JPH0775339B2 - Speech coding method and apparatus - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention reduces the amount of information in the case of converting a voice signal into a digital signal format for transmission during voice communication and transmitting it, thereby lowering the transmission speed and narrowing the band. The present invention relates to a speech coding method and apparatus for achieving efficient use of transmission band by utilizing the above method.

[0002]

2. Description of the Related Art Conventionally, a discretized audio signal is divided into processing units (frames), weighted by windowing, 3/4 frames are superposed, and a fast Fourier transform (FFT) is performed to change them from a time domain to a frequency domain. Orthogonal transformation is performed, and according to the Mel scale, which is one of the human auditory characteristics, the frequency [m] expressed from the physical frequency [Hz] to the Mel scale
[el], the mel frequency axis is equally divided into about 40 bands, the corresponding physical frequency axis is divided, and a plurality of spectrum information existing in the divided physical frequency axis bands There is a speech coding method in which representative spectrum information, that is, Mel scaled spectrum information is extracted and coded by ADPCM. However, with this method, the compression rate is low and effective use of the transmission band cannot be expected.

Therefore, the LO of the mel-scaled spectral information is
A method of calculating the mel-scaled cepstrum by taking G and again using FFT and vector-quantizing the low-order component (spectral envelopment information) has been devised, and it has become possible to significantly compress. However, this method is very difficult to realize because the amount of processing required for vector quantization is huge.

[0004]

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a speech coding method and apparatus which have a relatively high compression rate and are easy to realize.

[0005]

In order to solve the above problems, the present invention divides a discretized audio signal into processing units, performs orthogonal transform from the time domain to the frequency domain by fast Fourier transform, and performs physical transformation according to the Mel scale. In the case of replacing a specific frequency with a frequency represented by a mel scale, the mel frequency axis is equally divided into a plurality of bands, the corresponding physical frequency axis is divided, and a plurality of divided physical frequency axis bands exist. Of the maximum amplitude spectrum of the band and mel-scaled spectrum information composed of the combined spectrum power of the band is extracted from the spectrum information of the band, and is coded to reduce the amount of information. .

Further, in the above speech coding method, when transforming the discretized and framed speech signal from the time domain to the frequency domain by using the fast Fourier transform, the complex spectrum information is converted into the power spectrum information format. By expressing it, the phase information is ignored and the transmission parameters are reduced. At the time of decoding, the power spectrum information is arranged in the imaginary part, the real part is set to 0, and the fast inverse Fourier transform is performed. It is characterized in that the wave is synthesized.

Further, in the above speech coding method, the one having the largest absolute value is selected from the mel-scaled spectrum information as the maximum value information, and the maximum value information is compared with the absolute value of other mel-scaled spectrum information. , If the difference is ½ or more of the maximum value, 00, if ¼ or more and less than 1/2, 01,
It is characterized in that 2 bits are coded in 4 steps, that is, if it is 1/8 or more and less than 1/4, it is 10, and if it is less than 1/8, it is 11.

Also, an A / D converter that converts a voice signal into digital data, a first buffer that converts the digital data output from the A / D converter into processing units, and the first buffer Windower for weighting the data output from the buffer, the fast Fourier transformer for converting the data output from this windower into the spectrum information in the frequency domain, and the output from this fast Fourier transformer Mel scaler that obtains mel-scaled spectrum information by converting the generated spectral information from the physical frequency axis to the mel frequency axis represented by mel, and the mel-scaled spectrum information output from this mel-scaler A power spectrum encoder that obtains maximum value information and encoded mel-scaled spectrum information having the largest absolute value in the mel-scaled spectrum information from Max information, coding mel scale of spectral information extracted from the power spectrum encoder,
And a multiplexer for multiplexing position information of the mel-scaled spectrum output from the mel-scaler and transmitting it as digital data, and a mel-scaled spectrum for demultiplexing the digital data transmitted from the multiplexer. A demultiplexer for obtaining maximum value information having the largest absolute value in information, coded mel-scaled spectrum information, and position information of mel-scaled spectrum, and maximum value information and coded mel output from the demultiplexer. A power spectrum decoder for decoding Mel scaled spectrum information from the scaled spectrum information,
A mel-scale decoder that decodes mel-scaled spectrum information output from the power spectrum decoder and position information of the mel-scaled spectrum output from the demultiplexer into spectrum information on a physical frequency axis, and , A fast inverse Fourier transformer that transforms the spectral information output from the mel scale decoder into digital data in the frequency domain from the frequency domain, and the digital data output from the fast inverse Fourier transformer via a second buffer. And a D / A converter for converting into an analog audio signal.

[0009]

According to the present invention, the above means is one of the waveform coding in the frequency domain for the purpose of compressing by utilizing the human auditory characteristic. The audio signal is converted into the frequency domain and one of the human auditory characteristic is obtained. By replacing the physical frequency according to the Mel scale with the Mel frequency axis represented by the Mel scale, the phase information of the complex spectrum is ignored, and the Mel scaled spectrum information is encoded by 2 bits to obtain a large amount of information. The feature is that the amount is reduced.

[0010]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 shows a circuit configuration of an embodiment of the present invention.
In the figure, 20 is a speech encoder, and 21 is a speech decoder. First, the voice encoding process will be described.

A continuous audio signal S1 is sampled and quantized by an A / D (analog / digital) converter 1 to form a discrete signal S2. Next, the discrete signal S2 is converted into a signal S3 divided into processing units (frames) via the buffer 2. The length of this frame is 20 to 4 considering the steadiness of the audio signal.
It is about 0 msec. Next, the signal S3 is weighted by an appropriate windowing device 3 in order to improve the analysis accuracy when converting into the frequency domain. Examples of the windowing method include a Hamming window and a Hanning window. Next, the windowed processing unit signal S4 is transformed into the frequency domain. A fast Fourier transform (FFT) device 4 is used because a method having a high resolution and a high-speed execution is desirable as the conversion method. F
The number of FT analysis points is set to about 256 points from the trade-off between analysis time and resolution. 256 point FF
When T is performed, 128 pieces of complex spectrum information are obtained. The interval of the spectrum information is, for example, 31.25 Hz when the sampling frequency is 8 kHz and one frame is 32 mS.
Becomes Next, the mel scaler 5 replaces the spectrum information S5 and S6 calculated by the FFT unit 4 with the mel frequency axis represented by mel from the physical frequency axis. According to the Mel scale, the perception of pitch is not linear with the physical frequency as shown in FIG. 2 in the human auditory characteristic, and f _m = 1000 log ₂ (1 + f / 1000) f: physical frequency [ Hz] f _m: is approximated by Mel frequency [mel].

Specific means for achieving compression using the Mel scale will be described. First, the Mel frequency axis is equally divided. The smaller the number of divisions, the higher the compression rate. Next, as shown in FIG. 3, the physical frequency axis corresponding to the section divided by the mel frequency axis is divided. Representative spectrum (mel-scaled spectrum) from each section of the divided physical frequency axis
Calculate information. A method of calculating mel-scaled spectrum information will be described with reference to FIG. First, the spectrum information S5 and S6 in the complex format calculated by the FFT unit 4 are represented by power by the power spectrum calculator 50, and the spectrum information S7 is obtained by deleting the phase angle. This is because the human sense of hearing is insensitive to the instantaneous phase, and ignoring the phase does not affect the recognition of voice information. Next, the mel scale divider 51 divides the spectrum information S7 to obtain mel band spectrum information S8. Then, the maximum value position detector 52 detects the position information S9 of the maximum amplitude spectrum in the mel band of the spectrum information S8. Next, the combiner 53 obtains the combined spectrum power in the mel band. This combined spectrum power becomes Mel scaled spectrum information S10. As an example of the above, in the case of 15 divisions, 128 pieces of complex spectrum information are represented by 15 pieces of mel-scaled spectrum information and 15 pieces of position information, and the amount of information is about 1 at this point.
Compressed to / 4. Next, the power spectrum encoder 6 encodes the mel-scaled spectrum information S10. Details of the encoder 6 will be described. First, the maximum absolute value information S is extracted by extracting the maximum absolute value from the Mel scaled spectrum information.
11 Next, the maximum value information and the absolute value of the mel-scaled spectrum information are compared. If the difference is ½ or more of the maximum value, 00, ¼ or more and less than 1/2, 01, 1
If / 8 or more and less than ¼, it is 10, if less than ⅛, it is 11, and encoded mel-scaled spectrum information S12 is obtained by performing 2-bit encoding in four steps and greatly reducing the information amount. For example, in the case of 15 divisions, at this point, the information is compressed to about 1/25. Finally, the maximum value information S11, the encoded mel-scaled spectrum information (x division number) S12, and the mel-scaled spectrum position information (x division number) S9 are multiplexed by the multiplexer 7 and transmitted as digital data S13.

Next, the speech decoder 21 will be described. First, the received digital data S13 is demultiplexed by the demultiplexer 8. Next, the power spectrum decoder 9 sets the encoded mel-scaled spectrum information S12 based on the maximum value information S11 as 1/1, 1/2, 1/4, 1 of the maximum value.
Decoding is performed by expressing / 8 to obtain Mel scaled spectrum information S14. Next, the mel-scaled spectrum information S14 and the position information S9 obtained are arranged on the original physical frequency axis by arranging them at the position where the maximum amplitude spectrum information within the band divided by the mel-scale decoder 10 at the time of encoding was present. Restore to. By this operation, the pitch structure and intonation of the voice signal are almost preserved. Next, the fast inverse Fourier transform (IFFT) device 11 performs the conversion to convert the frequency domain to the time domain. At this time, the spectrum should originally have a phase angle in a complex format, but since it was expressed by the power value of that spectrum ignoring the instantaneous phase as described at the time of encoding, the frame when reconverted to the time domain The spectral information is arranged in the imaginary part S16 so that the smooth connection is made at the end, and the real part S1
5 is all 0. As a result, the data S17 re-converted to the time axis by the IFFT unit 11 becomes a sine wave composite, and the frame front end all starts from 0 in amplitude and ends in 0, and the frames before and after are smoothly connected. Therefore, it is understood that it is not necessary to perform overlap processing considering frame connection when dividing a frame during encoding. Next, the digital data S18 is transferred to the D / A via the buffer 12.
The converter 13 performs D / A conversion to obtain an analog decoded voice signal S19 and ends the processing.

[0014]

As described above, according to the present invention, in voice communication, when the voice signal is converted into a digital signal and transmitted, the Mel scale, which is one of the human auditory characteristics, is used. Is used, the spectrum information is compressed, the phase information is ignored, and the transmission parameter is quantized by 2 bits, so that the amount of information can be significantly reduced.

[Brief description of drawings]

FIG. 1 is a structural explanatory view showing an embodiment of the present invention.

FIG. 2 is a characteristic diagram showing an example of a relationship between a mel scale and a physical frequency according to the present invention.

FIG. 3 is a characteristic diagram showing an example of division of a physical frequency axis corresponding to division of a mel scale axis according to the present invention.

FIG. 4 is a structural explanatory view showing an example of a mel scaling processing according to the present invention.

[Explanation of symbols]

1 ... A / D converter, 2, 12 ... Buffer, 3 ... Window device, 4 ... FFT device, 5 ... Mel scaler, 6 ... Encoder,
7 ... Multiplexer, 8 ... Demultiplexer, 9 ... Decoder, 10 ... Mel scale decoder, 11 ... IFFT device, 13 ... D / A converter, 20 ... Speech encoder, 21 ... Speech decoder, 50 ... Power spectrum calculator, 51 ... Divider, 52 ... Maximum value position detector, 53 ... Combiner

Claims (4)

[Claims] 1. A case where a discretized audio signal is divided into processing units, orthogonal transformation is performed from a time domain to a frequency domain by a fast Fourier transform, and a physical frequency is replaced with a frequency represented by a mel scale according to the mel scale. In, in the mel frequency axis is divided into a plurality of bands on the physical frequency axis corresponding to it, the position information of the maximum amplitude spectrum of that band from the plurality of spectrum information existing in the band of the divided physical frequency axis and its A speech coding method characterized in that mel-scaled spectral information composed of band combined spectral power is extracted and coded to reduce the amount of information. 2. When transforming a discretized and framed speech signal from the time domain to the frequency domain by using the fast Fourier transform, the spectrum information in the complex format is represented in the format of the power spectrum information and the phase information is ignored. In order to reduce the transmission parameters, the power spectrum information is placed in the imaginary part at the time of decoding, the real part is set to 0, and the fast inverse Fourier transform is performed so that the sine wave is synthesized when re-transformed to the time axis. The speech coding method according to claim 1, wherein 3. The mel-scaled spectrum information having the largest absolute value is selected as maximum value information, and the maximum value information is compared with the absolute values of other mel-scaled spectrum information, and the difference is the maximum value. 00 if 1/2 or more, 1/2 or more 1/2
01 if less than 1 and 1 if 1/8 or more and less than 1/4
2. The speech coding method according to claim 1, wherein if the value is less than 0, 1/8, it is 11, and the coding is performed with 2 bits in 4 steps. 4. A / A in which a voice signal is digital data
A D converter, a first buffer that uses the digital data output from the A / D converter as a signal divided into processing units, and a windower that weights the data output from the first buffer. And a fast Fourier transformer that transforms the data output from this window multiplier into spectrum information in the time domain from the time domain, and the spectrum information output from this fast Fourier transformer is expressed in mel from the physical frequency axis. The mel scaler that obtains the mel-scaled spectrum information by converting it to the mel-scaled spectral axis that is output from this mel-scaled spectrum information has the largest absolute value in the mel-scaled spectrum information. A power spectrum coder for obtaining value information and coded mel-scaled spectrum information, and maximum value information and code extracted from the power spectrum coder. Mel-scaled spectrum information and a multiplexer for multiplexing position information of the mel-scaled spectrum output from the mel-scaler and transmitting it as digital data, and the digital data transmitted from this multiplexer. A demultiplexer that separates and obtains the maximum value information having the largest absolute value in the mel-scaled spectrum information, the encoded mel-scaled spectrum information, and the position information of the mel-scaled spectrum, and the output from this demultiplexer A power spectrum decoder for decoding the mel-scaled spectrum information from the maximum value information and the encoded mel-scaled spectrum information, the mel-scaled spectrum information output from the power spectrum decoder and the demultiplexer output Mel scale decoding for decoding from position information of mel scaled spectrum to spectrum information on physical frequency axis And a fast inverse Fourier transformer that transforms the spectrum information output from the Mel scale decoder into digital data in the frequency domain from the frequency domain, and the digital data output from the fast inverse Fourier transformer in a second buffer. A D / A converter for converting the signal into an analog speech signal via the speech coding apparatus.