KR19990068410A - On a Reduction of Pitch Search Time for IMBE Vocoder by Using the Spectral AMDF - Google Patents

Abstract

An improved method of shortening the pitch search time of a multi-band excitation vocoder using a spectral mean amplitude difference function that reduces the time required for initial pitch search when storing or transmitting a voice signal in a memory. The improved method for shortening the pitch search time of the multi-band excitation vocoder using the spectral mean amplitude difference function uses an IMBE speech coder that converts a signal into a frequency band and then band-divides a predetermined value for converting a predetermined value to the frequency domain. Taking a window of samples having; Calculating a spectral mean amplitude function for detecting the initial pitch in the frequency domain; Passing the initial pitch value as an argument of the pitch rechecking process. Therefore, when the voice signal is stored or transmitted to the memory, it is possible to reduce the time required for the initial pitch search.

Description

On-Reduction of Pitch Search Time for IMBE Vocoder by Using the Spectral AMDF}

The present invention relates to a method for shortening the pitch search time of an improved multi-band excitation vocoder using the spectral mean amplitude difference function, and to reduce the time required for initial pitch search when storing or transmitting a voice signal in a memory. An improved method of shortening the pitch search time of an improved multiband excitation vocoder using an amplitude difference function.

Voice coding methods for storing or transmitting voice signals in a memory include waveform coding, signal source coding, and hybrid coding.

Among them, the hybrid encoding method combines the memory efficiency of the signal source encoding method and the waveform encoding method of high quality sound.

Formant information in the hybrid coding method is encoded by the linear prediction coding method, and the RELP method, the VELP method, the MPLPC method, the CELP method, etc., have been proposed depending on how the remaining residual signals are encoded.

Among the hybrid coding methods, IMBE (Improved Multi-Band Excitation) method, which deals with speech in the frequency domain, is also recently adopted and used as a coding method for mobile and mobile communication.

1 shows a block diagram of a conventional IMBE vocoder.

The IMBE speech coder is a speech coder based on the MBE speech model, which makes it possible to synthesize speech more naturally than the conventional CELP speech coder by pitch prediction in the frequency domain.

Existing speech coder cannot distinguish each frame as voiced or unvoiced sound, but it can't make use of mixed region mixed with voiced / unvoiced sound, whereas IMBE voice coder converts the signal into frequency band and then divides the band into high-quality voice. Synthesis is possible, so natural speech synthesis can be performed.

However, since this requires a more sophisticated pitch search and requires a complicated pitch search process, it takes a lot of time to process the voice.

That is, natural pitch can be given by using the property that the pitch of voice is slowly changed in the initial pitch search process in the time domain. For this purpose, the initial pitch is found by performing the pitch estimation process before and after the current frame.

Since the pitch refinement process is performed in the frequency domain with respect to the found initial pitch, a large amount of time is required when searching for pitch.

In the existing IMBE, the initial pitch search selects an initial pitch value between 21 and 122 using the forward and backward pitch estimation algorithms in the time domain.

This method is to prevent sudden changes in pitch.

Since the pitch improvement process is performed again in the frequency domain with respect to the found initial value, a lot of time is required for the pitch search.

IMBE sets the pitch when the average square error of the pitch parameter value is minimum during pitch search.

For pitch search, first find the initial pitch in the time domain and then perform precise pitch search based on the initial pitch.

The initial pitch retrieval process involves firstly a backward measurement that maintains the continuity of the previous frame and pitch, Second, forward measurements to maintain continuity of future speech frames and pitches, , Is that.

The rear pitch measurement is calculated by the rear pitch estimation algorithm, and the front pitch measurement is calculated by the front pitch estimation algorithm.

The purpose of the pitch measurement is to detect the pitch P ₀ associated with the current speech frame S _w (n).

P ₀ is the fundamental frequency by equation (1) Associated with

here Is the radian value.

(One)

That is, the pitch estimation algorithm considers the pitch of the previous frame and the previous frame when the pitch of the current frame is determined.

The pitches associated with the two preceding voice frames are represented by P ₁ and P ₂ .

Similarly, the pitch of the two previous voice frames is represented by P- ₁ and P- ₂ .

Initial pitch measurements are limited to one of the sets {21, 21.5, ..., 121.5, 122}.

Initial pitch value is the fundamental frequency with quarter sample accuracy After the measurement, the desired pitch can be obtained.

The two part process is to improve the accuracy of the pitch detection.

One important aspect of the pitch detection algorithm is that the initial pitch detection algorithm uses a different window than the pitch improvement algorithm.

Window used for initial pitch detection (n) has a length of 301 samples.

Window used for pitch improvement (n) has a length of 221 samples.

Here the center points of the two windows must match.

When (n) is used, 61 samples overlap, When (n) is used, 141 samples overlap.

Pitch estimation is performed by comparing the result of E (P), and selecting the most suitable candidate within the range of 21 to 122. Named as

This process is referred to FIG. 2.

The E (P) function is defined as

(2)

here (n) is the initial pitch search window, and r (n) and s _LPF (n) equations are as follows.

Initial pitch value Is E ( Select the value that makes the smallest value.

However, if you find independent pitch values for the same frame, the measurements may change suddenly.

Sudden changes in pitch values can result in degraded synthesized sound.

Therefore, by using the property that the pitch changes slowly, the pitch measurement value obtained from neighboring frames is used to measure the pitch of the current frame.

The rear pitch estimation process is as follows.

Minimize the value of E (P) and obtain a P value that satisfies Equation 3. _It is defined as _B.

(3)

The forward pitch estimation process is as follows.

Forward pitch estimation is made in an attempt to preserve the continuity of the pitch between future speech frames.

The value of the _F CE (P ₀₎ to P ₀ value in the set to a minimum Defined by value.

After the front cumulative error and the back cumulative error have been calculated, compare each representative value. _F or One of the values of _B measures the initial pitch _Is selected by _I.

The pitch improvement algorithm improves the accuracy of pitch measurements from one half sample to one quarter samples.

Ten candidate pitches are generated from the initial pitch measurement.

These are each formally as follows.

(4)

Candidate pitches are shifted to the fundamental frequency.

Error function E _R (defined in equation 5) ) Is the candidate base frequency Calculated by substituting

E _R ( The candidate fundamental frequency obtained from the value of) is the modified fundamental frequency. Is selected.

(5)

As described above, the IMBE's complex pitch search method has an improvement in sound quality, but has a disadvantage in that a pitch search time is required.

If the processing time is long, the low cost water treatment DSP chip makes it difficult to process in real time, which increases the cost. Also, if the processing of the vocoder is complicated, the power consumption increases proportionally, which reduces the battery usage time in mobile phones.

An object of the present invention is an improvement using a spectral mean amplitude difference function to reduce the time required for initial pitch search by applying a Spectrum Average Magnitude Difference Function (SAMDF) when storing or transmitting a voice signal in a memory. To reduce the pitch search time of the multi-band excitation vocoder is provided.

1 is a block diagram of a conventional IMBE speech coder,

2 is a conventional initial pitch search configuration diagram;

3 is a block diagram according to the present invention;

4 is a result graph for voiced sounds according to the present invention;

5 is a comparison graph for the original voice and the synthesized voice according to the present invention;

6 is a comparison table of the results of processing time according to the present invention;

7 is a table of MOS results according to the present invention.

An improved multi-band excitation vocoder's pitch search time reduction method using the spectral mean amplitude difference function of the present invention provides a frequency domain using an IMBE speech coder that converts a signal into a frequency band and then bands it. Taking a window of a sample having a predetermined value to convert to; Calculating a spectral mean amplitude function for detecting an initial pitch in the frequency domain; And passing the initial pitch value as an argument of the pitch reconfirmation process.

Searching for the full range in which the fundamental frequency exists for odd frames of pitch; In even frames of the pitch, performing a search only within a predetermined range of the fundamental frequency of the previous frame.

A window of a sample having a predetermined value has a length of 221 samples.

The method may further include replacing the base frequency of the even frame with the base frequency of the previous frame when the base frequency of the even frame is out of a predetermined range of the base frequency of the previous frame.

According to the characteristics of the present invention described above, the improved pitch search time reduction method of the multi-band excitation vocoder using the spectral mean amplitude difference function according to the present invention is a method for comparing the spectral area when storing or transmitting a voice signal in a memory (SAMDF) By applying the Average Magnitude Difference Function, the time required for initial pitch search can be reduced.

Hereinafter, a method of shortening the pitch search time of an improved multi-band excitation vocoder using the spectral mean amplitude difference function of the present invention will be described in detail.

First, the pitch search method that can reduce the pitch search time while minimizing the deterioration of sound quality for the pitch search process that takes about 50% of the IMBE vocoder's processing time will be described.

3 is a block diagram showing a control process of the present invention.

After skipping the initial pitch search method in the time domain, we converted directly to the frequency domain, found the fundamental frequency using SAMDF, and then improved the pitch.

However, in this case, the sudden change of the pitch cannot be prevented. Therefore, the odd frame is searched for the entire range where the fundamental frequency exists and the even frame is searched only within the range of ± 20% of the fundamental frequency of the previous frame.

If the fundamental frequency of an even frame is out of ± 20% of the fundamental frequency of the previous frame, it is replaced by the fundamental frequency of the previous frame.

The SAMDF equation is as shown in equation 6.

This SAMDF method focuses on the fact that the energy of the first formant is about 10 dB greater than that of other formants in the case of voiced sound. Only process until

(6)

Where N is the length of one frame and S _w (·) is the amplitude spectrum of the audio signal.

Fig. 4 is a result graph for voiced sounds, Fig. 4 (a) is an original speech waveform, Fig. 4 (b) is an amplitude spectrum of original speech, and Fig. 4 (c) is an amplitude spectrum passing through SAMDF. .

Fig. 5 relates to the original voice and the synthesized voice, Fig. 5 (a) shows the original voice waveform, Fig. 5 (b) shows the conventional voice waveform of IMBE, and Fig. 5 (c) shows the synthesized voice waveform of the proposed IMBE. will be.

The equipment used for computer simulation is IBM-PC / 586 (200) system, which inputs data with sample rate of 8kHz by interfacing to commercialized 16-bit AD / DA converter.

For each sample, the length of one frame was treated with 160 specimens used by the IMBE vocoder.

In order to measure the performance of the treatment results, the following representative sentences were used as a sample by five speakers of various genders as follows.

Voice 1: Insu's little boy likes a genius boy.

Voice 2: / Jesus gave the lesson of creation.

Voice3: / Songsil University, Information and Communication Engineering, Voice Communication Research Team./

Voice 4: Thank you.

Pitch search was performed by implementing the pitch search process of IMBE vocoder in C-language.

For performance comparison, we compare the results of the proposed IMBE vocoder pitch search with the proposed method.

The conventional IMBE vocoder detects the fundamental frequency by performing both the initial pitch search in the time domain and the pitch improvement process in the frequency domain to detect the fundamental frequency, whereas the algorithm of the present invention uses the initial pitch search process in the interstitial domain. The pitch improvement process was performed after obtaining the initial fundamental frequency using the SAMDF method in the frequency domain.

6 shows a comparison between the processing time of the conventional IMBE and the processing time of the proposed IMBE.

7 shows the MOS test results.

An improved multi-band excitation vocoder's pitch search time reduction method using the spectral mean amplitude difference function has the following effects.

First, in order to solve the problem that the pitch search time in the IMBE vocoder is about 50% of the total coding processing time when using low-cost integer processing DSP, we use the SAMDF method to improve the pitch search time of the IMBE vocoder. Newly proposed.

Secondly, when IMBE coding was performed using the proposed pitch search for real speech, the processing time was reduced to 52.02% on average compared to the conventional IMBE pitch searching method, and there was almost no deterioration in sound quality as a result of MOS test.

Third, while the encoding stage of the conventional improved multi-band excitation vocoder uses a front pitch estimation block for initial pitch detection, a rear pitch estimation block, and a block for comparing the output values of the two blocks,

After taking a window of 221 samples for converting to the frequency domain, it has a spectral mean amplitude function block for detecting the initial pitch in the frequency domain, and passes the initial pitch value obtained as an argument of the pitch reconfirmation process for initial pitch search. This can reduce the time required.

Claims (4)

Taking a window of a sample having a predetermined value for conversion into a frequency domain using an IMBE speech coder that converts the signal into a frequency band and then band-divides; Calculating a spectral mean amplitude difference function for detecting an initial pitch in the frequency domain; And passing said initial pitch value as a factor of a pitch reconfirmation process. The improved method of shortening the pitch retrieval time of a multiband excitation vocoder using a spectral mean amplitude difference function. 4. The method of claim 1, wherein the window of samples having a predetermined value has a length of 221 samples. The method according to claim 1 or 2, further comprising: searching for the entire range in which the fundamental frequency exists for odd frames of the pitch; And performing a search only within a predetermined range of the fundamental frequency of the previous frame in the even frames of the pitch. 10. A method of shortening the pitch search time of an improved multi-band excitation vocoder using a spectral mean amplitude difference function. The improvement using the spectral mean amplitude difference function of claim 3, further comprising: substituting the fundamental frequency of the previous frame when the fundamental frequency of the even frame is out of a predetermined range of the fundamental frequency of the previous frame. A shorter pitch search time for a multiband excitation vocoder.