JP6731362B2 - Audio coding/decoding method - Google Patents

Description

The present invention relates to a method of coding voice at a transmitter and decoding voice at a receiver.

Due to the spread of high-speed lines such as optical fiber lines and the advent of large-capacity storage, the data size that we can handle is becoming unlimited, and in some cases the need for data compression is not recognized. On the other hand, data size is limited in data communication between outer space and the sea, or between space and the sea and the ground. In such a case, there is a demand for a voice compression method that does not cause interference and has a small data size.

MP3 and WMA are known as audio compression methods. These methods can compress not only linguistic information but also musical elements, and a compression rate of about 87% is possible.

In addition, a method disclosed in Patent Document 1 has been proposed as a method of clarifying a specific sound, for example, a broadcast sound of a station even if there is reverberation or background noise, although the sound is not compressed. This method divides the input voice signal into multiple frequency bands, divides the divided signals in each frequency band into multiple time frames, calculates the average power in each time frame, and calculates the power. The calculated average power in each time frame is compared with each other, the amplification degree of each signal divided by the time frame division unit is determined based on the comparison result, and the amplified signal in each frequency band is determined. This is the content to be added.

Japanese Patent No. 5115818

The compression rate of the existing method such as MP3 and WMA is not sufficient, and a higher compression rate is required in an environment where the data size is limited.

Further, in the embodiment shown in FIG. 6 of Patent Document 1, the voice is divided into four bands in the coding step, and the instantaneous amplitude value of the time waveform is converted into a code of +1 if it is positive, 0 if it is zero, and -1 if it is negative. A zero-cross wave generation process is performed and the zero-cross wave is further divided into four bands and then coded, but no compression process is performed.

In order to solve the above-mentioned problems, the present invention is configured for decoding by dividing the input voice into a plurality of bands and then compressing the input voice, and 0 and 1 encoding the input voice and then compressing the input voice. With regard to the above, the process of up-sampling the audio data compressed after being divided into the plurality of bands, and the up-sampling of the audio data compressed after the 0,1 encoding are made into the same band as the divided band in the coding. The division processing, the input voice is divided into a plurality of bands and compressed, and then the upsampled voice data of each band and the 0,1 encoding (rectangular clip voice) and the sampled voice data of each band It is composed of a process of synthesizing.

As a specific example of the process of compressing the input voice after dividing the input voice into a plurality of bands, a bandpass filter is used to divide the input voice into a plurality of bands, and a time for extracting an envelope of voice data in the divided frequency bands. It is conceivable to change the interval and then reduce the bit depth.

Further, after the input voice is divided into a plurality of bands and compressed, the amplitude information is digitized, the number of the band having the maximum value and its amplitude value are extracted as the amplitude information, and the band number having the maximum value is used in decoding. A mode in which the amplitude value is returned to the amplitude information is also conceivable.

Furthermore, in coding, the processing of dividing the input speech into multiple bands and then compressing is performed using the result of the factor analysis of the power change of the speech, and in decoding, the correlation value and factor loading amount for each factor obtained by coding. It is also conceivable that it is multiplied with.

ADVANTAGE OF THE INVENTION According to this invention, the compression rate at the time of audio|voice coding can be made high, and the audio|voice coding/decoding method of the audio|voice recognition (correct answer rate) in decoding can be provided.
Further, the method according to the present invention can be used as voice communication in any environment.

Diagram explaining an example of coding Diagram explaining an example of decoding Diagram explaining compression by changing the extraction time interval Diagram explaining compression by bit depth Diagram explaining 0,1 encoding A view similar to FIG. 1 for explaining another embodiment. A diagram similar to FIG. 2 for explaining another embodiment. A view similar to FIG. 1 for explaining another embodiment. A diagram similar to FIG. 2 for explaining another embodiment. (A) is a graph showing the relationship between the four factors and the center frequency, and (b) is a graph showing the relationship between the three factors and the center frequency. The figure explaining the method of factor analysis. Diagram explaining the calculation method during coding Diagram explaining the calculation method for decoding

A case where a voice represented by 16000 Hz and 16 bits is used as an original voice (input voice) will be described.
In coding, two processes of band division using a bandpass filter and 0,1 coding (rectangular clip voice) are performed in parallel on the original voice.

In band division using a bandpass filter, for example, it is divided into four bands of a first band 20 to 510 Hz, a second band 510 to 1270 Hz, a third band 1270 to 2700 Hz, and a fourth band 2700 to 6400 Hz. In this embodiment, the band is divided into four bands, but the number of bands may be 2, 3 or 5 bands or more.

Then, downsampling is performed for each of the divided bands. Downsampling is performed by changing the time interval for extracting the amplitude envelope of the voice curve as shown in FIG. In FIG. 3, the amplitude envelope is extracted every 20 ms (milliseconds). In the embodiment, the time interval is set to 40 ms so that 16000 Hz is compressed to 25 Hz by 1/640.

The audio data compressed by downsampling is further compressed by bit depth. The original voice is data of 16-bit depth (expressed in 65536 steps), but is compressed into 3-bit data (8 steps of 0 to 7) as shown in FIG. 4, for example. In the embodiment, the data is compressed into 8-bit (256 levels) data.

Further, as shown in FIG. 5, the 0, 1 encoding (rectangular clip voice) process in the coding is a process of setting 1 when the intensity of the input voice is 0 or more and 0 when the intensity of the input voice is less than 0. Compressed up to 1/10). That is, 0, 1 is determined every 1/1600 seconds.

As a method of determining 0 or 1, it is considered that an average value of values sampled 10 times is adopted, or an arbitrary (for example, first point) instantaneous value is adopted.

In this way, the amplitude information of the audio data of each band divided into bands using the bandpass filter is 25×8=200 bits, the number of bits after 0,1 encoding is 1600 bits, and the total number of bits is 200×. 4+1600=2400 bits.

In decoding, upsampling or the like is performed on the compressed audio data.
First, upsampling is performed on audio data compressed after being band-divided using a bandpass filter. In the embodiment, 640 times upsampling is performed.

As an upsampling method, it is conceivable to perform linear interpolation (interpolation) and then pass through a low-pass filter with a cutoff corresponding to the Nyquist frequency.

On the other hand, up-sampling is also performed on 0 and 1 encoded audio data transmitted simultaneously in parallel with the above. In this case, since it was compressed to 1/10, upsampling of 10 times is performed by the same method as described above.

Then, the 10-fold upsampled voice data is divided into 4 bands in the same band as the bandpass filter division, and the 4-divided voice data is combined with the 64-times upsampled voice data in each band. To do.

Evaluation conditions /Voice to be used: 15 sentences for each male and female of Japanese voice data of “NTT-AT multilingual voice database 2002” (sampling frequency 16 kHz, 16 bits), 15 sentences in total, and 30 sentences were used at random. ..
-Subjects: 12 students with normal hearing.
・Answer by keyboard.

Under the above conditions, when the total number of bits is 2400 by setting the number of bands to 4, the sampling frequency (0,1 coding) to 1600 Hz, the time interval to 40 ms, and the bit depth to 8 bits (compression rate 99. 1%), a mora correct answer rate of 80% or more and a perfect match correct answer rate of about 60% were obtained. This value is a value that can be sufficiently used for voice communication.

In FIG. 6 and FIG. 7, the amplitude information quantified by 8 bits in the above embodiment has four frequency bands, so that the band number can be expressed by 2 bits. Therefore, the band number having the maximum value and its amplitude value are taken out as amplitude information, and are restored to four pieces of amplitude information in decoding.
With such a configuration, it is possible to provide a coding/decoding method with high decoding even if the compression rate is increased.

FIGS. 8 and 9 explain still another embodiment, in which the coding of this embodiment uses a multilingual database to obtain from the output of the critical band filter as a band pass filter consisting of 20 bands. By using the result of the factor analysis of the power change of the generated speech, the characteristics of the speech are expressed by a small number of factors, and in decoding, the correlation value for each factor obtained by the coding is multiplied by the factor load amount. Is converted into amplitude information of 20 bands, and voice synthesis is performed.

It has been found that the amplitude spectrum of a voice can be represented by information using a load factor of three factors or four factors. FIG. 10A is a graph showing the relationship between 4 factors and the center frequency, and the number of factors may be 3 factors as shown in FIG. 10B.

Incidentally, FIG. 11 shows a method of deriving a factor load amount of voice by factor analysis. Further details of this method are disclosed in Science Reports (doi:10.1038/srep42468).

Further, FIG. 12 shows the calculation method in the correlation with the factor load of coding using the result of the factor analysis, and FIG. 13 shows the method of calculating the amplitude of each band using the factor load of the decoding. In the correlation with the factor load amount in the coding, the factor load amount of 4 factors or 3 factors for each band, which is prepared in advance, and the output of each band of the input speech are summed. On the other hand, when the amplitude of each band is calculated by using the factor loading amount in decoding, the correlation value for each factor obtained by coding is multiplied by the factor loading amount for each band.

Although the coding and decoding can be incorporated in separate devices, they are practically manufactured as one transceiver in which the coding unit and the decoding unit are incorporated.

Claims (2)

It is a method of coding voice and then decoding it.In the coding, the input voice is divided into a plurality of bands and then compressed, and the compressed amplitude information is digitized to determine the band number having the maximum value and its amplitude width. It consists of a process of extracting it as information and a process of compressing the input information after 0, 1 encoding, and the decoding is the compressed voice data in which the extracted band number having the maximum value and its amplitude width are digitized. And the process of up-sampling the digitized compressed voice data, and the process of up-sampling the voice data compressed after the 0,1 encoding and then dividing it into the same band as the division band in coding. , A process of synthesizing the voice data of each band that has been up-sampled after the input voice is divided into a plurality of bands and compressed, and the voice data of each band that has been up-sampled after being 0,1 encoded. Characteristic speech coding/decoding method. A method of coding voice and then decoding it. The coding is a process of dividing an input voice into a plurality of bands, a process of dividing the input voice using a result of a factor analysis of a power change of the voice, and input information. 0,1 encoding and then compression, the decoding upsamples the compressed speech data, and the upsampled speech data is correlated with each factor and factor load obtained by coding. A process of multiplying the amount for each band to convert into amplitude information; a process of up-sampling the voice data compressed after the 0,1 encoding and then dividing it into the same band as the division band in coding; And dividing the voice into a plurality of bands, compressing the voice, and then combining the voice data of each band that has been up-loaded with the voice data of each band that has been up-sampled after 0,1 encoding. Audio coding/decoding method.