JPH0664480B2 - Audio signal compression method and memory writing method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for compressing an audio signal and a method for writing the compressed audio signal in a memory.

[0002]

2. Description of the Related Art Conventionally, as a digital compression method of an audio signal, an input analog audio signal is converted by an A / D converter.
It is known that after making 6-bit digital data, a difference PCM is made by utilizing the property that the difference between adjacent sample values is small, and unequal length code coding in which the larger the occurrence probability is, the shorter the code is.

By the way, when a voice signal is compressed and expanded, it is obvious that some kind of signal deterioration occurs.
A method in which the deterioration of the audio signal with respect to the compression rate is small is desirable.

As for the deterioration, there is a case where the original sound and the compressed / decompressed audio signal are physically small in deterioration.
There are cases where deterioration in human hearing is small. In the latter case, the sound is physically different from the original sound and is the same in terms of hearing. However, in view of the purpose of compression / decompression of the audio signal, it can be said that the purpose of compression / decompression has been achieved if the problems are focused on the audibility and no audible problem occurs.

This is expressed by the concept that the feature points are the same, especially in speech recognition. That is, if the characteristic points are the same, those sounds are recognized as the same audible feeling. Therefore, compressing the audio signal while maintaining the audible identity means that the characteristics of the original sound are faithfully preserved, which is a more preferable compression method.

Conventionally, the compression / expansion method has been mainly focused on the point that the physical deterioration is small, and there has been no example which discloses the technology of compressing / expanding while leaving the characteristic points in the sense of hearing.

Further, even when the audio signal is stored in the storage medium, it is preferable that the audio signal is compressed and stored while maintaining the sameness in hearing, resulting in good hearing in reading. it is obvious. Especially in the field of speech recognition, it is possible to compare feature points with each other in the comparison calculation, and the recognition speed can be increased.

[0008]

SUMMARY OF THE INVENTION In view of the above points, the present invention prevents compression of the audibility by compressing / decompressing while leaving the audible characteristic points, and the memory is stored in the memory. An object of the present invention is to provide a method for compressing a voice signal and a method for writing a compressed voice signal in a memory, which can obtain a good audible characteristic even when writing, and can increase the recognition speed in the field of voice recognition.

[0009]

An audio signal compression method according to the present invention sequentially supplies digital audio signals to a register,
A digital audio signal is taken out from this register at every predetermined time, a power spectrum signal is formed by orthogonally transforming the taken digital audio signal, and the power spectrum signal is given a predetermined value according to the frequency mel scale matched to the auditory characteristics. The above-mentioned problem is solved by dividing into the bands.

The memory writing method according to the present invention is
In the method of writing an audio signal compressed by the above audio signal compression method into a memory, a power spectrum signal further band-divided is used as an acoustic parameter expressing an envelope characteristic of an audio spectrum, and a local parameter is obtained from this acoustic parameter. The above problem is solved by writing the parameters in the memory device.

[0011]

The power spectrum signal obtained by orthogonally transforming the digital audio signal is divided into predetermined bands according to the frequency mel scale matched to the auditory characteristics, and compression is performed while leaving the auditory characteristic points. . Also, by writing the local parameter obtained from the band-divided power spectrum signal as an acoustic parameter expressing the envelope characteristic of the speech spectrum into the memory device, the speech signal is compressed while maintaining the auditory identity. It can be stored in memory and the audibility at the time of reading is improved.

[0012]

Embodiments of the present invention will be described below with reference to the drawings. The audio signal compression method of the present invention compresses a digital audio signal while maintaining the audible identity of the digital audio signal by utilizing orthogonal transform by fast Fourier transform.

In FIG. 1, a voice signal supplied to a microphone 1 is amplified by a microphone amplifier 2 and supplied to an A / D conversion circuit 4 via a low pass filter (LPF) 3 having a pass band of 5.5 kHz or less. It Further, the sampling clock of 12.5 kHz (80 μsec interval) from the clock generator 5 is supplied to the A / D conversion circuit 4, and the audio signal is converted into a digital signal of a predetermined number of bits (= 1 word) at this timing. It The converted audio signal is supplied to the register 6 of 5 × 64 words. Further, the frame clock at an interval of 5.12 msec from the clock generator 5 is supplied to the quinary counter 7, and this count value is supplied to the register 6 to shift the audio signal by 64 words, and the shifted 4 × 64 word The signal is taken out of the register 6.

4 × 64 taken out from this register 6
A signal of 256 words is supplied to the fast Fourier transform (FFT) circuit 8. Here, in the FFT circuit 8, for example, when the waveform function represented by n _f sampling data included in the time length of T is represented by the following equation (1), this is Fourier transformed and the following (2 The signal of the formula is obtained.

[Equation 1]

[Equation 2]

Further, the signal from the FFT circuit 8 is supplied to the power spectrum detection circuit 9, and the following (3)
The expression signal is obtained.

[Equation 3]

Since the Fourier-transformed signal is symmetrical on the frequency axis, half of the n _f pieces of data extracted by the Fourier transform is redundant data. Therefore, half of the data is removed and n _f / 2 pieces of data are extracted. That is, the 256-word signal supplied to the above-described FFT circuit 8 is converted to obtain a 128-word power spectrum signal.

This power spectrum signal is supplied to the emphasis circuit 10 and weighted for audible correction. Here, as the weighting, for example, a correction for enhancing the high frequency component of the frequency is performed.

This weighted signal is the band division circuit 1
1 and is divided into, for example, 32 bands in accordance with the frequency mel scale (MEL-SCALES) adapted to the auditory characteristics. Here, when it is different from the division point of the power spectrum, the signal is apportioned into each band and a signal corresponding to the amount of the signal in each band is extracted. As a result, the 128-word power spectrum signal described above is
It is compressed to 32 words while preserving the acoustic features.

This signal is supplied to the logarithmic circuit 12 and converted into a logarithmic value of each signal. As a result, the redundancy due to weighting in the emphasis circuit 10 described above is eliminated. Here, the logarithmic power spectrum shown in the following equation (4) is converted into a spectrum parameter x _(i) (i = 0, 1, ... 3).
1).

[Equation 4]

This spectral parameter x _(i) is supplied to the discrete Fourier transform (DFT) circuit 13. In this DFT circuit 13, for example, when the number of divided bands is M, this M-dimensional spectrum parameter x
_(i) (i = 0, 1 ... M-1) is regarded as a real number symmetric parameter of 2M-1 points, and DFT of 2M-2 points is performed. Therefore, the following expression (5) is obtained.

[Equation 5]

Further, since the function for performing this DFT is regarded as an even function, the following equation (6) is obtained, and from these, the following equation (7) is obtained. Acoustic parameters expressing the envelope characteristic of the spectrum are extracted by this DFT.

[Equation 6]

[Equation 7]

With respect to the spectrum parameter x _(i) thus DFTed, 0 to P-1 (for example, P = 8)
If the P-dimensional value up to the next is taken out and is taken as the local parameter L _(p) (p = 0, 1, ... P-1), the following equation (8) is obtained, and the spectral parameter is symmetric here. Taking this into consideration, if the following equation (9) is set, the local parameter L _(p) becomes the following equation (10). In this way, a 32-word signal is compressed into P (eg, 8) words.

[Equation 8]

[Equation 9]

[Equation 10]

This local parameter L _(p) is supplied to the memory device 14. The memory device 14 has a storage unit of 1 row P words arranged in a matrix of 16 rows, for example, and the local parameters L _(p) are sequentially stored for each dimension, and at the same time, from the clock generator 5 described above. 5.12
A frame clock at msec intervals is supplied to sequentially shift the parameters of each row in the horizontal direction. As a result, the memory device 14 has 16 frames (81.92 msec) of P-dimensional local parameters L _{(p) at} 5.12 msec intervals.
Minutes are stored and updated to new parameters sequentially for each frame clock.

[0024]

According to the audio signal compression method of the present invention, a power spectrum signal obtained by orthogonally transforming a digital audio signal is divided into predetermined bands according to a frequency mel scale matched to the auditory characteristics. Therefore, the compression is performed while leaving the audible characteristic points, and good compression can be performed while faithfully preserving the characteristics of the original sound. According to the memory writing method of the present invention, when the signal compressed by the audio signal compression method is written in the memory, the band-divided power spectrum signal is used as an acoustic parameter expressing the envelope characteristic of the audio spectrum. Since the local parameter obtained from this acoustic parameter is written in the memory device, the audio signal can be compressed and stored while maintaining the sameness in hearing sense, and the listening feeling at the time of reading becomes good. Further, when applied to the field of voice recognition, only comparison between feature points is required at the time of comparison, and the recognition speed can be increased.

[Brief description of drawings]

FIG. 1 is a block circuit diagram for explaining an audio compression method and a memory writing method according to the present invention.

[Explanation of symbols]

 1 ... Microphone 3 ... Low-pass filter (LPF) 4 ... A / D conversion circuit 5 ... Clock generator 6 ... Register 7 ... -Five-ary counter 8 ... FFT (Fast Fourier Transform) circuit 9 ... Power spectrum detection circuit 10 ... Emphasis circuit 11 ... Band division circuit 12 ... Logarithmic circuit 13 ... DFT (Discrete Fourier Transform) circuit 14 ... Memory device

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hisao Nishioka 1-7-4 Konan, Minato-ku, Tokyo Inside Sony Corporation Corporate Research Laboratory (56) 51-8-15) IEICE P.P. 83-84

Claims (2)

[Claims] 1. A power spectrum signal is formed by sequentially supplying a digital audio signal to a register, extracting the digital audio signal from the register at predetermined time intervals, and orthogonally converting the extracted digital audio signal. And a step of dividing the power spectrum signal into a predetermined band according to a frequency mel scale matched to the auditory characteristic, the audio signal compression method. 2. A digital audio signal is sequentially supplied to a register, the digital audio signal is taken out from the register at every predetermined time, and the taken digital audio signal is orthogonally transformed to form a power spectrum signal. In a method of writing a signal obtained by dividing a spectrum signal into a predetermined band according to a frequency mel scale matched to an auditory characteristic to a memory, a sound that expresses the envelope characteristic of a voice spectrum by further dividing the power spectrum signal into a band A method for writing a digital audio signal in a memory, comprising: a step of setting a parameter, a step of obtaining a local parameter from the acoustic parameter, and a step of writing the local parameter in a memory device.