JP2005221811A - Device and method for converting speech speed - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve the problem that an optimum speech signal segment is not always selected and sufficient sound quality can not be obtained by a conventional device and method for speech speed conversion since positions of two speech signals to be weighted and added are unequivocally determined based upon a time axis conversion ratio α, a speech signal segment length Ts, and a shift time Tc with which a correlation function of last weighted addition becomes maximum. <P>SOLUTION: A device is equipped with a similarity calculating circuit which finds the similarity between the two speech signal segments, a judgement circuit which detects a value with a large similarity value among similarities that the similarity calculating circuit finds while shifting one of the two speech signal segments and delaying the segmentation time of the two speech signal segments, and a parameter storage circuit which stores a parameter when the similarity that the judgement circuit detects is high, and selects a combination of speech signal segments which have a high similarity and are suitable to weighted addition. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to an audio speed conversion apparatus and an audio speed conversion method that change only the duration length without changing the fundamental frequency of audio.

As an audio speed conversion apparatus and an audio speed conversion method that change only the duration length without changing the basic frequency (pitch) of audio, for example, audio for an MPEG audio layer 2 audio encoding system as described in Patent Document 1 is used. An audio speed conversion apparatus and an audio speed conversion method for an encoding apparatus or a PCM signal as described in Patent Document 2 or Non-Patent Document 1 are known.
Japanese Patent Application Laid-Open No. 11-194796 JP-A-4-104200 Suzuki, Misaki "Realization of high-quality voice speed conversion system by DSP", IEICE Technical Report, SP90-34 (1990)

  However, in the audio reproduction device disclosed in Patent Document 1, parameters for searching for a frame to be weighted and added are low energy, low audio quality, high continuity, and large amount of continuous masking based on the degree of change in energy. In contrast to the scale factor of the MPEG audio stream, the parameter for finding the position in phase for performing the weighted addition is a correlation function, and the processing is complicated because different parameters are used. There is a problem, and the time base conversion ratio is changed by changing the frequency of the frame with weighted addition and the frame without weighted addition based on the processing in units of frames such as MPEG audio. There is a problem that the ratio cannot be accurately changed in small increments.

  Further, in the audio speed conversion device and the audio speed conversion method disclosed in Patent Document 2 or described in Non-Patent Document 1, two audio signal segments are shifted to an optimal position using a correlation function and added. The positions of the two audio signal segments are uniquely determined by the time axis conversion ratio α, the audio signal segment length Ts, and the shift time Tc at which the correlation function in the previous weighted addition is maximized. The audio signal segment is not always selected, and the weighted addition length is shortened depending on the value of the shift time Tc at which the correlation function is maximized.

  The present invention solves such a problem, selects an audio signal segment from an optimal position, and performs weighted addition. Therefore, the audio quality is improved with less missing or duplicated audio, and the time axis conversion ratio is changed finely and accurately. Furthermore, it is an object of the present invention to provide an audio speed conversion device and an audio speed conversion method in which the sound quality is hardly deteriorated because the time length of weighted addition is constant.

  The audio speed conversion apparatus according to the present invention includes a similarity calculation circuit that obtains a similarity between an audio signal segment and a subsequent audio signal segment that allows the audio signal segment to partially overlap, and the similarity calculation circuit outputs the similarity. A determination circuit for detecting a high similarity value from the similarity, a parameter storage circuit for storing parameters relating to time information of two audio signal segments when the determination circuit detects a high similarity value, and a gradually decreasing window A window function generating circuit for outputting a function and a gradually increasing window function, and one window function output by the window function generating circuit for one audio signal segment based on a parameter stored in the parameter storage circuit. A first multiplier circuit to be multiplied and the window function for the other audio signal segment based on a parameter stored in the parameter storage circuit. A second multiplication circuit for multiplying the other window function output from the generation circuit, an addition circuit for adding the output of the first multiplication circuit and the output of the second multiplication circuit, and a desired time axis conversion ratio. And a switch circuit that switches and outputs the audio signal continuous at the front end of the output of the adder circuit and / or the audio signal continuous at the rear end of the output of the adder circuit and the output of the adder circuit. .

  The audio speed conversion apparatus according to the present invention includes a similarity calculation circuit for obtaining a similarity between two audio signal segments, and the similarity while shifting one of the two audio signal segments or delaying the cut-out time of the two audio signal segments. A determination circuit for detecting a value having a high similarity from the similarity calculated by the degree calculation circuit, and a parameter storage circuit for storing a parameter when the similarity detected by the determination circuit is high. Since an optimum set of audio signal segments can be selected, there is an effect that there are few audio omissions and audio duplications, and there is little audio quality degradation. Also, a switch for switching and outputting either the audio signal continuous at the front end of the output of the adder circuit and / or the audio signal continuous at the rear end of the output of the adder circuit and the output of the adder circuit so as to obtain a desired time axis conversion ratio Since the circuit is provided, an audio signal having an arbitrary time length can be output before and after the output of the adder circuit, so that there is an effect that the time axis conversion ratio can be finely and accurately changed. Further, a window function generation circuit that outputs a gradually increasing window function and a gradually decreasing window function, and one window that the window function generation circuit outputs for one audio signal segment based on the parameters stored in the parameter storage circuit A first multiplication circuit that multiplies the function, a second multiplication circuit that multiplies the other window function output from the window function generation circuit to the other speech signal segment based on the parameter stored in the parameter storage circuit, Since the adder circuit for adding the output of the first multiplier circuit and the output of the second multiplier circuit is provided, the audio signal having a certain time length and high similarity based on the parameter stored in the parameter storage circuit Since the set of segments is read and weighted and added, the time length of the weighted addition can be made constant in any case, and there is also an effect that the sound quality is hardly deteriorated.

  An audio speed conversion device according to the present invention includes a storage circuit in which an audio signal is recorded, a pointer control circuit that outputs an address value to the storage circuit, and an output destination of the audio signal output from the storage circuit as a first buffer. A first switch circuit selected from a memory circuit, a second buffer memory circuit, and a second switch circuit; a first buffer memory circuit that stores an audio signal segment output from the first switch circuit; A second buffer memory circuit for storing an audio signal segment that is allowed to partially overlap the contents of the first buffer memory circuit and that is output from the first switch circuit; and the contents of the first buffer memory circuit And a similarity calculation circuit for calculating the similarity between the second buffer memory circuit and the similarity output from the similarity calculation circuit. A determination circuit, a parameter storage circuit for storing parameters used by the pointer value calculation circuit to obtain an address value output by the pointer control circuit when the determination circuit detects a high similarity value, and the storage circuit A speed setting circuit for setting a time axis conversion ratio when reproducing an audio signal recorded in the recording medium, and the similarity calculation circuit obtains the similarity based on the time axis conversion ratio set in the speed setting circuit The address values of two audio signal segments to be calculated are calculated, or the address values of two audio signal segments having a high similarity and consecutive audio signals before and after that are calculated based on the parameters recorded in the parameter storage circuit. A pointer value calculation circuit that outputs to the pointer control circuit, and a window function generation that outputs a gradually increasing window function and a gradually decreasing window function. And the window function generation circuit outputs the audio signal segment output from the storage circuit based on the parameters stored in the parameter storage circuit and stored in the first buffer memory circuit. A first multiplication circuit for multiplying a window function; and a window for an audio signal segment output from the storage circuit and stored in the second buffer memory circuit based on a parameter stored in the parameter storage circuit A second multiplication circuit for multiplying the other window function output from the function generation circuit; an addition circuit for adding the output of the first multiplication circuit and the output of the second multiplication circuit; and the output of the addition circuit; A second switch circuit for selecting an output of the first switch circuit; an output buffer circuit for storing and outputting the output of the second switch circuit; When the similarity calculation circuit calculates the similarity, or when the addition circuit adds the output of the first multiplication circuit and the output of the second multiplication circuit, the first switch circuit is connected to the first buffer memory. When the output of the adder circuit is output to the output buffer circuit, the second switch circuit is inclined to the adder circuit side, and otherwise the output of the adder circuit is output to the circuit side or the second buffer memory circuit side. The speed setting circuit includes a control signal generation circuit that controls the first switch circuit and the second switch circuit so as to output audio signals continuous with the front and rear from the storage circuit to the output buffer circuit. Based on the set time axis conversion ratio, the pointer value calculation circuit shifts one of the two audio signal segments allowing some overlap within a certain range or 2 The start point addresses of two audio signal segments in various combinations that delay the start point of the audio signal segment are calculated, the similarity calculation circuit obtains the similarity between the two audio signal segments, and the determination circuit A value having a high similarity is detected from the similarity between two audio signal segments in various combinations, and the parameter storage circuit stores the parameters of the two audio signal segments having a high similarity detected by the determination circuit. From among the various combinations of audio signal segments within the range, it is possible to select the optimal audio signal segment set for weighted addition with a high degree of similarity. There is an effect that there is little.

  The pointer value calculation circuit calculates an address based on the time axis conversion ratio set in the speed setting circuit and the parameter stored in the parameter storage circuit, and the first switch circuit and the second switch circuit are The output of the adder circuit and the output of the audio signal from the memory circuit based on the address calculated by the pointer value calculation circuit are switched and output to the output buffer circuit. As a result, the audio signal continuous before and after the output of the adder circuit is output. Therefore, a continuous and smooth audio signal can be output, and an audio signal having a time length that satisfies a desired time axis conversion ratio can be output before and after the output of the adder circuit. There is an effect that it can be finely set and can be accurately changed to a desired time axis conversion ratio.

  Further, based on the parameters stored in the parameter storage circuit, the pointer value calculation circuit calculates an address, and the voice having a certain time length with high similarity from the storage circuit to the first buffer memory circuit and the second buffer memory circuit. A set of signal segments is read, a window function generating circuit outputs a gradually increasing window function and a gradually decreasing window function, and a first multiplying circuit generates a window function for the audio signal segment output from the first buffer memory circuit. The second multiplier circuit multiplies the audio signal segment output from the second buffer memory circuit by the other window function output from the window function generator circuit, and the adder circuit As a result of adding the output of the first multiplier circuit and the output of the second multiplier circuit so as to overlap, the time length of the output of the adder circuit can be made constant in any case, and the sound quality is lowered. The effect of hard to be certain.

  As an evaluation measure in the similarity calculation circuit applicable to the present invention, for example, the small square error, the correlation function, or the shift times of two audio signal segments when the similarity is high are the same for a certain time or more. Is applicable.

The voice speed conversion method of the present invention
Reading the time axis conversion ratio α,
Setting a first pointer and a second pointer after Ts of the first pointer;
Calculating a maximum delay time Td_max based on the time axis conversion ratio α, the audio signal segment length Ts, and the initial value Tc_min of the shift time;
Initializing the similarity,
The shift time Tc is changed from the initial value Tc_min within the range of the delay time Td, the audio signal segment length Ts and the time axis conversion ratio α, and the delay time Td is changed within the range of 0 to the maximum delay time Td_max. The start points of the audio signal segment X1 and the audio signal segment X2 are obtained by using the pointer 1, the second pointer, the shift time Tc and the delay time Td as parameters, and the audio signal segment X1 and the audio signal having the time length Ts from the respective start points. Inputting a segment X2, calculating a similarity between X1 and X2, and searching for a delay time Td_opt and a shift time Tc_opt when the similarity is high;
Inputting a voice signal having a time length Td_opt from the first pointer or the second pointer as a starting point and outputting it as it is;
The audio signal segment X1 (1 to Ts) and the audio signal segment X2 (1 to Ts) having high similarity are input using the first pointer, the second pointer, Tc_opt, and Td_opt as parameters, and the window function W1 (1 ~ Ts) and the gradually decreasing window function W2 (1 ~ Ts), the audio signal segment X1 and the audio signal segment X2 are weighted and output, and
Based on the time axis conversion ratio α, the audio signal segment length Ts, and the shift time Tc_opt having a high similarity, the time length Tt for outputting the audio signal as it is is calculated, and the first pointer or the second pointer and Tc_opt, Td_opt, and Ts are calculated. Obtaining a starting point on the basis of a voice signal having a time length (Tt−Td_opt) and outputting it as it is;
A step of setting the first pointer and the second pointer after Ts of the first pointer based on the audio signal segment length Ts, the time length Tt for outputting the audio signal as it is, and the shift time Tc_opt when the similarity is high When,
If it is not completed, a step of returning to the step of initializing the similarity is provided, so that the shift time Tc ranges from the initial value Tc_min to the maximum shift time Tc_max determined by the delay time Td, the audio signal segment length Ts, and the time axis conversion ratio α. The delay time Td is changed in the range from 0 to the maximum delay time Td_max, and the start points of the audio signal segment X1 and the audio signal segment X2 are set using the first pointer, the second pointer, the shift time Tc, and the delay time Td as parameters. And input the audio signal segment X1 and the audio signal segment X2 of time length Ts from each starting point, calculate the similarity between X1 and X2, and search for the delay time Td_opt and the shift time Tc_opt when the similarity is high Step within a certain range and shift time Tc. While changing the ray time Td, the degree of similarity between the two audio signal segments X1 and X2 that allow some overlap is obtained, and the shift time Tc and the delay time Td when a high value of the similarity is detected are set as Tc_opt and Td_opt, respectively. As a result of storing, it is possible to select a voice signal segment pair that has a high degree of similarity and is optimal for weighted addition from among a variety of voice signal segment combinations within a certain range. There is an effect that there is little sound quality degradation.

A step of inputting a voice signal having a time length Td_opt using the first pointer or the second pointer as a starting point and outputting the voice signal as it is;
The audio signal segment X1 (1 to Ts) and the audio signal segment X2 (1 to Ts) having high similarity are input using the first pointer, the second pointer, Tc_opt, and Td_opt as parameters, and the window function W1 (1 ~ Ts) and the gradually decreasing window function W2 (1 ~ Ts), the audio signal segment X1 and the audio signal segment X2 are weighted and output, and
Based on the time axis conversion ratio α, the audio signal segment length Ts, and the shift time Tc_opt having a high similarity, the time length Tt for outputting the audio signal as it is is calculated, and the first pointer or the second pointer and Tc_opt, Td_opt, and Ts are calculated. A start point is obtained based on the step, a voice signal having a time length (Tt−Td_opt) is input and output as it is;
Therefore, the audio signal having the time length Td_opt is output continuously with the front end of the weighted signal, the audio signal having the weighted time length Ts is output, and the time length ( (Tt−Td_opt) is input and output as it is. As a result, a continuous audio signal is output before or after the weighted audio signal, so that a continuous and smooth audio signal can be output. Since a sound signal having a time length that gives a desired time axis conversion ratio can be output before and after the weighted and added sound signal, the time axis conversion ratio α can be set finely and accurately changed to the desired time axis conversion ratio. There is also an effect that can be done.

  Then, the audio signal segment X1 (1 to Ts) and the audio signal segment X2 (1 to Ts) having high similarity are input using the first pointer, the second pointer, Tc_opt, and Td_opt as parameters, and the window function W1 is gradually increased. (1 to Ts) and a step function of gradually decreasing the window function W2 (1 to Ts), and the step of weighting and adding the audio signal segment X1 and the audio signal segment X2 are output. Are input using a window function W1 with an increasing segment length Ts and a window function W2 with a decreasing segment length Ts so that the audio signal segment X1 and the audio signal segment X2 overlap. As a result, in any case, the time length of the weighted and added audio signal can be a constant segment length Ts, and the sound quality is low. There is also an effect that is hard.

A step of calculating a maximum delay time Td_max based on the time axis conversion ratio α, the audio signal segment length Ts, and the initial value Tc_min of the shift time;
The shift time Tc is changed from the initial value Tc_min to the maximum shift time Tc_max, the delay time Td is changed from 0 to the maximum delay time Td_max, and the first pointer, the second pointer, the shift time Tc, and the delay time Td are changed. The starting points of the audio signal segment X1 and the audio signal segment X2 are obtained as parameters, the audio signal segment X1 and the audio signal segment X2 having the time length Ts are input from the respective starting points, the similarity between X1 and X2 is calculated, and the similarity Searching for a delay time Td_opt and a shift time Tc_opt when the degree is high;
Therefore, when searching for Tc_opt and Td_opt when the similarity between the audio signal segment X1 and the audio signal segment X2 is high, the shift time Tc is limited to the range from the initial value Tc_min to the maximum shift time Tc_max, and the delay As a result of limiting the time Td to a range from 0 to the maximum delay time Td_max, there is also an effect that a sound signal having a time axis conversion ratio α can be stably output.
As an evaluation scale for obtaining the similarity applicable to the present invention, for example, two audio signal segments whose similarity is increased even when the square error is small, the correlation function is large, or the delay time Td is changed by a certain time or more. It can be applied that the shift times Tc_opt are the same or change little.

  Next, an embodiment of an audio speed conversion device and an audio speed conversion method of the present invention will be described with reference to the drawings.

(Embodiment 1)
FIG. 1 is a block diagram according to an embodiment of an audio speed conversion apparatus of the present invention, in which 101 is a storage circuit, 102 is a first switch circuit, 103 is a first buffer memory circuit, and 104 is a second buffer memory. Circuit, 105 similarity calculation circuit, 106 determination circuit, 107 window function generation circuit, 108 third switch circuit, 109 fourth switch circuit, 110 first multiplication circuit, and 111 second circuit Multiplication circuit, 112 addition circuit, 113 second switch circuit, 114 output buffer circuit, 115 speed setting circuit, 116 parameter storage circuit, 117 pointer value calculation circuit, 118 pointer control circuit, 119 control It is a signal generation circuit.

  An audio signal is recorded in the memory circuit 101, and an audio signal having a desired start point and time length is output based on the address value and time length output by the pointer control circuit 118.

  The first switch circuit 102 selects the output destination of the audio signal output from the storage circuit 101 from the first buffer memory circuit 103, the second buffer memory circuit 104, and the second switch circuit 113.

  The first buffer memory circuit 103 stores the audio signal segment output from the first switch circuit 102, and the second buffer memory circuit 104 is output from the first switch circuit 102 and is stored in the first buffer memory circuit 103. Accumulate subsequent audio signal segments with some overlap in content.

  The similarity calculation circuit 105 obtains the similarity between the contents of the first buffer memory circuit 103 and the contents of the second buffer memory circuit 104, and the determination circuit 106 determines the similarity based on the similarity output from the similarity calculation circuit 105. The parameter storage circuit 116 uses the parameter (shift) used by the pointer value calculation circuit 117 to obtain the address value output by the pointer control circuit 118 when the determination circuit 106 detects a high similarity value. Time: Tc, delay time: Td) are stored as optimum shift time: Tc_opt and optimum delay time: Td_opt. At this time, the third switch circuit 108 and the fourth switch circuit 109 are open, and the content of the first buffer memory circuit 103 and the content of the second buffer memory circuit 104 are the same as those of the first multiplier circuit 110 and the second 2 is not output to the multiplication circuit 111.

  The speed setting circuit 115 sets a time axis conversion ratio α when reproducing an audio signal recorded in the storage circuit 101.

  The pointer value calculation circuit 117 calculates the address values of two audio signal segments from which the similarity calculation circuit 105 should obtain the similarity based on the time axis conversion ratio α set in the speed setting circuit 115, or a parameter Based on the parameters (optimum shift time: Tc_opt, optimum delay time: Td_opt) recorded in the memory circuit 116, two speech signal segments having high similarity, and address values and time lengths of speech signals consecutive before and after the speech signal segments are obtained. Calculate and output to the pointer control circuit 118.

  The window function generation circuit 107 outputs a gradually increasing window function and a gradually decreasing window function, and the first multiplication circuit 110 is output from the storage circuit 101 based on the parameters stored in the parameter storage circuit 116, and the first multiplication circuit 110 is output. The audio signal segment stored in the buffer memory circuit 103 is multiplied by one window function output from the window function generation circuit 107, and the second multiplication circuit 111 is based on the parameters stored in the parameter storage circuit 116. The audio signal segment output from the storage circuit 101 and stored in the second buffer memory circuit 104 is multiplied by the other window function output from the window function generation circuit 107, and the addition circuit 112 is connected to the first multiplication circuit 110. The output and the output of the second multiplication circuit 111 are added. At this time, the third switch circuit 108 and the fourth switch circuit 109 are closed, and the contents of the first buffer memory circuit 103 and the contents of the second buffer memory circuit 104 are the same as those of the first multiplier circuit 110 and This is output to the second multiplication circuit 111.

  The second switch circuit 113 selects the output of the adder circuit 112 and the output of the first switch circuit 102, and the output buffer circuit 114 stores and outputs the output of the second switch circuit 113.

  When the similarity calculation circuit 105 calculates the similarity, the control signal generation circuit 119 moves the first switch circuit 102 to the first buffer memory circuit 103 side or the second buffer memory circuit 104 side, and The switch circuit 108 and the fourth switch circuit 109 are opened, and when the adder circuit 112 adds the output of the first multiplier circuit 110 and the output of the second multiplier circuit 111, the first switch circuit 102 is changed to the first switch circuit 102. The third switch circuit 108 and the fourth switch circuit 109 are closed while the buffer memory circuit 103 side or the second buffer memory circuit 104 side is closed, the second switch circuit 113 is brought down to the adder circuit 112 side, and the memory circuit 101 When the audio signal output from the first switch circuit 102 is output to the output buffer 114 as it is, the first switch circuit 102 is connected to the second switch circuit. Fold 113 side is controlled to defeat the second switching circuit 113 to the first switch circuit 102 side.

  FIG. 2 is a block diagram of the similarity calculation circuit 105 when the similarity evaluation function is a square error according to an embodiment of the audio speed conversion apparatus of the present invention, where 201 is a first shift register memory circuit, 202 is The second shift register memory circuit, 203_1 to 203_Ts are subtraction circuits, 204_1 to 204_Ts are multiplication circuits, and 205 is an addition circuit.

  Audio signal segments stored in the first buffer memory circuit 103 in FIG. 1 are sequentially input to the first shift register memory circuit 201, and the second buffer in FIG. 1 is input to the second shift register memory circuit 202. The audio signal segments stored in the memory circuit 104 are sequentially input. The subtraction circuits 203_1 to 203_Ts are connected to the audio signal segment X2 (1 to Ts) stored in the second shift register memory circuit 202 from the audio signal segment X1 (1 to Ts) stored in the first shift register memory circuit 201. Ts) is subtracted, the multiplication circuits 204_1 to 204_Ts square the outputs of the subtraction circuits 203_1 to 203_Ts, and the addition circuit 205 calculates the sum of the outputs of the multiplication circuits 204_1 to 204_Ts, and outputs the result as a square error. Equation 1 shows the square error calculation performed by the similarity calculation circuit 105. However, in Equation 1, for simplicity, the unit time and the sampling period are expressed as being equal.

  FIG. 3 is a block diagram of the similarity calculation circuit 105 when the similarity evaluation function is a correlation function according to an embodiment of the audio speed conversion apparatus of the present invention. 301 is a first shift register memory circuit, 302 is A second shift register memory circuit, 303_1 to 303_Ts are multiplication circuits, and 304 is an addition circuit.

  Audio signal segments stored in the first buffer memory circuit 103 in FIG. 1 are sequentially input to the first shift register memory circuit 301, and the second shift register memory circuit 302 has the second shift register memory circuit 301 in FIG. The audio signal segments stored in the buffer memory circuit 104 are sequentially input. The multiplication circuits 303_1 to 303_Ts include the audio signal segment X1 (1 to Ts) stored in the first shift register memory circuit 301 and the audio signal segment X2 (1 to 2) stored in the second shift register memory circuit 202. Ts) and the adder circuit 304 calculates the sum of the outputs of the multiplier circuits 303_1 to 303_Ts, and outputs the result as a correlation function. Equation 2 shows the calculation of the correlation function performed by the similarity calculation circuit 105. However, in Equation 2, for simplicity, the unit time and the sampling period are expressed as being equal.

  FIG. 4 is a block diagram of the determination circuit 106 according to an embodiment of the audio speed conversion apparatus of the present invention. 401 is a similarity memory circuit, 402 is a comparison circuit, and 403 is a maximum / minimum memory circuit.

  The similarity memory circuit 401 receives and stores the similarity output from the similarity calculation circuit 105 in FIG. The comparison circuit 402 compares the current similarity output from the similarity memory circuit 401 with the maximum value or minimum value of past similarities output from the maximum / minimum memory circuit 403, and outputs the output from the similarity memory circuit 401. Is larger than the maximum value output from the maximum / minimum memory circuit 403 or smaller than the minimum value, the output of the similarity memory circuit 401 is stored again in the maximum / minimum memory circuit 403, and the parameter storage in FIG. The circuit 116 is instructed to store the current parameters as candidates for the optimum shift time: Tc_opt and the optimum delay time: Td_opt. The comparison circuit 402 searches for the minimum value when the evaluation function is a square error, and the comparison circuit 402 searches for the maximum value when the evaluation function is a correlation function.

  FIG. 5 is a processing diagram in the case of time axis expansion (time axis conversion ratio α = 5/4) according to an embodiment of the audio speed conversion apparatus of the present invention.

  The audio signal 501 recorded in the storage circuit 101 is converted into the first buffer memory circuit 103 and the first buffer memory circuit 103 with the first pointer 502_i and the second pointer 503_i output from the pointer control circuit 118 in the i-th processing unit as a reference. The second buffer memory circuit 104 reads the audio signal segment X1 (1 to Ts) and the audio signal segment X2 (1 to Ts). The audio signal segment X2 when the delay time Td is 0 is a section indicated by 507, and the audio signal segment X1 is a period indicated by 504_0 preceding the audio signal segment X2 indicated by 507 by the audio signal segment length Ts. Is the standard. Using the section indicated by 504_0 as a reference, an audio signal segment X1 that is shifted by one sample from the section indicated by 504_min that precedes by Tc_min to the section indicated by 504_max by Tc_max, and an audio signal segment X2 The similarity calculation circuit 105 obtains the similarity to. Here, Tc_min is a predetermined constant and is shorter than the audio signal segment length Ts. Tc_max is obtained by the pointer value calculation circuit 117 according to Equation 3, so that a wide range of similarities can be searched and an accurate time axis conversion ratio α can be realized.

  Then, while increasing the delay time Td, the shift time Tc is changed in the range from Tc_min to Tc_max, and the similarity calculation circuit 105 and the determination circuit 106 calculate the similarity and search for a value with a high similarity. The maximum delay time Td_max is obtained by the pointer value calculation circuit 117 according to Equation 4, so that a wide range of similarities can be searched and an accurate time-axis conversion ratio α can be realized.

The delay time Td may be increased from 0 to Td_max every sample, but there is no problem in sound quality even if it is performed every few samples in order to reduce the amount of calculation.

  The delay time Td becomes Td_max, and the audio signal segment X2 becomes an interval indicated by 509. The audio signal segment X1 is shifted from the interval indicated by 506_min to the interval indicated by 506_max with reference to the interval indicated by 506_0. Sometimes the similarity search ends. When the similarity evaluation function is a square error, the determination circuit 106 detects the minimum value of the square error output from the similarity calculation circuit 105, and when the similarity evaluation function is a correlation function, the determination circuit 106 106 detects the maximum value of the correlation function output from the similarity calculation circuit 105. The parameter storage circuit 116 stores a delay time Td and a shift time Tc when the determination circuit 106 detects the highest similarity, and uses them as an optimum delay time Td_opt and an optimum shift time Tc_opt. Based on the optimum delay time Td_opt stored in the parameter storage circuit 116, the pointer control circuit 118 reads the audio signal X0 (516) having the time length Td_opt from the second pointer 503_i as a starting point from the storage circuit 101, and outputs the output buffer. Output to the circuit 114. Next, based on the optimum delay time Td_opt and the optimum shift time Tc_opt stored in the parameter storage circuit 116, the pointer control circuit 118 indicates the audio signal segment X2 (511) in the section indicated by 508 and indicated by 505_opt. The audio signal segment X1 (510) in the section is read from the storage circuit 101 and output to the second buffer memory circuit 104 and the first buffer memory circuit 103. The window function generation circuit 107 outputs a gradually increasing window function 512 and a gradually decreasing window function 513, and the first multiplication circuit 110 applies the audio signal segment X 1 (510) stored in the first buffer memory circuit 103. On the other hand, X1 ′ is output by multiplying the gradually increasing window function 512 output from the window function generating circuit 107, and the second multiplier circuit 111 outputs the audio signal segment X2 (511) stored in the second buffer memory circuit 104. Is multiplied by a gradually decreasing window function 513 output from the window function generation circuit 107 to output X2 ′, and the adder circuit 112 outputs the output X1 ′ of the first multiplier circuit 110 and the output X2 ′ of the second multiplier circuit 111. Is output to the output buffer circuit 114. Based on the optimum delay time Td_opt and the optimum shift time Tc_opt stored in the parameter storage circuit 116, the pointer control circuit 118 starts from the sample following the audio signal segment X1, and the time length (Tt−Td_opt) is reached. The audio signal X3 (517) is read from the storage circuit 101 and output to the output buffer circuit 114. Here, the time length Tt for outputting the input audio signal as it is is obtained by the pointer value calculation circuit 117 according to Equation 5, so that an accurate time axis conversion ratio α can be realized.

Thus, the i-th processing unit is completed, and for the i + 1-th processing unit, the pointer value calculation circuit 117 obtains the second pointer 503_i + 1 and the first pointer 502_i + 1 by Equations 6 and 7, and performs pointer control. Output to the circuit 118.

  FIG. 6 is a processing diagram in the case of time axis compression (time axis conversion ratio α = 4/5) according to an embodiment of the audio speed conversion apparatus of the present invention.

  The audio signal 601 recorded in the storage circuit 101 is compared with the first buffer memory circuit 103 and the first buffer memory circuit 103 based on the first pointer 602_i and the second pointer 603_i output from the pointer control circuit 118 in the i-th processing unit. The audio signal segment X1 (1 to Ts) and the audio signal segment X2 (1 to Ts) are read into the second buffer memory circuit 104. The audio signal segment X1 when the delay time Td is 0 is a section indicated by 604, and the audio signal segment X2 is indicated by 607_0 following the audio signal segment length Ts from the audio signal segment X1 indicated by 604. The section is the standard. The interval indicated by 604_0 is used as a reference, the audio signal segment X2 that is shifted by one sample from the interval indicated by 607_min following the Tc_min and preceded by Tc_max, and the interval indicated by 607_max, and the audio signal segment X1 The similarity calculation circuit 105 obtains the similarity to. Here, Tc_min is a predetermined constant and is shorter than the audio signal segment length Ts. Tc_max is obtained by the pointer value calculation circuit 117 according to Equation 8, so that a wide range of similarities can be searched and an accurate time axis conversion ratio α can be realized.

  Then, while increasing the delay time Td, the shift time Tc is changed in the range from Tc_min to Tc_max, and the similarity calculation circuit 105 and the determination circuit 106 calculate the similarity and search for a value with a high similarity. The maximum delay time Td_max is obtained by the pointer value calculation circuit 117 according to Equation 9, so that a wide range of similarities can be searched and an accurate time axis conversion ratio α can be realized.

  The delay time Td may be increased from 0 to Td_max every sample, but there is no problem in sound quality even if it is performed every few samples in order to reduce the amount of calculation.

  The delay time Td becomes Td_max, and the audio signal segment X1 becomes a section indicated by 606, and the audio signal segment X2 is shifted from the section indicated by 609_min to the section indicated by 609_max with reference to the section indicated by 609_0. Sometimes the similarity search ends. When the similarity evaluation function is a square error, the determination circuit 106 detects the minimum value of the square error output from the similarity calculation circuit 105, and when the similarity evaluation function is a correlation function, the determination circuit 106 106 detects the maximum value of the correlation function output from the similarity calculation circuit 105. The parameter storage circuit 116 stores a delay time Td and a shift time Tc when the determination circuit 106 detects the highest similarity, and uses them as an optimum delay time Td_opt and an optimum shift time Tc_opt. Based on the optimum delay time Td_opt stored in the parameter storage circuit 116, the pointer control circuit 118 reads out the audio signal X0 (616) of the time length Td_opt from the storage circuit 101 starting from the first pointer 602_i, and outputs the output buffer. Output to the circuit 114. Next, based on the optimal delay time Td_opt and the optimal shift time Tc_opt stored in the parameter storage circuit 116, the pointer control circuit 118 is indicated by the audio signal segment X1 (610) in the section indicated by 605 and 608_opt. The audio signal segment X2 (611) in the section is read from the storage circuit 101 and output to the first buffer memory circuit 103 and the second buffer memory circuit 104. The window function generation circuit 107 outputs a gradually decreasing window function 612 and a gradually increasing window function 613, and the first multiplication circuit 110 outputs the audio signal segment X 1 (610) stored in the first buffer memory circuit 103. On the other hand, X1 ′ is output by multiplying the gradually decreasing window function 612 output from the window function generating circuit 107, and the second multiplier circuit 111 outputs the audio signal segment X2 (611) stored in the second buffer memory circuit 104. Is multiplied by a gradually increasing window function 613 output from the window function generation circuit 107 to output X2 ′, and the adder circuit 112 outputs the output X1 ′ of the first multiplier circuit 110 and the output X2 ′ of the second multiplier circuit 111. Is output to the output buffer circuit 114. Then, based on the optimum delay time Td_opt and the optimum shift time Tc_opt stored in the parameter storage circuit 116, the pointer control circuit 118 starts from a sample following the audio signal segment X2, and has a time length (Tt−Td_opt). The audio signal X3 (617) is read from the storage circuit 101 and output to the output buffer circuit 114. Here, the time length Tt for outputting the input audio signal as it is is obtained by the pointer value calculation circuit 117 using Equation 10, so that an accurate time axis conversion ratio α can be realized.

Thus, the i-th processing unit is completed, and for the i + 1-th processing unit, the pointer value calculation circuit 117 obtains the first pointer 602 — i + 1 and the second pointer 603 — i + 1 using Equations 11 and 12, and performs pointer control. Output to the circuit 118.

  In this way, based on the time axis conversion ratio α set in the speed setting circuit 115, the pointer value calculation circuit 117 shifts one of the two audio signal segments that allow some overlap within a certain range. The start point addresses of two audio signal segments in various combinations, such as delaying the start points of two audio signal segments, and the similarity calculation circuit 105 obtains the similarity between the two audio signal segments, The determination circuit 106 detects a high similarity value from the similarities between two audio signal segments of various combinations of the delay time Td and the shift time Tc, and the parameter storage circuit 116 detects the two segments detected by the determination circuit 106. As a result of storing parameters (Tc_opt, Td_opt) when the similarity between them is high, various audio signal segments within a certain range are stored. From the combinations of cement, since the optimum set of audio signal segments for summing the similarity is high weighting can be chosen, less voice missing or audio overlap, there is the effect that less sound quality.

  The pointer value calculation circuit 117 calculates an address based on the time axis conversion α set in the speed setting circuit 115 and the parameters (Tc_opt, Td_opt) stored in the parameter storage circuit 116, and the first switch circuit 102 and the second switch circuit 113 switch the output of the adder circuit 112 and the output of the audio signal from the storage circuit 101 based on the address calculated by the pointer value calculation circuit 117 and output to the output buffer circuit 114 as a result. Since continuous audio signals (X0, X3) are output before and after the output of the circuit 112, a continuous and smooth audio signal can be output, and a desired time axis can be output before and after the output of the adder circuit 112. Since the audio signal (X0, X3) with a time length that can be converted to α can be output, the time axis conversion ratio α can be set finely. Come, and there is also an effect that can be changed accurately in α desired time axis conversion ratio.

  Further, the pointer value calculation circuit 117 calculates an address based on the parameters (Tc_opt, Td_opt) stored in the parameter storage circuit 116, and the first buffer memory circuit 103 and the second buffer memory circuit 104 from the storage circuit 101. , A set (X1, X2) of a speech signal segment having a certain time length Tc having a high degree of similarity is read out, and the window function generation circuit 107 outputs a gradually increasing window function and a gradually decreasing window function, and the first multiplication circuit 110 The audio signal segment X1 output from the first buffer memory circuit 103 is multiplied by one window function output from the window function generation circuit 107, and the second multiplier circuit 111 outputs the audio output from the second buffer memory circuit 104. The signal segment X2 is multiplied by the other window function output from the window function generation circuit 107, and the adder circuit 112 As a result of adding the output X1 ′ of the multiplier circuit 110 and the output X2 ′ of the second multiplier circuit 111 so as to overlap each other, the time length of the output of the adder circuit 112 can be made a constant segment length Ts in any case, There is also an effect that the sound quality is hardly deteriorated.

  As the evaluation scale in the similarity calculation circuit 105 of the present embodiment, the small square error shown in FIG. 2 and the correlation function shown in FIG. 3 are used, but the delay time Td is changed by a certain time or more. However, it is also possible to use an evaluation measure that the shift times Tc of two audio signal segments having high similarity are the same or change little. In this case, the continuity of the audio signal is taken into consideration, and improvement in sound quality can be expected.

  Note that, as the evaluation measure in the similarity calculation circuit 105 of the present embodiment, the size of the unnormalized square error shown in FIG. 2 and the size of the unnormalized correlation function shown in FIG. 3 are used. It is also possible to use a normalized squared error or a normalized correlation function. In this case, the amount of computation increases, but the evaluation scale does not depend on the amplitude of the audio signal, so that the similarity can be obtained without being influenced by the amplitude of the audio signal, and improvement in sound quality can be expected. .

In the first buffer memory circuit 103 and the second buffer memory circuit 104 of the present embodiment, the audio signal is read from the storage circuit 101 in units of the audio signal segment length Ts, but may be read in units of larger processing units. good. For example, in the case of the time axis extension shown in FIG. 5, from the start point of 504_min to the end point of 509, in the case of time axis compression shown in FIG. 6, from the start point of 604 to the end point of 609_min, When the similarity between two audio signal segments is obtained while changing the delay time Td and the shift time Tc by reading them into the first buffer memory circuit 103 and the second buffer memory circuit 104, and when the similarity is high When the two audio signal segments in the shift time Tc_opt and the delay time Td_opt are weighted and added, the storage circuit 101 can be prevented from being accessed. In this case, since the number of transfers from the memory circuit 101 to the first buffer memory circuit 103 and the second buffer memory circuit 104 can be reduced, the processing time can be shortened (Embodiment 2).
FIG. 7 is a flowchart in the case of the time base extension (α ≧ 1) of the voice speed conversion method of the present invention, a step of reading the time base conversion ratio α of 702, and a step of initializing the first pointer of 703. , The step of setting a value after the audio signal segment length Ts relative to the first pointer to the second pointer of 704, and the time axis conversion ratio α, the audio signal segment length Ts, and the shift time based on Equation 4 of 705 A step of calculating a maximum delay time Td_max from an initial value Tc_min, a step of initializing a minimum square error R_min of 706 to an initial value N, a step of initializing a delay time Td of 707 to an initial value of 0, and a shift of 708 The step of setting the initial value Tc_min of the shift time to the time Tc and the delay time Td based on Equation 3 of 709 A step of calculating a maximum shift time Tc_max; a step of inputting Ts audio signal segments X1 (1 to Ts) starting from 710 (first pointer + Tc + Td); and 711 (second pointer + Td). The step of inputting Ts audio signal segments X2 (1 to Ts) starting from, and the square of the audio signal segment X1 and the audio signal segment X2 at the time of the shift time Tc and the delay time Td based on Equation 712 The step of calculating the error R (Tc, Td) is compared with the least square error R_min and the square error R (Tc, Td) of 713, and if R_min is larger than the square error R (Tc, Td), the process proceeds to step 714. Go to step 717, otherwise go to step 717 and change the square error R (Tc, Td) of 714 to the new minimum A step of updating the multiplication error R_min, a step of updating the shift time Tc of 715 as the optimal shift time Tc_opt, a step of updating the delay time Td of 716 as the optimal delay time Td_opt, and a shift time Tc of 717 of only one sample. The step of increasing is compared with the shift time Tc of 718 and the maximum shift time Tc_max. When the shift time Tc is not greater than the maximum shift time Tc_max, the process returns to step 710, and the shift time Tc is greater than the maximum shift time Tc_max. If larger, the process proceeds to step 719 to change the shift time Tc within the range from the initial value Tc_min to the maximum shift time Tc_max, to increase the delay time Td of 719 by ΔTd samples, and to 720 The ray time Td and the maximum delay time Td_max are compared. If the delay time Td is not greater than the maximum delay time Td_max, the process returns to step 708. If the delay time Td is greater than the maximum delay time Td_max, the process returns to step 721. By proceeding, the step of changing the delay time Td in the range from 0 to the maximum delay time Td_max, the step of inputting Td_opt audio signal samples starting from the second pointer 721, and outputting them as they are, A step of inputting Ts audio signal segments X1 (1 to Ts) starting from 1 pointer + Tc_opt + Td_opt), and Ts audio signal segments X2 (1 to 1) starting from 723 (second pointer + Td_opt) Step to enter Ts) And a window function W2 (i) that gradually decreases with respect to the audio signal segment X2 (i), which is multiplied by the window function W1 (i) that gradually increases with respect to the audio signal segment X1 (i) based on Equation 13 of 724. Calculating and outputting a signal Y (i) obtained by adding the product of i = 1 to Ts;

A step of calculating a time length Tt for outputting the input voice signal as it is based on Equation 5 of 725, and (Tt−Td_opt) number of voice signals as input starting from (first pointer + Tc_opt + Td_opt + Ts) of 726 A step of outputting; a step of setting a second pointer in the next processing unit based on Formula 6 of 727; a step of setting a first pointer in the next processing unit based on Formula 7 of 728; If not, the process returns to step 706 to repeat the process. If completed, the process ends at step 730.

  However, in this flowchart, for simplicity, the unit time and the sampling period are expressed as being equal.

  In step 708, step 717 and step 718, the shift time Tc is changed in the range from the initial value Tc_min to the maximum shift time Tc_max. In steps 707, 719 and 720, the delay time Td is changed in the range from 0 to the maximum delay time Td_max. In step 710, the start point of the audio signal segment X1 is obtained and the audio signal segment X1 having the time length Ts is input. In step 711, the start point of the audio signal segment X2 is obtained and the audio signal segment X2 having the time length Ts is inputted. In step 712, the square error between the audio signal segment X1 and the audio signal segment X2 is calculated as an evaluation function of the similarity. In steps 713, 714, 715, and 716, the delay time Td_op when the square error value is small is calculated. By searching the shift time Tc_opt and the shift time Tc and the delay time Td within a certain range determined by the initial value Tc_min of the shift time Tc, the maximum shift time Tc_max, and the maximum delay time Td_max, As a result of obtaining the square error of the two audio signal segments X1 and X2 that allow overlap and storing the shift time Tc and delay time Td when the square error is the smallest as Tc_opt and Td_opt, respectively, various audio signals within a certain range Since a combination of audio signal segments having a high degree of similarity and optimal for weighted addition can be selected from among the combinations of segments, there is an effect that there are few voice omissions and voice duplications and little deterioration in sound quality.

  In step 721, the audio signal having the time length Td_opt is input and output as it is with the second pointer as a starting point, and in step 722, step 723, and step 724, the first audio signal segment X1 ( 1 to Ts) and the second audio signal segment X2 (1 to Ts) are input, and the gradually increasing window function W1 (1 to Ts) and the gradually decreasing window function W2 (1 to Ts) are used to input the audio signal segment X1. And the audio signal segment X2 are weighted and output. In step 725, the time length Tt for outputting the audio signal as it is is calculated. In step 726, the starting point is calculated based on the first pointer, Tc_opt, Td_opt, and Ts. Weighted addition is performed by inputting a sound signal of time length (Tt-Td_opt) and outputting it as it is. An audio signal having a time length Td_opt that is continuous with the front end of the signal is output, an audio signal having a weighted addition of the time length Ts is output, and an audio signal having a time length (Tt−Td_opt) is continuous with the rear end of the weighted and added signal. As a result of inputting and outputting as it is, a continuous audio signal is output before or after the weighted and added audio signal, so that a continuous and smooth audio signal can be output, and before and after the weighted and added audio signal Since the audio signal having the total time length Tt is output, the time axis conversion ratio α can be finely set and can be accurately changed to a desired time axis conversion ratio.

  Then, in step 722, step 723, and step 724, the first audio signal segment X1 (1 to Ts) and the second audio signal segment X2 (1 to Ts) having high similarity are input, and the window function W1 that gradually increases is input. (1 to Ts) and the gradually decreasing window function W2 (1 to Ts) are used for weighted addition of the audio signal segment X1 and the audio signal segment X2 to output the audio signal having a high similarity and a segment length Ts. The segment sets X1 and X2 are input, and the first audio signal segment X1 and the second audio signal segment X2 are obtained by using the window function W1 that gradually increases the segment length Ts and the window function W2 that gradually decreases the segment length Ts. As a result of weighted addition so as to overlap, in any case, the time length of the weighted audio signal can be set to a constant segment length Ts, resulting in low sound quality. There is also an effect that is hard.

  Further, in step 705, the maximum delay time Td_max is calculated, in step 709, the maximum shift time Tc_max is calculated. In steps 708, 717, and 718, the shift time Tc is within the range from the initial value Tc_min to the maximum shift time Tc_max. In step 707, step 719, and step 720, the delay time Td is changed in the range of 0 to the maximum delay time Td_max. In steps 710, 711, and 712, the first pointer, the second pointer, and the shift time are changed. The start points of the audio signal segment X1 and the audio signal segment X2 are obtained using Tc and the delay time Td as parameters, and the audio signal segment X1 and the audio signal segment X2 having the time length Ts are input from the respective start points. And X2 are calculated, and the delay time Td_opt and the shift time Tc_opt when the square error is small are searched in Step 713, Step 714, Step 715, and Step 716, thereby obtaining the audio signal segment X1 and the audio signal segment X2. When searching for Tc_opt and Td_opt when the degree of similarity is high, the shift time Tc is limited to the range from the initial value Tc_min to the maximum shift time Tc_max, and the delay time Td is limited to the range from 0 to the maximum delay time Td_max. As a result, there is also an effect that an audio signal having a time axis conversion ratio α can be output stably.

  Although the small square error is used as an evaluation scale when obtaining the similarity applicable to the present invention, two similarities can be obtained even if the magnitude of the correlation function or the delay time Td is changed over a certain time. It is also applicable that the shift time Tc_opt of the audio signal segment is the same or little changed.

FIG. 8 is a flowchart in the case of time axis compression (α ≦ 1) of the voice speed conversion method of the present invention, a step of reading the time axis conversion ratio α of 802, and a step of initializing the first pointer of 803. , The step of setting a value after the audio signal segment length Ts relative to the first pointer to the second pointer of 804, and the time axis conversion ratio α, the audio signal segment length Ts, and the shift time based on Equation 9 of 805 A step of calculating a maximum delay time Td_max from an initial value Tc_min, a step of initially setting a minimum square error R_min of 806 to an initial value N, a step of initializing a delay time Td of 807 to an initial value of 0, and a shift of 808 The step of setting the initial value Tc_min of the shift time to the time Tc, and the delay time Td based on Expression 8 of 809 A step of calculating a maximum shift time Tc_max; a step of inputting Ts audio signal segments X1 (1 to Ts) starting from 810 (first pointer + Td); and 811 (second pointer −Tc + Td). ) As a starting point and inputting Ts audio signal segments X2 (1 to Ts), and based on Equation 812, the audio signal segment X1 and audio signal segment X2 at the shift time Tc and delay time Td The step of calculating the square error R (Tc, Td) is compared with the least square error R_min of 813 and the square error R (Tc, Td). If R_min is larger than the square error R (Tc, Td), step 814 is performed. Go to step 817, otherwise go to step 817;
The step of updating the square error R (Tc, Td) of 814 as a new least square error R_min, the step of updating the shift time Tc of 815 as the optimum shift time Tc_opt, and the delay time Td of 816 as the optimum delay time Td_opt The step of updating, the step of increasing the shift time Tc of 817 by one sample, the step of comparing the shift time Tc of 818 and the maximum shift time Tc_max, and the step if the shift time Tc is not greater than the maximum shift time Tc_max Returning to 810, if the shift time Tc is greater than the maximum shift time Tc_max, the process proceeds to step 819 to change the shift time Tc in the range from the initial value Tc_min to the maximum shift time Tc_max, and the delay time Td of 819 is changed. △ Td The delay time Td is compared with the maximum delay time Td_max. If the delay time Td is not greater than the maximum delay time Td_max, the process returns to step 808, and the delay time Td is the maximum delay time Td_max. If the delay time Td is greater than the delay time Td, the process proceeds to step 821 to change the delay time Td in the range from 0 to the maximum delay time Td_max, and Td_opt audio signal samples are input from the first pointer 821 as a starting point. A step of outputting, a step of inputting Ts audio signal segments X1 (1 to Ts) using (first pointer + Td_opt) at 822 as a starting point, and a step 823 (second pointer−Tc_opt + Td_opt) as a starting point. Ts voices Audio signal segment X2 (1 to Ts), audio signal segment X2 (i) multiplied by window function W2 (i) that gradually decreases with respect to audio signal segment X1 (i) based on Equation 824 A signal Y (i) obtained by adding the window function W1 (i) that gradually increases with respect to) to the signal Y (i) is calculated and output;

A step of calculating a time length Tt for outputting the input voice signal as it is based on Formula 10 of 825, and inputting (Tt−Td_opt) number of voice signals starting from 826 (second pointer −Tc_opt + Td_opt + Ts) A step of outputting as it is, a step of setting a first pointer in the next processing unit based on Formula 11 of 827, a step of setting a second pointer in the next processing unit based on Formula 12 of 828, If it is not the end of 829, the process returns to step 806 to repeat the process, and if it is the end, the process ends at step 830.

  However, in this flowchart, for simplicity, the unit time and the sampling period are expressed as being equal.

  In step 808, step 817, and step 818, the shift time Tc is changed from the initial value Tc_min to the maximum shift time Tc_max. In steps 807, 819, and 820, the delay time Td is changed from 0 to the maximum delay time Td_max. In step 810, the start point of the audio signal segment X1 is obtained and the audio signal segment X1 having the time length Ts is input. In step 811, the start point of the audio signal segment X2 is obtained and the audio signal segment X2 having the time length Ts is inputted. In step 812, the square error between the audio signal segment X1 and the audio signal segment X2 is calculated as an evaluation function of the similarity, and in steps 813, 814, 815, and 816, the delay time Td_op when the square error value is small. By searching the shift time Tc_opt and the shift time Tc and the delay time Td within a certain range determined by the initial value Tc_min of the shift time Tc, the maximum shift time Tc_max, and the maximum delay time Td_max, As a result of obtaining the square error of the two audio signal segments X1 and X2 that allow overlap and storing the shift time Tc and delay time Td when the square error is the smallest as Tc_opt and Td_opt, respectively, various audio signals within a certain range Since a combination of audio signal segments having a high degree of similarity and optimal for weighted addition can be selected from among the combinations of segments, there is an effect that there are few voice omissions and voice duplications and little deterioration in sound quality.

  In step 821, the audio signal having the time length Td_opt is input and output as it is with the first pointer as a starting point, and in step 822, step 823, and step 824, the first audio signal segment X1 ( 1 to Ts) and the second audio signal segment X2 (1 to Ts) are input, and the gradually decreasing window function W2 (1 to Ts) and the gradually increasing window function W1 (1 to Ts) are used to input the audio signal segment X1. And the audio signal segment X2 are weighted and output. In step 825, the time length Tt for outputting the audio signal as it is is calculated. In step 826, the start point is calculated based on the second pointer, Tc_opt, Td_opt, and Ts. Weighted addition is performed by inputting a sound signal of time length (Tt-Td_opt) and outputting it as it is. An audio signal having a time length Td_opt that is continuous with the front end of the signal is output, an audio signal having a weighted addition of the time length Ts is output, and an audio signal having a time length (Tt−Td_opt) is continuous with the rear end of the weighted and added signal. As a result of inputting and outputting as it is, a continuous audio signal is output before or after the weighted and added audio signal, so that a continuous and smooth audio signal can be output, and before and after the weighted and added audio signal Since the audio signal having the total time length Tt is output, the time axis conversion ratio α can be finely set and can be accurately changed to a desired time axis conversion ratio.

  Then, in step 822, step 823, and step 824, the first audio signal segment X1 (1 to Ts) and the second audio signal segment X2 (1 to Ts) having high similarity are input, and the window function W2 gradually decreases. (1 to Ts) and the gradually increasing window function W1 (1 to Ts) are used for weighted addition of the audio signal segment X1 and the audio signal segment X2 to output the audio signal having a high similarity and a segment length Ts. The segment sets X1 and X2 are input, and the first audio signal segment X1 and the second audio signal segment X2 are obtained by using the window function W2 gradually decreasing the segment length Ts and the window function W1 gradually increasing the segment length Ts. As a result of weighted addition so as to overlap, in any case, the time length of the weighted audio signal can be set to a constant segment length Ts, resulting in low sound quality. There is also an effect that is hard.

  Further, in step 805, the maximum delay time Td_max is calculated, in step 809, the maximum shift time Tc_max is calculated, and in steps 808, 817, and 818, the shift time Tc is within the range from the initial value Tc_min to the maximum shift time Tc_max. In step 807, step 819 and step 820, the delay time Td is changed in the range of 0 to the maximum delay time Td_max. In steps 810, 811 and 812, the first pointer, the second pointer and the shift time are changed. The start points of the audio signal segment X1 and the audio signal segment X2 are obtained using Tc and the delay time Td as parameters, and the audio signal segment X1 and the audio signal segment X2 having the time length Ts are input from the respective start points. And the square error of X2, and by searching the delay time Td_opt and the shift time Tc_opt when the square error is small in Step 813, Step 814, Step 815, and Step 816, the audio signal segment X1 and the audio signal segment X2 When searching for Tc_opt and Td_opt when the degree of similarity is high, the shift time Tc is limited to the range from the initial value Tc_min to the maximum shift time Tc_max, and the delay time Td is limited to the range from 0 to the maximum delay time Td_max. As a result, there is also an effect that an audio signal having a time axis conversion ratio α can be output stably.

  In this embodiment, the smallness of the square error in step 712 or step 812 is used as the evaluation scale when obtaining the similarity, but even if the correlation function size or the delay time Td is changed over a certain time, It can also be applied that the shift times Tc_opt of two audio signal segments having high similarity are the same or change little.

  Note that the smallness of the unnormalized square error in step 712 and step 812 is used as the similarity evaluation scale of the present embodiment, but the small normalized square error or the normalized correlation function is large. Can also be used. In this case, the amount of computation increases, but the evaluation scale does not depend on the amplitude of the audio signal, so that the similarity can be obtained without being influenced by the amplitude of the audio signal, and improvement in sound quality can be expected. .

  In this embodiment, when the square error between audio signal segments is obtained, the audio signal is input in units of audio signal segment length Ts in steps 710 and 711 or steps 810 and 811. You may input for every unit. For example, in the case of the time base extension shown in FIG. 5, the start point from 504_min to the end point of 509 is input, and in the case of time base compression shown in FIG. 6, the end of 609_min from the start point of 604. By inputting up to a point, when two audio signal segments are input while changing the delay time Td and the shift time Tc as in steps 710 and 711 and steps 810 and 811, or in steps 721 and When inputting an audio signal of time length Td_opt as in 821, or when inputting two audio signal segments in the shift time Tc_opt and delay time Td_opt as in steps 722 and 723 and steps 822 and 823, , Time length (Tt, as in step 726 and step 826 When entering the audio signal Td_opt), it can be prevented re-enter the audio signal. In this case, it is only necessary to cut out the voice signal that has already been input, and the number of times the voice signal is input can be reduced, so that the processing time can be shortened.

  The voice speed conversion device and voice speed conversion method of the present invention can change only the duration time without changing the fundamental frequency of the voice, and even if the speed is changed, the intelligibility is not easily lowered. The recorded audio signal can be applied to an application where it is necessary to reproduce the audio signal at a speed at which the user can easily hear it or at a desired speed.

The block diagram of one Embodiment of the audio | voice speed converter of this invention Block diagram of similarity calculation circuit of the same embodiment Block diagram of similarity calculation circuit of the same embodiment Block diagram of the determination circuit of the same embodiment Processing diagram in the case of time axis extension of the same embodiment (time axis conversion ratio α = 5/4) Processing diagram for time axis compression of the embodiment (time axis conversion ratio α = 4/5) The flowchart in the case of time-axis expansion | extension ((alpha)> = 1) of one Embodiment of the audio | voice speed conversion method of this invention. The flowchart in the case of time-axis expansion | extension ((alpha)> = 1) of one Embodiment of the audio | voice speed conversion method of this invention. The flowchart in the case of time-axis expansion | extension ((alpha)> = 1) of one Embodiment of the audio | voice speed conversion method of this invention. The flowchart in the case of time-axis expansion | extension ((alpha)> = 1) of one Embodiment of the audio | voice speed conversion method of this invention. The flowchart in the case of time-axis compression ((alpha) <= 1) of one Embodiment of the audio | voice speed conversion method of this invention. The flowchart in the case of time-axis compression ((alpha) <= 1) of one Embodiment of the audio | voice speed conversion method of this invention. The flowchart in the case of time-axis compression ((alpha) <= 1) of one Embodiment of the audio | voice speed conversion method of this invention. The flowchart in the case of time-axis compression ((alpha) <= 1) of one Embodiment of the audio | voice speed conversion method of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Memory circuit 102 1st switch circuit 103 1st buffer memory circuit 104 2nd buffer memory circuit 105 Similarity calculation circuit 106 Judgment circuit 107 Window function generation circuit 108 3rd switch circuit 109 4th switch circuit 110 1st 1 multiplication circuit 111 second multiplication circuit 112 addition circuit 113 second switch circuit 114 output buffer circuit 115 speed setting circuit 116 parameter storage circuit 117 pointer value calculation circuit 118 pointer control circuit 119 control signal generation circuit

Claims (6)

A similarity calculation circuit that obtains a similarity between an audio signal segment and a subsequent audio signal segment that allows a part of the audio signal segment to overlap, and a value that is high from the similarity output by the similarity calculation circuit. A determination circuit for detection, a parameter storage circuit for storing parameters relating to time information of two audio signal segments when the determination circuit detects a high similarity value, and a gradually decreasing window function and a gradually increasing window function are output. A window function generating circuit that performs multiplication of one window function output from the window function generating circuit on one audio signal segment based on a parameter stored in the parameter storage circuit, The other window output from the window function generation circuit for the other audio signal segment based on the parameter stored in the parameter storage circuit A second multiplication circuit for multiplying the number, an addition circuit for adding the output of the first multiplication circuit and the output of the second multiplication circuit, and the output of the addition circuit so as to obtain a desired time axis conversion ratio. And a switch circuit for switching and outputting the output of the adder circuit and / or one of the audio signal continuous at the front end of the adder circuit and the output of the adder circuit. apparatus. A storage circuit in which an audio signal is recorded, a pointer control circuit that outputs an address value to the storage circuit, and an output destination of the audio signal output from the storage circuit is a first buffer memory circuit and a second buffer memory circuit The first switch circuit selected from the first switch circuit, the first buffer memory circuit for storing the audio signal segment output from the first switch circuit, and the contents of the first buffer memory circuit A second buffer memory circuit for accumulating an audio signal segment that is allowed to be partially duplicated and subsequently output from the first switch circuit; and contents of the first buffer memory circuit and the second buffer memory circuit A similarity calculation circuit for obtaining a similarity to the content, a determination circuit for detecting a value having a high similarity from the similarity output from the similarity calculation circuit, and the determination circuit A parameter storage circuit for storing parameters used for the pointer value calculation circuit to obtain an address value output by the pointer control circuit when a value having a high degree of similarity is detected, and an audio signal recorded in the storage circuit A speed setting circuit for setting a time axis conversion ratio at the time of reproduction, and addresses of two audio signal segments from which the similarity calculation circuit should obtain a similarity based on the time axis conversion ratio set in the speed setting circuit A value is calculated, or based on the parameters recorded in the parameter storage circuit, two voice signal segments having high similarity and address values of voice signals continuous before and after are calculated and output to the pointer control circuit. A pointer value calculation circuit; a window function generation circuit that outputs a gradually increasing window function and a gradually decreasing window function; and the parameter storage circuit. Is multiplied by one of the window functions output from the window function generating circuit to the audio signal segment output from the storage circuit and stored in the first buffer memory circuit based on the parameter stored in the first buffer memory circuit. The other of the multiplication circuit and the window function generation circuit that outputs the audio signal segment output from the storage circuit and stored in the second buffer memory circuit based on the parameter stored in the parameter storage circuit A second multiplication circuit that multiplies the window function, an addition circuit that adds the output of the first multiplication circuit and the output of the second multiplication circuit, the output of the addition circuit, and the first switch circuit A second switch circuit that selects an output; an output buffer circuit that stores and outputs the output of the second switch circuit; and the similarity calculation circuit calculates the similarity Or when the adder circuit adds the output of the first multiplier circuit and the output of the second multiplier circuit, the first switch circuit is connected to the first buffer memory circuit side or the second buffer memory circuit. When the output of the adder circuit is output to the output buffer circuit, the second switch circuit is tilted to the adder circuit side, and in other cases, the audio signal continuous with before and after the output of the adder circuit is stored in the memory. An audio speed conversion apparatus comprising: a control signal generation circuit that controls the first switch circuit and the second switch circuit so as to output from the circuit to the output buffer circuit. Between two audio signal segments cut out while changing the shift time Tc and the delay time Td within a certain range determined by the time axis conversion ratio α, the audio signal segment length Ts, and the initial value Tc_min of the shift time with reference to the start point. Calculating a delay time Td_opt and a shift time Tc_opt when the similarity is high;
Inputting a voice signal having a time length Td_opt from the start point and outputting it as it is;
The audio signal segment X1 (1 to Ts) and the audio signal segment X2 (1 to Ts) are input using Tc_opt and Td_opt as parameters, and a gradually increasing window function W1 (1 to Ts) and a gradually decreasing window function W2 (1 to Ts). A step of weighting and outputting the audio signal segment X1 and the audio signal segment X2 using
The time length Tt of the audio signal to be output as it is is calculated based on the time axis conversion ratio α, the audio signal segment length Ts, and the shift time Tc_opt having a high degree of similarity, and the time length following the weighted and added audio signal (Tt−Td_opt) Input the audio signal of and output as it is,
Setting a starting point for the next process;
Step to return to the first step if not finished,
An audio speed conversion method characterized by comprising: Reading the time axis conversion ratio α,
Setting a first pointer and a second pointer after Ts of the first pointer;
Calculating a maximum delay time Td_max based on the time axis conversion ratio α, the audio signal segment length Ts, and the initial value Tc_min of the shift time;
Initializing the similarity,
The shift time Tc is changed from the initial value Tc_min in the range of the maximum shift time Tc_max determined by the delay time Td, the audio signal segment length Ts, and the time axis conversion ratio α, and the delay time Td is changed in the range of 0 to the maximum delay time Td_max. The start points of the audio signal segment X1 and the audio signal segment X2 are obtained by using the pointer 1, the second pointer, the shift time Tc, and the delay time Td as parameters, and the audio signal segment X1 and the audio signal having the time length Ts from the respective start points. Inputting a segment X2, calculating a similarity between X1 and X2, and searching for a delay time Td_opt and a shift time Tc_opt when the similarity is high;
Inputting a voice signal having a time length Td_opt from the first pointer or the second pointer as a starting point and outputting it as it is;
The audio signal segment X1 (1 to Ts) and the audio signal segment X2 (1 to Ts) having high similarity are input using the first pointer, the second pointer, Tc_opt, and Td_opt as parameters, and the window function W1 (1 ~ Ts) and the gradually decreasing window function W2 (1 ~ Ts), the audio signal segment X1 and the audio signal segment X2 are weighted and output, and
Based on the time axis conversion ratio α, the audio signal segment length Ts, and the shift time Tc_opt having a high similarity, the time length Tt for outputting the audio signal as it is is calculated, and the first pointer or the second pointer and Tc_opt, Td_opt, and Ts are calculated. Obtaining a starting point on the basis of a voice signal having a time length (Tt−Td_opt) and outputting it as it is;
A step of setting the first pointer and the second pointer after Ts of the first pointer based on the audio signal segment length Ts, the time length Tt for outputting the audio signal as it is, and the shift time Tc_opt when the similarity is high When,
If not finished, return to the step of initializing the similarity,
An audio speed conversion method characterized by comprising: Reading a time axis conversion ratio α (≧ 1.0);
Setting a start point in the first pointer;
Setting the value of the first pointer + Ts to the second pointer;
Calculating a maximum delay time Td_max = (Ts−α × Tc_min) / (α−1);
Initializing the similarity,
The shift time Tc is changed from the initial value Tc_min to the maximum shift time Tc_max = (Ts + Td) / α-Td, the delay time Td is changed from 0 to the maximum delay time Td_max, and (first pointer + Tc + Td) is the starting point Is input with the audio signal segment X1 with the time length Ts, and the audio signal segment X2 with the time length Ts is input with (second pointer + Td) as the starting point, and the similarity between X1 and X2 is calculated, and the similarity is high Searching for time delay time Td_opt and shift time Tc_opt;
Inputting a voice signal having a time length Td_opt with the second pointer as a starting point and outputting it as it is;
An audio signal segment X1 (1 to Ts) having a time length Ts is input with (first pointer + Tc_opt + Td_opt) as a starting point, and an audio signal segment X2 (1 to Ts) having a time length Ts with (second pointer + Td_opt) as a starting point. Ts) is input, and a gradually increasing window function W1 (1 to Ts) and a gradually decreasing window function W2 (1 to Ts) are used, and W1 (i) × X1 (i) + W2 (i) × X2 (i) is set to i Calculating and outputting within a range of = 1 to Ts;
Calculating Tt = (Ts−α × Tc_opt) / (α−1), inputting a voice signal having a time length (Tt−Td_opt) from (first pointer + Tc_opt + Td_opt + Ts) as a start point, and outputting the speech signal as it is;
Setting the first pointer + Tc_opt + Ts + Tt to the second pointer;
Setting a second pointer -Ts to the first pointer;
If not finished, return to the step of initializing the similarity,
An audio speed conversion method characterized by comprising: Reading a time axis conversion ratio α (≦ 1.0);
Setting a start point in the first pointer;
Setting the value of the first pointer + Ts to the second pointer;
Calculating a maximum delay time Td_max = ((2 × α−1) Ts−α × Tc_min) / (1−α);
Initializing the similarity,
The shift time Tc is changed from the initial value Tc_min to the maximum shift time Tc_max = 2 × Ts + Td− (Ts + Td) / α, the delay time Td is changed from 0 to the maximum delay time Td_max, and (first pointer + Td) is changed. An audio signal segment X1 having a time length Ts is input as a starting point, an audio signal segment X2 having a time length Ts is input using (second pointer −Tc + Td) as a starting point, and the similarity between X1 and X2 is calculated. Searching for a delay time Td_opt and a shift time Tc_opt when the degree is high;
Inputting a voice signal having a time length Td_opt starting from the first pointer and outputting it as it is;
An audio signal segment X1 (1 to Ts) having a time length Ts is input starting from (first pointer + Td_opt), and an audio signal segment X2 (1) having a time length Ts starting from (second pointer -Tc_opt + Td_opt) ˜Ts), and gradually increasing window function W1 (1 to Ts) and gradually decreasing window function W2 (1 to Ts), W2 (i) × X1 (i) + W1 (i) × X2 (i) calculating and outputting in the range of i = 1 to Ts;
Tt = ((2 × α−1) Ts−α × Tc_opt) / (1−α) is calculated, and an audio signal having a time length (Tt−Td_opt) is input starting from (second pointer−Tc_opt + Td_opt + Ts). And output as it is,
Setting a second pointer −Tc_opt + Ts + Tt to the first pointer;
Setting the first pointer + Ts to the second pointer;
If not finished, return to the step of initializing the similarity,
An audio speed conversion method characterized by comprising:

Images