KR100322704B1 - Method for varying voice signal duration time - Google Patents

Abstract

PURPOSE: A method for varying voice signal duration time is provided to classify a voice signal into a vocal sound, a voiceless sound and a voiceless plosive sound in the time domain and change duration time of each of the classified sounds according to a pitch cycle and pitch time obtained through a pitch searching method and GCI detection method. CONSTITUTION: A voice signal is classified into a vocal sound, a voiceless sound and a voiceless plosive sound in the time domain(202). On the basis of pitch time, the vocal sound is repeated and removed in pitch cycle units to change the duration time of the vocal sound. The voiceless sound is repeated and removed in frame units to vary the duration time of the voiceless sound. The voiceless plosive sound is repeated and removed in mute units to change the duration time of the voiceless plosive sound(208).

Description

How to change the duration of voice signals

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech signal processing, such as speech recognition, synthesis, and analysis, and more particularly, to a method of changing the duration of a speech signal that classifies the speech signal into voiced sounds, unvoiced sounds, and unvoiced burst sounds to change respective durations.

Conventional pitch detection methods for detecting pitch periods of speech signals can be classified into a time domain method, a frequency domain method, and a time-frequency domain method.

The time domain detection method is a method of detecting a pitch by crystal logic after emphasizing the periodicity of a waveform. Since the time domain detection method is usually performed in the time domain, it is not necessary to change the domain, and the resolution of the pitch detection is high. However, when the phoneme spans the transition period, it is difficult to detect the pitch because the level change in the frame is severe and the pitch period varies. In particular, in the case of speech mixed noise, the decision logic for the pitch detection is complicated, which increases the detection error.

As the pitch detection method in the frequency domain, harmonic analysis, lifter, comb-filtering, and the like have been proposed as a method of detecting fundamental frequencies of voiced sound by measuring harmonic spacing of a speech spectrum. In general, since the spectrum is obtained in units of one frame (20 to 40 ms), even if the phoneme is shifted or fluctuates or the background noise is averaged, the spectrum is attenuated. However, the calculation is complicated because the conversion to the frequency domain is required in the processing process, and if the number of pointers of the fast fourier transform (FFT) is increased to increase the precision of the fundamental frequency, the processing time becomes long.

The time-frequency hybrid domain method takes the advantage of accurate calculation of the time-domain calculation time, the precision of the pitch, and the background noise and the phoneme change of the frequency domain method. These methods include the Cepstrum method and the spectral comparison method, and the error is increased when reciprocating the time and frequency domains, which may affect the pitch extraction, and the calculation process is complicated because the time and frequency domains are applied simultaneously. Although there is a problem in that the domain is not required, only simple operations such as sum, difference, and comparison logic are required. However, when the phoneme spans the transition period, it is difficult to detect the pitch because the level change in the frame is severe and the pitch period varies. In particular, in the case of speech mixed with noise, the decision logic for pitch detection is complicated, which leads to an increase in detection error.

On the other hand, three methods for detecting the pitch point of the conventional method using the measurement of the log matrix of the covariance matrix of the speech waveform, P pole (M) window in the M point window of the speech waveform ( Order of linear prediction coefficients) and a method of analyzing a linear prediction error sequence and using an electro glottal graph (EGG) signal. First, the method is not applicable to all signals because it is very difficult to determine GCI (Glottal Closure Instant, GCI) in some vowels, and there is a problem that it takes a lot of processing time. Second, the method has a problem that it is difficult to obtain an accurate closed phase when the closed phase of the gate is a high frequency or respiratory voice having a very short interval. In the third method, a microphone is attached to the larynx to directly measure the movement of the vocal cords.

The conventional method for detecting a section of unvoiced burst sound uses a method of changing the slope of a signal for a roughly divided voiced sound section, and a linear prediction coefficient (LPC Linear Prediction Coefficient) to indicate a relative amount of a periodic component and a noise source. There is a method using autocorrelation for the residual signal of the analysis. In the first method, the test for unvoiced rupture sound is performed only for the section judged as voiced sound, so there is a problem in that unvoiced rupture sound with the dominant voice component is not found. There is a problem in that it is not possible to distinguish between unvoiced and unvoiced burst sounds.

As a method of changing the duration of a conventional synthesized sound, there are a method using signal source coding and a method using waveform coding.

In the method using the signal source coding, each component of the signal is separated at the time of analysis and synthesized using the information of the separated signal components. Therefore, the error of analysis and the error at synthesis are added together, and thus the intelligibility of the synthesized sound is greatly reduced. have. In addition, the method using the waveform coding has a difficulty of separately processing a closed sound or a ruptured sound that cannot determine the pitch.

Also, in both methods, the accuracy of pitch detection greatly affects the naturalness of the synthesized sound, and the pitch time becomes an important variable for pitch repetition. That is, the conventional method of changing the duration of the voice signal has a problem that it is difficult to accurately detect the pitch and the pitch time and difficult to classify the sound source.

An object of the present invention is to classify a speech signal into voiced sounds, unvoiced sounds and unvoiced burst sounds in the time domain, and to change the duration of each classified sound source by a pitch period and a pitch time point obtained by a pitch search method and a GCI detection method, respectively. To provide a way to change the duration of the.

In order to achieve the above object, a pitch detection method of a speech signal according to the present invention comprises: a speech classification step of classifying the speech signal into voiced, unvoiced, and unvoiced sound intervals in a time domain; and pitching the voiced sound based on a pitch time point. Repeating and removing the voiced sound to change the duration of the voiced sound, repeating and removing the voiced sound by frame to change the duration of the unvoiced sound, repeating and removing the unvoiced burst sound by silence unit of the unvoiced burst sound And a changing step of changing the duration.

Hereinafter, a pitch detection method of a voice signal according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of a conventional voice signal processing apparatus for performing a method of changing a duration of a voice signal according to the present invention, and includes a microphone 100, a first attenuator 102, and a first low pass filter (LPF). : Low Pass Filter, LPF (104), Analog / Digital (A / D: Analogue / Digital, A / D) Converter 106, Input Port 108, Memory 110, Control Unit 112, The first output port 114, the D / A converter 116, the second LPF 118, the second amplifier 120, the speaker 122 and the second output port 124. Reference numeral 126 denotes a transport channel.

2 is a flowchart for explaining a method of changing a duration of a voice signal according to the present invention, comprising: classifying a voice signal into voiced, unvoiced, and unvoiced sound intervals (steps 200 to 202), a pitch period, A pitch point is obtained (steps 204 to 206), and a duration is changed (step 208) for each classified voice section.

Before explaining the operation and effect of the method for changing the duration of the voice signal shown in FIG. 2, the operation of the voice signal processing apparatus of FIG. 1 will be described. Once converted, it is amplified to a constant level in the first amplifier 102. The voice signal input through the microphone 100 is composed of components having a frequency in the range of 20 Hz to 20 KHz. In order to obtain the pitch among these components, the audio signal only needs to include the pseudo-transmission information component, and thus, the low-pass filter removes the frequency component of 4 KHz or more, which is the range of the pseudo-transmission information component frequency. The reason for removing components above a certain frequency is to reduce the number of processed data per second when this voice signal is converted to digital.

On the other hand, the low-pass filtered signal leaving only 4KHz or less signal components to be converted into a digital signal for processing by the computer, which is sampled by the A / D converter 106 converting the analog signal to a digital signal.

Since the sampling rate of the digital signal should be 8KHz, which is twice the maximum signal frequency (4KHz here) according to Nyquist's sampling theory, a clock frequency of 8KHz is input through the input terminal IN1. We also need to quantize the voltage level per sample, which uses a 16-bit (2 ¹⁶ = 65,536) level.

The digital voice signal thus processed is input to the controller 112 through the input port 108 to be calculated and processed. The input voice signal data is processed by the duration change method according to the present invention, and then stored in the memory 110 as necessary or through a decoding process using the data input through the transmission channel 126. Synthesize the audio signal.

In this way, the synthesized speech signal, which has been decoded by the controller 112, is transmitted to the first output port 114 in order to hear through the speaker 122 whether or not it is well processed.

The first output port 114 transmits the input digital signal to the D / A converter 116 converting the analog signal into an analog signal. In this case, the digital signal is processed and converted into an analog value at a sampling rate of 8KHz. Since the converted signal is still represented as an individual signal including harmonics, the signal is passed through the second low pass filter 118 to process only the signal of the base band. The signal thus processed may be amplified by the second amplifier 120 and output to the speaker 122 so that the speaker 122 may be output. As a result, the speaker 122 converts the signal whose duration has been changed into a sound pressure wave, thereby allowing the user to listen through the human ear.

A method of changing the duration of a voice signal according to the present invention performed in the above-described conventional apparatus will be described as follows.

First, a window function is applied to a voice signal input through the input port 108 or stored in the memory 110 (operation 200). That is, a process of applying a window function that separates and processes an average of 20 ms from speech signals continuously changing with time in order to classify the analysis section of the speech signal in time into small units.

After step 200, the voice signal is classified into a voiced sound zone, an unvoiced sound zone, and an unvoiced sound sound zone in the time domain (step 202). In this case, in order to detect an unvoiced rupture sound section, the characteristics of the unvoiced rupture sound are used.

One of the methods for detecting the interval of unvoiced rupture sound is to use the characteristic of unvoiced rupture sound having a high peak valley rate (PVR). Here, the PVR means the number of peaks and valleys in the waveform during a given negative interval. The other method has the transition and discontinuity characteristics in the speech signal, is accompanied by the silent section in one frame, and moves from the unvoiced voice to the voiced sound, so the energy function increases, the linear prediction coefficient error is large, and the safe pitch interval. The unvoiced burst sound section is detected by using the feature of the unvoiced sound sound in which the periodic section is detected.

The pitch period is detected using the pitch search method in the voiced sound section classified in step 202 (step 204). Here, unlike the pitch detection method, the pitch search method is a method of detecting the pitch period by repeatedly comparing the pitch duration value and the pitch gain that optimally satisfy the pitch duration condition (pitch period section) between the original voice and the synthesized voice.

After step 204, the pitch point is detected by applying the G-peak method (step 206). G-peaks are called G-peaks (after G at the beginning of the voice) because they appear in a waveform during a pitch period of voiced sound, very similar to the shape of the voice. In particular, since the G-peak coincides with the moment when the glottal is opened in the waveform of the voiced sound, when the G-peak is detected from the voiced sound, the pitch period of the voiced sound is obtained at every G-peak period, and the starting point of the voiced sound can be obtained at each peak position. have.

After step 206, the duration of each voice section is changed for each voice classified (step 208).

In order to change the duration in the voiced sound interval, the waveform is inserted or deleted in units of pitch periods obtained in step 204 based on the pitch time point obtained in step 206 by changing the duration. Conversely, reducing the duration means removing the waveform in one pitch period while maintaining the duration of the waveform. In order to increase or decrease the duration of the current voice waveform, the waveform must be inserted and deleted in pitch periods while maintaining the voice characteristics of the waveform. That is, waveforms of adjacent pitch periods are inserted and deleted. When the waveform of one pitch period is repeated or deleted, it is preferable to take a waveform of one pitch period from the time when the gate is started for natural connection with the adjacent waveform. Otherwise, when connecting waveforms of the same pitch period, the pitch period may be changed to an unexpected value.

Similarly, if you want to change the duration of the unvoiced section, change the duration by inserting or deleting the desired duration in frame unit, and change the duration by the desired duration by inserting or deleting the silent burst section in silence unit. do.

An embodiment of the present invention will be described as follows.

Change the duration of the 250 pronouncing words pronounced by the male and female announcers to 50%, 70%, 100%, 150% and 200%, and receive 20 randomly selected words from each of the five unlearned male and female listeners. Clarity was evaluated as shown in Table 1 below.

In addition, naturalness was evaluated as shown in Table 2 by allowing them to express the protrusion and unnaturalness of the sound they listened to.

As can be seen from Table 1 and Table 2, the evaluation result of clarity and naturalness was maintained at an average of 96.9%, and the naturalness was maintained at 96.6% to obtain a very good performance.

As described above, the method for changing the duration of the voice signal according to the present invention changes the duration by classifying the sound source, accurately detects the pitch and the pitch time, and thus has excellent naturalness and clarity even when the duration of the voice signal is changed. It works.

1 is a block diagram of a conventional voice signal processing apparatus for performing a method for changing the duration of a voice signal according to the present invention.

2 is a flowchart for explaining a method of changing a duration of a voice signal according to the present invention.

Claims (3)

In the method of changing the duration of the audio signal, Speech classification step of classifying the speech signal into voiced sound, unvoiced sound and unvoiced burst sound interval in the time domain: The duration of the voiced sound is changed by repeating and removing the voiced sound on a pitch period basis based on a pitch time point, and changing the duration of the voiced sound by repeating and removing the voiced sound on a frame basis, and the silent burst sound unit And changing the duration of the unvoiced rupture sound by repeating and removing the voice signal. The method of claim 1, wherein the voice classification step A method of changing the duration of an unvoiced rupture sound using a P.V. (PVR), a Z.R. (ZCR), a pitch gradient and a silent section as parameters. . The method of claim 1, wherein said changing step The pitch period is detected by a pitch search method, and the pitch time point is detected by a G.I. (GCI) detection method.