KR19990079718A - Speech Signal Reproduction Using Multiband Stimulation Algorithm - Google Patents

Abstract

The present invention relates to a speech signal reproducing method using a multi-band excitation (MBE) speech signal parameter extraction algorithm. Adjust the playback speed by changing the playback time by repeating or deleting frames without changing. Therefore, the problem of adjusting the playback speed without changing the pitch and spectral envelope parameters in the same frequency range as the original sound when reproducing the voice signal, and improves the function of the vocoder having the same.

Description

SPEECH PLAYBACK METHOD USING MULTI BAND EXCITATION SPEECH PRODUCTION MODEL

The present invention relates to a method for reproducing a voice signal, and more particularly, to a method for reproducing a voice signal using multi band excitation (MBE) in a voice signal parameter extraction algorithm.

Sometimes a voice signal needs to be played back quickly or slowly. For example, in a telephone answering machine, the user wants to listen to the recording in detail, and the telephone answering machine can be operated slowly. In addition, even if the user knows exactly the subject of the recorded content, the user can quickly play back and prevent unnecessary waste of time. And the reproduction speed of such a telephone answering machine is made by a reproduction algorithm.

The general reproduction method is recognized by the reproduction time or the relative magnitude of the frequency of the synthesized speech signal. However, this technique has a problem of adjusting the pitch and spectral envelope in the same frequency range as the original sound. And voice signal processors using this technology, for example, the improvement of many vocoders such as IMBE, LPC, etc. is not easy.

The LPC model is typical of speech signal analysis. It is also used in many voice codecs, such as LPC-10 and CELP. This model is used to extract voice signal parameters, voice signal characteristics, and so on.

Recently, the MBE speech model has been used, which more efficiently extracts speech parameters. And in practice, many high quality vocoders are often based on the MBE model.

The reproduction technique does not change the pitch of the original audio signal, but increases the reproduction time of the original sound when it is played slowly and decreases it when it is played fast.

According to the audio signal recording / reproducing algorithm, any audio signal is recorded and reproduced depending on two parameters. One is pitch, the other is the spectral envelope. In order to reproduce the voice signal, it must be able to be analyzed and can be used as a voice signal of clear sound quality. And the spectral envelope parameters must remain the same. Therefore, speech signal reproduction can reproduce speech data of quality close to the original sound by changing time intervals in a state where the pitch and spectral envelope parameters do not change.

Thus a typical playback method is recognized by the relative magnitude of time or frequency. However, this technique has a problem of adjusting the pitch and spectral envelope in the frequency range. And it is not easy to improve many vocoders, such as voice signal processors using IMBE, for example IMBE.

An object of the present invention is to solve the above-described problem, the present invention is to implement a speech signal reproduction algorithm using the MBE parameter extraction algorithm.

1 is a flowchart showing a procedure of a voice signal reproducing method according to the present invention;

FIG. 2 is a flowchart showing the procedure of extracting the voice signal parameter shown in FIG. 1; FIG.

3A illustrates an analog voice signal according to an embodiment of the present invention;

FIG. 3B is a diagram showing an audio signal waveform at a slow reproduction speed of the analog audio signal shown in FIG. 3A; FIG. And

FIG. 3C is a diagram showing an audio signal waveform at a high reproduction speed of the analog audio signal shown in FIG. 3A.

* Description of symbols on the main parts of the drawings *

1: analog voice signal 2: first composite signal

3: second composite signal 4: third composite signal

5: fourth composite signal 6: fifth composite signal

7: sixth composite signal

According to a feature of the present invention for achieving the above object, a voice reproduction method for reproducing a voice signal using the MBE algorithm, comprising: extracting a voice signal parameter in a specific frame unit by receiving the analog signal; Synthesizing a speech signal based on the extracted parameters; Determining a first or second reproduction speed for reproducing the voice signal; A first reproducing step of repeatedly inserting at least one frame and reproducing the inserted frame and the frames before and after the frame by using the first reproducing speed; In the case of using the second reproduction speed, a second reproduction step of deleting at least one frame and concatenating and reproducing the phases of the frames before and after the deleted frame is performed.

The extracting step includes: extracting an initial pitch parameter; Subdividing the pitch parameter to reduce the error of the extracted pitch parameter; Classifying V / UV information through the subdivided pitch parameters; Improving pitch using the extracted and subdivided pitch parameters and the V / UV information.

Therefore, according to the present invention, the original analog voice signal is received and voice signal parameters are extracted in units of frames. The error is reduced and synthesized by the extracted parameters and information. Subsequently, during reproduction of the audio signal, a specific frame is inserted or deleted according to the reproduction speed, and the frames before and after the changed frame are connected in the same phase.

DETAILED DESCRIPTION Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 shows a procedure of a voice reproducing method according to an embodiment of the present invention.

Referring to the figure, in step S10, the speech signal parameters are extracted from the speech signal using the MBE speech extraction algorithm model. As shown in FIG. 2, in detail, the initial pitch parameter is extracted in step S30. In step S32, the extracted pitch parameters are subdivided. This is to reduce the error of the speech signal. Then, the V (Voiced) / UV (UnVoiced) information is classified from the pitch parameters subdivided in step S34. In step S36, the error of the voice signal is minimized using the pitch parameter and the V / UV information.

Subsequently, referring to FIG. 1, in step S12, the synthesized voice signal is of the same quality as the original voice signal in step S10. In step S14, the speed for reproducing the synthesized speech signal is determined. That is, when using a slow playback method, the flow advances to step S16 to insert at least one current frame repeatedly. Through this, the original voice signal should not be greatly modified. In step S18, the inserted current frame and the phase of the frames before and after the inserted frame are equally connected. In step S24, the audio signal at the reproduction speed is reproduced. If the playback is to be performed at a high speed, the process proceeds to step S20 so as to be close to the original audio signal and to delete at least one frame. In step S22, the phases of the frames before and after the deleted frame are equally connected. Subsequently, the audio signal is reproduced at a high reproduction speed.

And an example of one audio signal waveform will be described with reference to Figs. 3A to 3C. Fig. 3A is the original analog voice signal 1. One audio signal 1 is generated by repetition of the synthesized signals 2, 3, 4 shown in Fig. 3B at a slow reproduction speed. That is, the first synthesized signal 2 is an audio signal synthesized by the first harmonic which is the pitch frequency. The second synthesized signal 3 is synthesized by three times the frequency of the first harmonic. The third synthesized signal 4 is an audio signal synthesized by all harmonics.

In addition, due to the high reproduction speed, one voice signal 1 inserts a specific frame to generate the synthesized signals 5, 6 and 7 shown in FIG. 3C.

Specifically, the MBE voice parameters include pitch, V / UV information, and spectral width in all subbands.

Pitch extraction includes initial pitch evaluation and pitch improvement. In the initial pitch evaluation, an approximate pitch value can be obtained based on the automatic correlation between the front and rear pitch tracks and the current frame. And the frame length that is always processed is 160 sampling points or 20 ms.

Pitch improvement is achieved by a finer pitch that minimizes the error that occurs between the speech spectrum and the synthesized speech spectrum.

This error in all subbands causes the subbands to categorize voice or non-voice signals. Therefore, the same number of V / UV information as the number of subbands or harmonics is obtained.

By Fast Fourier Transform (FFT) the spectral amplitudes in all subbands can be calculated.

That is, after parameter extraction, in the parameters of voice signal pitch ω 0, voice V / UV information and voice spectral amplitude A _l (l is the subband index), if these parameters are very accurate and sufficiently extracted to determine the voice characteristics If so, a very high quality voice signal can be synthesized by the following equation.

[Equation 1]

S (t) of the above [Formula 1] means a synthesized speech signal, A ₁ is the l-th harmonic magnitude, w ₀ represents the initial phase of the frequency of the speech signal, φ Is the start phase of the current frame, which is the same as the end phase of the previous frame. And t is the sampling time for evaluating the parameter. Thus, these two frame parameters are inserted to maintain continuity between the previous frame and the current frame.

In the speech parameter extraction method using the MBE model, the slow playback method can be improved by iterative synthesis of the current frame. If iterative synthesis is made, the parameters of the current frame are reused, and the phase of the current synthesized frame that is the same as the currently synthesized frame is considered as the phase of the previous frame. This method has the same voice pitch and spectral envelope as the original sound. Therefore slow motion is improved. If the slow operation rate is r = p / q, p and q are positive integers. In other words, the time of slow motion has the same time r as the time of the original sound (p> q). Then we have q frames combined with q frames. p-q frames are iteratively synthesized. These repeated frames are arranged sequentially or randomly.

And in the voice parameter extraction method using the MBE model, if r = p / q <1, the fast operation can be improved by means of deleting some frames. For example, delete q-p frames from q frames. Their parameters are not used when frames are deleted. And the synthesizer waits for the next frame. Therefore, the current frame is deleted and then the phase between the previous frame and the next frame is made continuous.

Assuming a playback algorithm with a frame length of 20 ms, some deleted frames or some repeated frames do not change the perceptual sound quality of the speech signal. In general, speech signals synthesized by slow motion and fast motion become useful. . Therefore, better sound quality can be confirmed.

As described above, the present invention can easily implement a vocoder having the same frame by repeating or deleting a specific frame when the voice signal is reproduced using the MBE model.

Claims (2)

In the method of reproducing a speech signal using the MBE algorithm: Receiving one voice signal and extracting voice signal parameters in units of frames (S10); Synthesizing a speech signal based on the extracted parameters (S12); Determining a first or second reproduction speed for reproducing the synthesized speech signal (S14); In the case of using the first reproducing speed, the first reproducing step (S16, S18, S24) of repeatedly inserting at least one frame and continuously connecting the inserted frame and the phases of the preceding and following frames of the frame and ; In the case of using the second reproduction speed, a second reproduction step (S20, S22, S24) of deleting at least one frame and continuously connecting the phases of the frames before and after the deleted frame is reproduced. Audio playback method. The method of claim 1, The extraction step (S10) is: Extracting an initial pitch parameter (S30); Subdividing the pitch parameter to reduce an error of the extracted pitch parameter (S32); Classifying voice (Voiced) and noise (UnVoiced) information through the subdivided pitch parameters (S34); And improving the pitch by using the extracted pitch parameter and the voice and noise information (S36).