KR100445342B1 - Time scale modification method and system using Dual-SOLA algorithm - Google Patents

Abstract

According to the present invention, the dual SOLA algorithm is used to change the voice speed so that the information that is deleted is reduced in the compression mode so that the " clicking " phenomenon occurs less. In the extended mode, the repeated information is reduced so that the " reverberation " The present invention relates to a voice speed conversion method and system.

A voice speed conversion method using a dual SOLA algorithm according to the present invention comprises the steps of: (a) multiplying a voice signal by a window and dividing the voice signal into a frame having a predetermined length; (b) selecting a composite shift value of the output signal fixed to the predetermined value according to the rate of change of the frame, and changing the analysis shift value of the input signal to obtain a next frame; (c) rearranging the frames by moving by the synchronization length obtained by cross correlation in the overlap period between the frames; And (d) outputting a voice signal whose speed is converted by adding weights to the overlapped frame samples in the overlap period.

Description

Time scale modification method and system using Dual-SOLA algorithm

The present invention relates to a speech processing method and system, and more particularly, to a method and system for converting speech speed using a dual SOLA algorithm.

Time Scale Modification (TSM) is a method of converting the playback speed of a signal by compressing or extending an input signal on the time axis. The music tempo conversion of karaoke equipment, the speed of voice playback for foreign language learning, and data compression And it can be applied to various fields such as restoration. In particular, in recent years, portable audio devices such as MP3 players have been used as a specific voice decoder to implement additional functions such as foreign language learning.

The TSM algorithm is a method of transforming the time axis and can be divided into a time domain method and a frequency domain method.

The typical time domain method uses the OLA (Overlap-Add) algorithm, which compresses or expands the input signal by overlapping and adding the adjacent signals by segmenting the input signal in window units, and using the pitch synchronization between neighboring windows. There are various SOLA algorithms and various SOLA transformation algorithms that can obtain more natural output sound by overcoming the disadvantages of OLA "clicking" (compression) and "reverberation" (expansion).

Typical frequency domain methods include the Griffin and Lim algorithm using STFT.

Algorithms generally used among the TSM algorithms are SOLA algorithms.

The SOLA algorithm is a representative method of converting tempo in the time domain and improves the disadvantages of the existing OLA method by performing overlap-add operation using pitch information between neighboring windows.

1 shows each parameter defined and used in the SOLA algorithm. In FIG. 1, winlen represents a frame length having a constant size by multiplying the original input signal by a window, Sa is an Analysys Shift, an analysis segment unit of the input signal, and Ss is a Synthesis Shift. The segment unit, Kmax, is used to match the pitch synchronization between two consecutive frames, and defines the maximum range of pitch search.

Also, the rate of change rate α is defined as the value of Ss / Sa. If the value of α is smaller than 1 (α <1), the voice speed becomes faster than the original sound due to the voice compression effect, and is larger than 1 (α>). 1) In case of this, the sound expansion effect is slower than the original sound speed. In general, the rate of change (α) is reproduced as a value between 0.5 (2 times faster) and 2.0 (2 times slower).

SOLA first multiplies the original signal by the window and cuts it into a frame with a constant size (winlen), then copies the first frame of the input signal x (n) into the first frame of the transformed output signal y (n), and then When the frame is obtained, the next frame is obtained by overlapping the window while moving at regular intervals as Sa, an analysis unit of the input signal.

When two consecutive frames are defined, the next step is to obtain a value of k, the pitch synchronization length between neighboring frames, rearrange the frames, add weights to the duplicated frame samples, and finally obtain a speed-converted output signal.

The method of obtaining the synchronization length value k between two consecutive frames is determined by maximizing the cross-correlation value of Equation 1.

In Equation 1 silver Wow The overlapping length of. The sync length value When is found, the final output signal Is in the overlapping interval by Equation 2 Wow Weights on Then, the frames are rearranged to obtain a signal having a length different from that of the original input signal, that is, a signal whose speed is converted.

FIG. 2 shows the operating principle of the compressed (α <1, fast) mode and the extended (α> 1, slow) mode of the SOLA algorithm according to the rate of change α = Ss / Sa.

In Fig. 2A and Fig. 2B, (i) is the original signal (ii) is the output signal rearranged by the rate of change rate (α) or the signal is not yet synchronized, and (i) is synchronized by the synchronization length value k. This is the final output signal produced.

In the conventional SOLA algorithm, since the current window is moved in the previous window direction and overlapped by the synchronization length value k for pitch synchronization between two consecutive frames, the voice information overlaps with each other. This is one cause of the phenomenon and "reverberation" in the extended mode.

The technical problem to be achieved in the present invention is to reduce the information to be deleted in the compression mode to reduce the "clicking" phenomenon, in the extended mode to reduce the repetitive information to convert the voice speed so that less "reverberation" phenomenon occurs In order to provide a voice speed conversion method and system using a dual SOLA algorithm.

1 shows each parameter defined and used in the SOLA algorithm.

Figure 2 shows the operating principle of the compression (α <1, fast) mode and the expansion (α> 1, slow) mode of the SOLA algorithm according to the rate of change (α = Ss / Sa).

3 illustrates a correlation between a pitch and a window length in an input voice signal.

4 shows the relationship between the composite shift Ss, the analysis shift Sa and the window length.

5 is a flowchart illustrating a voice speed conversion method using the dual SOLA algorithm according to the present invention.

6 is a block diagram of a voice speed conversion system incorporating a voice decoder and dual SOLA algorithms according to the present invention.

7 shows each test tone used in the simulation.

Figure 8 illustrates the buffer processing before applying the dual SOLA algorithm to ITU G.729.

9 is a compression mode of the speech rate conversion system applying the dual SOLA algorithm The output resulting speech waveforms for English male speech at = 0.500, 0.750 and 0.875 are shown.

FIG. 10 shows the output result speech waveform for the English male voice in the extended modes α = 1.250, 1.500, 1.750, 2.000 of the speech rate conversion system to which the dual SOLA algorithm is applied.

11 is a compression mode of the speech rate conversion system using dual SOLA algorithm including interpolation preprocessing At 0, 500, 0.750, and 0.875, the output waveforms for English speech are shown.

FIG. 12 is a comparison diagram of a case of including the case where there is no interpolation process when α = 0.5 for the same English voice.

13 is an extended mode of a speech rate conversion system using dual SOLA algorithms including interpolation preprocessing. At = 1.250, 1.500, 1.750, and 2.000, the output waveform for English speech is shown.

FIG. 14 is a comparison diagram of a case in which there is no interpolation process and the case of α = 2.0 for the same English voice.

In order to solve the above technical problem, a voice speed conversion method using a dual SOLA algorithm according to the present invention is a method of outputting a voice signal according to a voice speed desired by a user. Dividing into frames having; (b) selecting a composite shift value of the output signal fixed to the predetermined value according to the rate of change of the frame, and changing the analysis shift value of the input signal to obtain a next frame; (c) rearranging the frames by moving by the synchronization length obtained by cross correlation in the overlap period between the frames; And (d) outputting a voice signal whose speed is converted by adding weights to the overlapped frame samples in the overlap period.

The method may further include a preprocessing step of adjusting a sampling rate of the voice signal before step (a).

In addition, in the step (a), the window has a length including two pitch periods of the input voice signal, and has a length of 3 to 4 pitch periods in consideration of synchronization between neighboring frames.

In addition, the step (b) is characterized by using a different synthesis shift window length in the extended mode (slow speed) and compression mode (fast speed) of the voice.

In addition, the synchronization length is

(here, silver Wow Overlapping length of, Means an analysis shift which is an analysis segment unit of the input signal, Denotes a synthesis shift that is a synthesis segment unit of the output signal.) The cross-correlation value is obtained and determined as a value maximizing the cross-correlation value.

In addition, the synchronization length is a value in which the pitch information can be aligned, and is characterized in that it is larger than one pitch period and smaller than the window length.

The voice speed conversion system using the dual SOLA algorithm according to the present invention for solving the other technical problem in the voice speed conversion system for converting the speed of the voice signal output through the voice decoder, the voice signal output from the voice decoder A frame dividing unit dividing the frame into frames having a predetermined window length; A frame calculator which selects a composite shift value of the output signal fixed to the predetermined value according to the rate of change of the frame and changes the analysis shift value of the input signal to obtain a next frame; A synchronization length calculation unit obtaining a synchronization length by cross correlation in the overlap period between the frames; A frame repositioning unit for repositioning the frame by moving the synchronization length; And a weighting unit for outputting a voice signal whose speed is converted by adding weights to the overlapping frame samples in the overlap period.

The apparatus may further include a preprocessor configured to adjust a sampling rate of the voice signal before the voice signal output from the voice decoder is input to the frame divider.

In addition, the synchronization length calculation unit

(here, silver Wow Overlapping length of, Means an analysis shift which is an analysis segment unit of the input signal, The cross-correlation value is obtained by using a synthesizing shift which is a synthesis segment unit of the output signal. The synchronization length is determined by maximizing the cross-correlation value.

Hereinafter, the present invention will be described in detail with reference to the following examples, and the present invention will be described in detail with reference to the accompanying drawings.

First, the optimization of parameters applied to the dual SOLA algorithm according to the present invention will be described.

That is, the optimized value of the parameter is obtained, and the optimized parameter is applied to the algorithm according to the present invention. The parameters include window length, synchronization length, analysis shift, synthesis shift, and the like.

First, the window length will be described.

The window for segmenting the input signal should generally be large enough to include two pitch periods of the input audio signal, and up to 3-4 pitch periods in consideration of synchronization between neighboring windows. It may include a length.

In general, pitch periods vary from person to person, with women in the 50-250 Hz frequency range and men in the 120-500 Hz frequency range. Therefore, when the sampling rate of the input signal is 8KHz, assuming a pitch period of 100 samples on average, the window length should be at least 200 samples and not exceed 400 samples.

Figure 3 shows the correlation between the pitch and the window length in the input speech signal.

3A shows that when the window length is smaller than 2 pitch periods (assuming 1 pitch), the window length is too small and the overlapping area between neighboring windows is reduced, so that the probability of having common pitch information is almost 0%. This makes pitch alignment impossible in the overlapping section.

On the other hand, if the window length is too large (assuming 4 pitches) as shown in FIG. 3B, the overlapping area is increased and common pitch information is increased between neighboring windows. Therefore, "clicking" and "reverberation" occur in this area, thereby degrading the sound quality. Can be

As a result of various simulations on the actual window length, both males and females were able to obtain the optimal sound quality when the window length was 300 samples. However, when the window length is 320 samples or more, the "reverberation" phenomenon is severe in the female voice extension mode. Could see.

This occurs because the female voice has a shorter pitch period than the male voice. In the present invention, a window length of 300 samples is determined on the assumption that a common algorithm is applied to both male and female voices.

Second, the synchronization length Kmax will be described.

Kmax is to match the pitch synchronization between two consecutive frames and defines the maximum range of pitch search. Kmax is the minimum value that the pitch information can be aligned using a cross-correlation operation. It should be larger than 1 pitch period and basically exist between [0-window length]. If Kmax is smaller than 1 pitch period and pitch search probability is lowered, it causes significant sound quality degradation. Therefore, when the input voice signal is 8KHz sampling rate, Kmax is limited to at least 100 samples, and the maximum value of Kmax is simulated in the overlapping area of the neighboring window. When the Kmax value is more than 200 samples, jitters occur a lot. Therefore, it was limited to 2 pitch cycle or less.

In the present invention, the simulation results of various voices of males and females in the 100 Kmax and 200 sample intervals show that the voice output is good for listening when the Kmax = 100 is less "clicking" and "reverberation". Therefore, in the present invention, the optimum value of Kmax was 100 samples.

Third, the analysis shift Sa and the synthesis shift Ss will be described.

The dual SOLA algorithm according to the present invention, in contrast to SOLA, reduces the overlapping voice information between windows by reversing the moving direction of the window for retrieving the synchronization length k value for frame synchronization. Do this.

That is, as the speed change rate α = Ss / Sa is changed, the output signal is changed by changing the Sa value of the input signal. The Ss and Sa parameters should be basically larger than 1 sample, and their maximum value can be obtained by calculating the other parameter value only by obtaining the maximum value of either Ss or Sa by α = Ss / Sa.

4 shows the relationship between Ss, Sa and the window length.

For example, when Sa, the analysis shift of the input signal of FIG. 4 is larger than the window length, in the compression mode, the cross-correlation section is reduced by null to reduce negative smoothness, and in the extended mode, the sound is cut off by including a null value. Losing symptoms appear. Therefore, Sa must be smaller than the window length, and from the relation α = Ss / Sa, the composite shift Ss should be smaller than the window length / 2 considering the case where α is 0.5 (compression mode).

Simulation results under the above conditions In the present invention, good results can be obtained at Ss = 75, 50 samples.

Table 1 lists the optimal simulation results for each SOLA parameter.

S _s α = 2.0 α = 1.5 α = 0.875 α = 0.5 win300 win250 win200 win300 win250 win200 win300 win250 win200 win300 win250 win200 150 C B B C 125 B C B B B B C C 100 B B C B B B B B B B B C 75 B B C A A B B B C B B C 50 C A A B B B C

The simulation for each parameter assumes an input voice with ITU G.729 8Khz sampling rate, window lengths of 200, 250, and 300 samples, kmax is 100 and 200 samples, and synthetic shift Ss is 150 when the window length is 300 samples. , 125, 100, 75 samples, 7250 samples, 125, 100, 75, 50 samples, 200 samples 100, 75, 50 samples the rate of voice rate change, = 2.0 (twice slower), = 1.500 (1.500 times slower), = 0.875 (1 / 0.875x faster), Performed at = 0.5 (2 fold faster).

For the simulation, three voices were used as inputs: Korean male and female voices and foreign males. For the final subjective performance evaluation, three male and two female listeners were selected. . Subjective performance assessments have no problem in recognizing the overall context, even if there is some partial noise when the actual result is heard, so that each listener hears the rate-converted result sound and partially "clicking" or "reverberation". According to the degree of recognition of the same noise phenomenon, three grades were evaluated: A (nothing at all), B (littlely), and C (many).

As shown in the results of the simulated listening experiment in Table 1, the SOLA is calculated as the window length approaches 300 and the composite shift. The overall performance is good at = 75 samples and 50 samples. Especially in composite mode, = 75 samples and in compressed mode At 50 samples, the sound quality was good in each mode. The exception is the rate of change of the extreme speed of the algorithm. = 2.0 (twice slower), At 0.5 (2 times faster), partial "clicking" and "reverberation" noise phenomena could be recognized, especially in extended mode, where the noise was mainly noticeable in burst sounds such as "ㅋ, ㅌ, ㅎ". .

Therefore, in the present invention, based on the simulation results of Table 1, the optimal value of the SOLA parameter is obtained by window length = 300 samples, = 100 samples, and composite shift Is in extended mode = 75 samples in compressed mode The dual SOLA algorithm uses an optimal composite shift value in each compression and extension mode when the algorithm is set to = 50 samples.

Therefore, the dual SOLA algorithm used in the present invention enables the construction of a high performance speech rate conversion system by using different synthesized shift window lengths in the extended mode (slow speed) and the compressed mode (fast speed) of the voice.

5 is a flowchart illustrating a voice speed conversion method using the dual SOLA algorithm according to the present invention.

The voice signal output from the voice decoder is multiplied by a window and divided into a frame having a predetermined length (S510).

Here, the length of the window is preferably a length including two pitch periods of the input voice signal, and has a length of 3 to 4 pitch periods in consideration of synchronization between neighboring frames.

The method further includes a preprocessing step of adjusting the sampling rate of the voice signal before step 510 (S505).

The pretreatment process is described as follows.

The dual SOLA algorithm is designed to have optimal conditions for each parameter, but is still at the extreme rate of change of the algorithm. = 2.0 (twice slower), At = 0.5 (twice faster), partial "clicking" and "reverberation" noise phenomena were recognized. This noise phenomenon can improve the performance at the algorithm's extreme rate of change by interpolating the input voice signal with 8Khz sampling rate to have 16Khz sampling rate and using it as the input of dual SOLA.

The interpolation method used in the present invention uses linear interpolation, which is a simple structure and can be easily implemented in an algorithm, as a preprocessing procedure of dual SOLA. When applying the actual interpolation process, the sampling rate of the input signal is changed from 8Khz to 16Khz so that the optimal SOLA parameters, window length, Composite shift , The length is doubled.

The synthesized shift value of the output signal fixed to the predetermined value according to the rate of change of the frame is selected, and the next frame is obtained by changing the analysis shift value of the input signal (S520).

The frame is rearranged by moving by the synchronization length obtained by the cross correlation in the overlap period between the frames (S530).

The synchronization length is determined as a value maximizing the cross-correlation value obtained by Equation 1. In addition, the synchronization length is a value in which the pitch information can be aligned, which is larger than one pitch period and smaller than the window length.

In operation 540, a voice signal having a converted speed is output by adding weights to the overlapping frame samples in the overlap period.

In the present invention, the 8kHz, 80 sample / frame unit voice signal of the ITU G.729 voice decoder is input, and the output voice is optimized up to 2 times faster or 2 times slower according to the user's desired voice speed through the dual SOLA algorithm. A speech rate conversion system that enables reproduction in speech quality will be described.

The voice speed conversion system according to the present invention can be used for various purposes by implementing additional functions such as foreign language learning and portable recorder (voice recorder) in a portable audio device such as an MP3 player.

6 is a block diagram of a voice speed conversion system incorporating a voice decoder and dual SOLA algorithms according to the present invention.

The speech decoder 600 is a speech compression coder of 8 kbps rate using CS-ACELP (Conjugated Structured Algebraic Code Excited Linear Predictive). Here, the CS-ACELP speech coding method determines pitch and codebook parameters by an analysis / synthesis structure, and the excitation signal is vector quantized and stored in the codebook. This method has the advantage of improving the sound quality and reducing the amount of calculation compared to the previous CELP method and can encode the voice signal without degrading the voice quality. ITU G.729 is a two-step vector quantization that extracts 10th linear linear predictive coding (LPC) coefficients and analyzes them with LSP (Line Spectrum Pair) coefficients by constructing a frame of 8ms sampled input speech signal with 10ms and 80 samples. Through the process, the pitch value is quantized and transmitted, and the decoder recovers 80 samples / frame speech in 10 ms units. In order to integrate the dual SOLA algorithm into the ITU G.729 voice decoder, 80ms / frames of 10ms are accumulated in the buffer as much as the window length of the dual SOLA, and the algorithm is applied.

The frame dividing unit 610 divides the voice signal output from the voice decoder into a frame having a predetermined window length.

The frame calculator 620 selects the composite shift value of the output signal fixed to the predetermined value according to the rate of change of the frame, and changes the analysis shift value of the input signal to obtain the next frame.

The synchronization length calculator 630 calculates a synchronization length by cross correlation in the overlap period between the frames. At this time, the synchronization length is determined as a value maximizing the cross-correlation value obtained by Equation 1. In addition, the synchronization length is a value in which the pitch information can be aligned, which is larger than one pitch period and smaller than the window length.

The frame repositioning unit 640 rearranges the frames by moving by the synchronization length.

The weighting unit 650 outputs the voice signal whose speed is converted by adding weights to the overlapping frame samples in the overlap period.

The preprocessor 660 adjusts the sampling rate of the voice signal before the voice signal output from the voice decoder 600 is input to the frame divider 610. This preprocessing process is performed through an interpolation process.

In the present invention, the speech rate conversion system may be implemented using a preprocessing process, and the speech rate conversion system may be implemented without using the preprocessing process.

Therefore, in the voice rate conversion system without interpolation procedure, 10 SO80 samples / frames of ITU G.729 are stored in a buffer as long as 300 sample window lengths corresponding to 37.5 msec, and dual SOLA is applied. On the other hand, the input signal that has undergone the interpolation process is applied to the algorithm after accumulating data by using the window buffer of 75msec unit which is twice the length.

In the present invention, the simulation and performance evaluation of the speech rate conversion system is performed. The optimal parameters of the Dual SOLA used for the simulation were window length = 300 samples, = 100 samples, and composite shift Was set to 75 samples in the extended mode and 50 samples in the compressed mode.

The test voices used in the simulation were inputted into three voices: Korean male (3.79sec), female voice (4.01sec), and foreign male (3.05sec), and each test voice was 8Khz sampling rate, 16bit, It is mono.

Male voice: "The seasons are full of autumn sky."

Femininity: "The seasons are filled with autumn sky."

English: "We have just dock on the beautiful shores of the same problem."

7 shows each test tone used in the simulation.

Figure 8 illustrates the buffer processing before applying the dual SOLA algorithm to ITU G.729.

ITU G.729 consists of 8khz sampled digital input signal with 10msec 80samples / frame, so proper buffer adjustment is required to apply Dual SOLA with window length of 300 samples.

In FIG. 8, since dual SOLA has to store input data by 300 samples of window length, 320 samples corresponding to 4 frames of ITU G.729 are accumulated in a buffer, and 20 samples of which the window length is exceeded are stored in another buffer. To be used for the next window operation.

Dual SOLA has a difficult disadvantage of index adjustment because the buffer index varies depending on the rate of change α. For example, in FIG. 8, when S_s is 75 and α is 1.75, S_a is 43. To obtain the next 300 samples after processing one frame, G is 238 samples except for the remaining 42 samples and the 20 samples stored in the remaining buffer. It is input from the .729 decoder. However, since the G.729 decoder is always fixed at 80 samples / frame, the next 240 samples are input and 238 of them are used as input buffer samples, and the remaining 2 samples are stored in the remaining buffer and then processed. Used when This index adjustment requires a proper buffer adjustment according to the actual speed change rate since the buffer index changes every time since the speed change rate α is changed by the user.

9 and 10 show the compression mode of the speech rate conversion system applying the dual SOLA Output waveforms for the English male voice are shown in the = 0.500, 0.750, 0.875 and extended modes α = 1.250, 1.500, 1.750, 2.000.

In the present invention, the output results for the Korean male voice and the female voice show similar conclusions, and thus the output voice waveform is omitted and explained based on the English male voice.

The simulation results in the compression mode of FIG. 9 show that the overall sound quality is good according to the rate of change. However, in the case of α = 0.5 (2 times faster), the clicking phenomenon in the "dock" portion indicated by the dotted line is particularly exceptional. I was able to confirm it and was able to perceive it in the actual voice listening test. On the other hand, in Figure 9 of the extended mode, it can be seen that reverberation occurs over the entire context when α = 2.0 (twice slow).

Ultimate rate rate of these algorithms The clicking and reverberation phenomena at = 2.0 `,` 0.5 are also consistent with the simulation results and can be overcome by preprocessing the input signal.

The performance of the voice speed conversion system using the dual SOLA algorithm described above is the extreme rate of change of the algorithm. = 2.0 `,` Except for 0.5, the overall performance is good at the remaining change rate.

It shows that the algorithm can improve the performance at the rate of change of the extreme speed of the algorithm by converting the 8Khz sampling rate of the ITU G.729 decoder into 16Khz sampling rate and using it as the input of dual SOLA. The interpolation technique used in the present invention uses linear interpolation, which is a simple structure and can be easily implemented in an algorithm, as a preprocessing procedure of dual SOLA. When the actual interpolation process is applied, the sampling rate of the input signal is changed from 8Khz to 16Khz, so that the optimal dual SOLA parameters are doubled, and the window length is 300 samples to 600 samples, 100 samples to 200 samples, Is changed from 75 samples to 150 samples in the extended mode and 50 samples to 100 samples in the extended mode. In this case, the window length is doubled, which increases the cross-correlation calculation of the frame synchronization search that has the largest amount of computation in the dual SOLA.

Therefore, in the present invention, in order to reduce the calculation amount of the algorithm, the simulation of the dual SOLA parameter set is changed to 1.25 times, 1.5 times, 1.59 times, 1.67 times, 1.75 times, and the result is almost equivalent to the condition that the parameter set is twice even at 1.75 times. As a result, sound quality can be obtained, and as a result, one-eighth of the total calculation amount can be reduced. The dual SOLA parameter set has a window length of 525 samples, = 175 samples, 131 samples in extended mode and 88 samples in compressed mode.

11 is a compression mode of the speech rate conversion system using dual SOLA algorithm including interpolation preprocessing At 0, 500, 0.750, and 0.875, the output waveforms for English speech are shown.

FIG. 12 is a comparative view illustrating a case in which there is no interpolation process in the case of α = 0.5, and FIG. 12A is an enlarged portion of a “dock” in which a clicking phenomenon occurs in FIG. 9A. As a result of the processing, the "dock" portion of the voice signal is enlarged. As can be distinguished from the figure, the clicking phenomenon was reduced a lot in the 0.44 ~ 0.46 sec section of the output signal which passed the pre-interpolation process, and the break between the sound and the sound disappeared.

13 is an extended mode of a speech rate conversion system using dual SOLA algorithms including interpolation preprocessing. At = 1.250, 1.500, 1.750, and 2.000, the output waveform for English speech is shown.

FIG. 14 is a comparative view illustrating a case where there is no interpolation process for the same English voice and when there is no interpolation process. FIG. 14A is an enlarged view of a "dock" in which reverberation occurs in FIG. 10D. As a result of passing the pre-interpolation process, the "dock" portion of the speech signal is enlarged. As shown in FIG. 14, it can be seen that FIG. 14A has a larger number of repeated signals than FIG. 14B, and reverberation is reduced in the output signal passing through the pre-interpolation process.

The drawings and specification are merely exemplary of the invention, which are used for the purpose of illustrating the invention only and are not intended to limit the scope of the invention as defined in the appended claims or claims. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible from this. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

According to the present invention, we propose an optimized dual SOLA algorithm for the ITU G.729 decoder speech signal based on simulation and theoretical analysis of various SOLA parameters. reverberation "can improve the overall performance by including a simple linear interpolation process in the speech input stage.

In addition, the system according to the present invention can be applied to a variety of digital language learning system that can increase the language learning efficiency of the user.

In addition, the present invention can implement a system in which the ITU G.729 decoder and the dual SOLA algorithm are interconnected and integrated in the MP3 player when the digital language learning function of the MP3 player is implemented as well as the voice rate conversion algorithm itself.

Claims (9)

In the method for outputting a voice signal at a desired voice speed, (a) multiplying a voice signal by a window to divide the speech signal into frames having a predetermined length; (b) selecting a composite shift value of the output signal fixed to the predetermined value according to the rate of change of the frame, and changing the analysis shift value of the input signal to obtain a next frame; (c) rearranging the frames by moving by the synchronization length obtained by cross correlation in the overlap period between the frames; And (d) adding a weight to the overlapping frame samples in the overlap period and outputting a voice signal whose speed is converted; And a preprocessing step of adjusting a sampling rate of the voice signal before step (a), In step (a), the length of the window has a length including two pitch periods of the input voice signal, and has a length of 3 to 4 pitch periods in consideration of synchronization between neighboring frames, In the step (b), the voice speed conversion method using a dual SOLA algorithm, characterized in that using a different synthesis shift window length in the expansion mode (slow speed) and compression mode (fast speed) of the voice. delete delete delete The method of claim 1, wherein the synchronization length is (here, silver Wow Overlapping length of, Means an analysis shift which is an analysis segment unit of the input signal, Denotes a composite shift unit which is a composite segment unit of the output signal.) Obtaining the cross-correlation value using, and determines the value to maximize the cross-correlation value, the voice speed conversion method using a dual SOLA algorithm. The method of claim 5, wherein the synchronization length is Pitch information is a sortable value, the voice speed conversion method using a dual SOLA algorithm characterized by greater than 1 pitch period, less than the window length. In the voice speed conversion system for converting the speed of the voice signal output through the voice decoder, A frame dividing unit dividing the voice signal output from the voice decoder into a frame having a predetermined window length; A frame calculator which selects a composite shift value of the output signal fixed to the predetermined value according to the rate of change of the frame and changes the analysis shift value of the input signal to obtain a next frame; A synchronization length calculation unit obtaining a synchronization length by cross correlation in the overlap period between the frames; A frame repositioning unit for repositioning the frame by moving the synchronization length; And A weighting unit configured to output a voice signal whose speed is converted by adding weights to the overlapping frame samples in the overlap period, And a preprocessing unit for adjusting a sampling rate of the voice signal before the voice signal output from the voice decoder is input to the frame divider. delete The method of claim 7, wherein the synchronization length calculation unit (here, silver Wow Overlapping length of, Means an analysis shift which is an analysis segment unit of the input signal, Denotes a synthesis shift, which is a synthesis segment unit of the output signal.), And the synchronization length is determined by a value maximizing the cross correlation value, using the dual SOLA algorithm. Speed conversion system.