the first segment and the next segment are substantially coincident with each other, and a waveform deviation amount calculator that calculates the deviation amount; A second correlator that calculates a correlation function between the contents of one memory and the contents of the second memory based on the output of the waveform shift amount calculator to calculate a second correlation delay amount; A generated window function generator, a first multiplier for multiplying the content of the first memory by the weighting factor, a second multiplier for multiplying the content of the second memory by the weighting factor, An adder for adding the output of the first multiplier and the output of the second multiplier at a position where the value of the correlation function is large based on the output of the second correlator or the first correlator; (2) a changeover switch for switching the output from the correlator according to the flag in the flag storage unit and outputting the output; and the output from the adder and, subsequently, the second segment and subsequent segments when the speed is input from the divider. 3rd segment for high speed A speech speed converter comprising: a combiner that outputs a through section segment composed of subsequent segments; and a speed control unit that controls the divider and the combiner based on a speed ratio that represents a conversion ratio of the speed of sound. apparatus. 2. The audio speed conversion device according to claim 1, wherein the speed control unit changes the length of the segment based on the pitch information. 3. The length of the segment is set to an initial segment length based on a female voice, and when the pitch of the pitch information exceeds the initial segment length, the length of the segment is set to 2 of the initial segment. 3. The audio speed conversion device according to claim 2, wherein the length is doubled and the length of the block is doubled. 4. The audio speed conversion device according to claim 3, wherein in the block in which the length of the segment is switched, the switch outputs the output of the first correlator to the adder. . 5. The switch outputs the output of the second correlator to the adder when the pitch of the audio data in the flag storage unit is clear, and outputs the first correlation when the pitch is unclear. 2. The audio speed conversion device according to claim 1, wherein the output of the audio device is output to the adder. 6. The changeover switch outputs an output of the second correlator to the adder when data of a first segment and a next segment of the block are similar based on the pitch polarity information. The audio speed conversion device according to claim 1, wherein 7. Double the initial segment as a new segment, wherein the first new segment comprises a first initial segment and a second initial segment in data transmission order, and the next new segment comprises a third initial segment and a second initial segment in data transmission order. If the first initial segment is similar to the second initial segment, the second initial segment is similar to the third initial segment, and the fourth initial segment is similar to the third initial segment, The switch is the second
The audio speed conversion device according to claim 3, wherein an output of a correlator is output to the adder. 8. The audio speed conversion device according to claim 1, wherein a head of data input to the first memory and the second memory indicates a head address of the segment. 9. A correlation delay amount integrating means for integrating the first correlation delay amount and the second correlation delay amount, and a delay flag generator for generating a delay flag when the integrated value substantially reaches a segment length. The audio speed conversion apparatus according to claim 1, further comprising: a speed control unit that adds or deletes a segment of a through section of the block based on the delay flag. 10. The speed controller, based on the pitch polarity information, wherein the first segment of the current block is similar to the next segment, and the last segment of the current block is similar to the first segment of the next block. If you have
2. The apparatus according to claim 1, wherein a last segment of the current block is removed to shorten a through section of the current block. 11. The speed controller according to claim 1, wherein the first segment of the current block is similar to the last segment of the immediately preceding block, and the first segment of the next block is determined based on the pitch polarity information. If the next segment is similar, the last segment of the previous block is taken into the current block as the first segment, the old first segment of the current block is added as the second segment, and the through section of the current block is added. 2. The audio speed conversion device according to claim 1, wherein the number is increased by one segment. 12. The speed controller according to claim 1, wherein the first segment of the current block is similar to the last segment of the immediately preceding block, and the last segment of the current block is determined based on the pitch polarity information. If it is similar to the first segment, the current block is set to the segment of the previous block as the first segment, the old first segment of the current block is added as the second segment, and the last segment is set to the first segment of the next block. The audio speed conversion device according to claim 1, wherein the addition is performed as: 13. A decoder for decoding encoded audio data, and pitch information of audio data obtained from the decoder and data in one pitch are similar to data of an adjacent temporally later pitch. Or a pitch storage unit for storing pitch polarity information indicating whether the data is similar to the data of the previous pitch, and dividing the data from the decoder into segments of a predetermined length and dividing a predetermined number of continuous segments into one block. A first memory for storing the first segment of each block output by the divider; a second memory for storing the first segment of each block output by the divider; and a second memory of each block output by the divider. A second memory for storing a segment, and a shift of the waveform of the first segment and the waveform of the next segment based on the data of the pitch storage unit so that the waveform substantially matches the waveform of the next segment. And a waveform shift amount calculator for calculating the shift amount of the overlap, and calculating a correlation function between the content of the first memory and the content of the second memory based on the output of the waveform shift amount calculator. , A window function generator for generating a weight coefficient for the data of the segment, a first multiplier for multiplying the content of the first memory by the weight coefficient, and a weight for the content of the second memory. A second multiplier for multiplying a coefficient, an adder for adding an output of the first multiplier and an output of the second multiplier at a position where a value of a correlation function is large, based on an output of the correlator; A combiner for outputting a through section segment composed of the second and subsequent segments at the time of low-speed input and the third and subsequent segments at the time of high-speed input from the divider; Audio speed conversion ratio A speed control unit for controlling the coupler and the splitter based on a speed ratio representative, voice speed converting apparatus having a. 14. A divider that inputs audio data, divides it into segments of a predetermined length, divides a predetermined number of continuous segments into one block, and divides and outputs the segments of each block in the order of input. A first memory for storing the first segment of each block output by the divider, a second memory for storing a second segment of each block output by the divider, the contents of the first memory and the second memory And a window function generator for generating a weighting factor for the data of the segment, and a first multiplication for multiplying the content of the first memory by the weighting factor. A multiplier, a second multiplier for multiplying the content of the second memory by the weighting coefficient, and a correlation function between an output of the first multiplier and an output of the second multiplier based on an output of the correlator. An adder for adding at a position where the value of is larger, an output of the adder, and a subsequent segment input from the divider when the speed is low, and after the third segment when the speed is high. And a speed control unit that controls the splitter and the combiner based on a speed ratio representing a conversion ratio of the speed of voice, and the speed control unit includes the speed control unit. An audio speed conversion device characterized in that the length of the segment is made variable by controlling a dividing unit.