JP3351746B2 - Audio signal compression method, audio signal compression device, audio signal compression method, audio signal compression device, speech recognition method, and speech recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to efficient transmission of an audio signal obtained by converting music into an electric signal or an audio signal obtained by converting human voice into an electric signal through a small-capacity transmission line to a recording medium. In particular, when compressing an audio signal or a voice signal based on the weighting on the frequency corresponding to the auditory sensitivity characteristic which is a human auditory characteristic, a higher sound quality than before can be achieved. TECHNICAL FIELD The present invention relates to an audio signal compression method and an audio signal compression apparatus or an audio signal compression method and an audio signal compression apparatus which can perform compression while maintaining the same. In addition, the present invention provides a high-performance speech recognition device, in which the resolution is changed for each frequency, which is obtained by a linear prediction analysis method that incorporates the auditory sensitivity characteristic, which is a human auditory characteristic. The present invention relates to a speech recognition method and a speech recognition device that can obtain a higher recognition rate than in the past when performing recognition using an amount.

[0002]

2. Description of the Related Art A variety of audio signal compression methods of this type have been proposed in the past, and an example will be described below.

First, a time series of an input audio signal is represented by, for example, an MDCT (modified discrete cosine trans).
form: Modified discrete cosine transform) or FFT (Fast Fourier Transform) is converted into a frequency characteristic signal sequence for each fixed period (frame), and the input audio signal is further subjected to linear prediction analysis (LPC analysis) for each frame. , The LPC coefficient (linear predictive coeffi
cient; linear prediction coefficient and LSP coefficient (line spectrum)
pair coefficient) or PARCOR coefficient (pa
rtial auto-correlation coefficient) and the like, and an LPC spectrum envelope is obtained from these coefficients. Next, the calculated frequency characteristic signal sequence is
By dividing by C spectrum envelope and normalizing,
The frequency characteristics are flattened, and the power is normalized based on the maximum value or the average value of the power. In the following description, the output coefficient at the time when the power is normalized is also referred to as a residual signal. Further, the flattened residual signal is vector-quantized using the spectral envelope as a weight. An example of such an audio signal compression method is TwinVQ (Iwagami, Moriya, Miki: "Audio coding by frequency weighted interleaved vector quantization (TwinVQ)", Proceedings of the Acoustical Society of Japan, 1-P-1, pp. 339. -34
0, (1994)).

Next, a conventional example of the audio signal compression method will be described below. First, the time series of the input audio signal is obtained by linear prediction analysis (LPC analysis) for each frame.
By doing, LPC coefficient (linear prediction coefficient) and LSP
Coefficient (line spectrum pair coefficient) or P
An LPC spectrum envelope component such as an ARCOR coefficient (partial autocorrelation coefficient) is separated into a residual signal having a flattened frequency characteristic. Then, the LPC spectrum envelope component is scalar-quantized, and the flattened residual signal is quantized by a sound source codebook prepared in advance to be converted into a digital signal. As an example of such an audio signal compression method, CELP (MR Schroeder
and BS Atal: “Code-excited linear prediction (C
ELP) high quality speech atvery low rates ", Proc.
There is ICASSP-85 (March 1985).

A conventional example of a speech recognition method will be described below. In general, in a speech recognition device, a standard model for each phoneme or word is created in advance using reference speech data, and a feature amount corresponding to a spectral envelope is obtained from input speech. Is calculated, and a phoneme or word corresponding to the standard model having the highest similarity is found to perform speech recognition. The standard model in this case is
For example, the hidden Markov model (hidden Markov model;
HMM) and the time series of representative feature values themselves are used as standard models (Seiichi Nakagawa, “Speech Recognition by Stochastic Model”, edited by IEICE, pp. 18-20).

Conventionally, as a time series of feature amounts obtained from an input speech, the time series of the input speech is determined by, for example, a linear prediction analysis (LPC analysis) having a predetermined period length (frame).
LPC cepstrum coefficients obtained by converting the linear prediction coefficients (LPC coefficients) into cepstrum transforms (Kiyohiro Kano, Satoshi Nakamura, Shiro Ise, "Digital Signal Processing of Voice / Sound Information", Sho Kodo, p10-16)
Or the input voice is converted into a power spectrum for each length (frame) of a certain period by a DFT or a band-pass filter bank, and recognition is performed using cepstrum coefficients obtained by cepstrum conversion of the power spectrum. I have.

[0007]

In a conventional example of the audio signal compression method, a normalized residual signal is obtained by dividing a frequency characteristic signal sequence calculated by MDCT or FFT by an LPC spectrum envelope. On the other hand, in the conventional example of the audio signal compression method, the input audio signal is separated into the LPC spectrum envelope and the residual signal calculated by the linear prediction analysis, and the conventional audio signal compression method and the audio signal compression method are separated. This is the same as in the conventional example in that the spectral envelope component is removed from the input signal by ordinary linear prediction analysis, that is, the input signal is normalized (flattened) by the spectral envelope to obtain the residual signal. .
Therefore, if the performance of this linear prediction analysis is improved, or if the estimation accuracy of the spectrum envelope obtained by the linear prediction analysis can be increased, information can be compressed more efficiently than before, while maintaining high sound quality.

By the way, in the ordinary linear prediction analysis, the envelope is estimated with the same precision frequency resolution for any frequency band, so that the frequency resolution of a low frequency band that is important for hearing is increased, that is, In order to accurately obtain the spectrum envelope of a low frequency band, it is necessary to increase the order of analysis, and as a result, there is a problem that the amount of information increases. In addition, when the analysis order is increased, the resolution of a high frequency band, which is not so important for the sense of hearing, is unnecessarily increased, so that a spectrum envelope having a peak in a high frequency band may be calculated. In the end, there is a problem that the sound quality is deteriorated.

Also, as in the conventional audio signal compression method, when performing vector quantization, weighting at the time of quantization is performed based only on the spectral envelope. There is a problem that it is not possible to efficiently quantize using the auditory properties of the sound.

On the other hand, in the conventional example of the speech recognition method, for example, the LPC cepstrum coefficient obtained by the ordinary linear prediction analysis does not perform the linear prediction analysis method incorporating the auditory sensitivity characteristic which is a human auditory characteristic. Therefore, there is a possibility that sufficient recognition performance is not exhibited. It is generally known that, in the first place, human hearing tends to place importance on low-frequency components and less on high-frequency components than low frequencies. Therefore, this LPC cepstrum (ce
pstrum) LP obtained by performing mel conversion of coefficients
C-mel coefficient (by Kiyohiro Kano, Satoshi Nakamura, Shiro Ise, "
Digital signal processing of sound information ", Shokodo, p.39-4
Although there is a method of performing recognition using 0), in the first place LPC
The cepstrum coefficients themselves do not sufficiently take into account the characteristics of human hearing in linear predictive analysis. For this reason, low-frequency information that is auditory important is not sufficiently reflected in the mel-transformed LPC mel-cepstral coefficient.

[0011] The Mel scale is a scale obtained from the perceptual characteristics of human pitch, and the pitch is an amount that largely depends on the frequency, but it affects not only the frequency but also the intensity of the sound. It is also well known that 1000 Hz, 40
The pure sound of dB SPL is set to 1000 mel as the reference sound.
The sound perceived at double height or half height is measured by the magnitude measurement method etc., and it is 2000 mel and 500m respectively.
However, as described above, the LPC cepstrum coefficient itself does not sufficiently take into account the characteristics of human hearing in linear prediction analysis, so that the LPC cepstrum coefficient is essentially recognized even if mel conversion, that is, mel transformation is performed. No improvement in performance can be expected.

Further, in the ordinary linear prediction analysis, the spectral envelope is estimated with the same frequency resolution for any frequency band. Therefore, when trying to increase the frequency resolution of a low frequency band that is important for hearing, that is, When trying to accurately determine the spectral envelope of the lower frequency band,
It is necessary to increase the order of analysis, which results in a problem that the amount of features increases and the amount of processing required for recognition increases. In addition, when the analysis order is increased, the resolution of a high frequency band is unnecessarily increased, so that the high frequency band has an unnecessary feature, and the recognition performance is rather deteriorated.

There is also a method of performing speech recognition using a cepstrum coefficient or a mel cepstrum coefficient obtained from a DFT or a band pass filter bank as a feature amount. There is also a problem that the amount is much larger than that of the linear prediction analysis.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to improve the performance of linear predictive analysis, that is, to improve the performance of linear predictive analysis incorporating auditory sensitivity characteristics which are human auditory characteristics. (Hereinafter, referred to as mel linear prediction analysis method (MLPC analysis method)), and the resulting melded linear prediction coefficients (hereinafter, referred to as mel linear prediction coefficients) are used for speech recognition, or Melted PARCOR coefficients that can be determined from mel linear prediction coefficients by a known method similar to that for determining PARCOR coefficients from ordinary linear prediction coefficients (hereinafter referred to as mel PARCOR coefficients)
LSP coefficients that can be obtained from the mel linear prediction coefficients by a known method similar to that for obtaining LSP coefficients from ordinary linear prediction coefficients (hereinafter, referred to as coefficients).
The mel LSP coefficients are referred to as mel LSP coefficients, and mel LPC cepstrum coefficients obtained by cepstrum conversion of mel linear prediction coefficients are used for speech recognition, thereby making it possible to further improve recognition performance. It is a thing. Although it has been conventionally assumed that the use of this type of melted coefficient improves the compression performance of audio signals and audio signals and the performance of speech recognition,
In reality, the amount of calculation became enormous, and it was never used for actual use. The inventor of the present invention has conducted intensive studies in view of the current situation, and as a result, it was originally necessary to perform an infinite number of calculations to calculate this kind of coefficient, and if this was censored by a finite number of times, the calculation error was reduced. What has been involved is that, just by performing the desired number of calculations, it is possible to perform calculations equivalent to performing infinite calculations, and there is a completely new calculation that does not involve errors in this calculation. I found The present invention uses such a new operation to perform weighting on a frequency corresponding to the auditory sensitivity characteristic, which is a human auditory property, to improve the compression performance of an audio signal and a voice signal and to improve the voice recognition performance. An object of the present invention is to provide an audio signal compression method, an audio signal compression device, an audio signal compression method, an audio signal compression device, a speech recognition method, and a speech recognition device that can be improved.

That is, the present invention improves the performance of linear prediction analysis by obtaining a spectrum envelope based on frequency weighting corresponding to the auditory sensitivity characteristic, which is a human auditory characteristic, or by performing linear prediction analysis. An audio signal compression method, an audio signal compression apparatus or an audio signal compression method, and an audio signal compression apparatus capable of increasing the accuracy of estimating the obtained spectrum envelope and compressing it while maintaining high sound quality more efficiently than before. The purpose is to provide.

In addition, since the characteristic amount corresponding to the spectral envelope is obtained by the mel-linear predictive analysis based on the weighting on the frequency corresponding to the auditory sensitivity characteristic which is a human auditory characteristic, even a small characteristic amount is efficient. A speech recognition method and a speech recognition device capable of realizing high recognition performance with a smaller processing amount than before by using this feature amount for speech recognition. The purpose is to provide.

[0017]

In order to solve the above-mentioned problem, an audio signal compression method according to the present invention (claim 1) encodes an input audio signal,
And, in Odeo signal compression method of compressing the information amount, and the audio signal that is the input, which is the input
Corresponds to human auditory sensitivity characteristics for audio signals
Using the audio signal with the frequency axis expanded and contracted,
The autocorrelation function on the frequency axis
Find the mel linear prediction coefficient from the autocorrelation function
Whether the linear prediction coefficient itself is the spectral envelope,
Or obtain a spectral envelope from the mel linear prediction coefficient,
The input audio signal is smoothed for each frame using the spectrum envelope .

[0018]

Also, the audio signal compression method according to the present invention (claim 2) is a method for compressing an input audio signal.
An audio signal that encodes and compresses the amount of information
Signal compression method, the input audio signal
From the audio signal of a certain time length,
Multi-stage all-pass filter for audio signals with a long length
To obtain the filter output signal of each stage,
Audio signal and the filter output signal for each stage
The human sense of hearing is obtained by multiply-accumulate (Equation 1) with the signal
Mel frequency axis with expansion and contraction of frequency axis corresponding to degree characteristics
Calculate the autocorrelation function on the
Find the mel linear prediction coefficient from the function,
The coefficient itself can be used as the spectral envelope or the
Calculate the spectral envelope from the linear prediction coefficients
Using the torque envelope, the input audio signal is
This is to smooth each frame . However, (Equation 1) is

(Equation 1) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0020]

The audio signal compression method according to the present invention (claim 3) is the audio signal compression method according to claim 2 , wherein the all-pass filter is a first-order audio signal.
Filter.

[0022]

The audio signal compression method according to the present invention (claim 4) is described in claim 2 or claim 3.
Audio signal compression method.
Use Bark scale or Mel scale for filter coefficients
Weighting on frequencies corresponding to human auditory sensitivity characteristics
Is what you do .

[0024]

Further, the audio signal compression apparatus according to the present invention (claim 5) performs
In an audio signal compression device for performing encoding and compressing the amount of information, the input audio signal
Is converted to a frequency domain signal and output.
Means, the input audio signal, and the input audio signal.
Corresponding to the human auditory sensitivity characteristics
Using an audio signal that has undergone expansion and contraction of the frequency axis,
Find an autocorrelation function on the mel frequency axis,
The mel linear prediction coefficient obtained from the autocorrelation function
Vector envelope or the mel linear prediction coefficient
Spectrum envelope calculation means for obtaining a spectrum envelope from
Normalize the frequency domain signal with the spectral envelope
Normalizing means for obtaining a residual signal; and
Normalization based on the maximum or average value of
Power normalizing means for obtaining a residual signal;
Is quantized by the residual codebook, and the residual
And a vector quantization means for converting into a code .

[0026]

The audio signal compression apparatus according to the present invention (Claim 6) provides the audio signal compression apparatus according to Claim 5.
In the decompression device, human beings
Weighting on the frequency corresponding to the auditory sensitivity characteristics
Perceptual weighting means for outputting as perceptual weighting coefficients
Wherein said vector quantization means comprises:
The quantization of the normal residual signal is performed using a number .

[0028]

[0029]

[0030]

Further, the audio signal compression apparatus according to the present invention (Claim 7), in the audio signal compression apparatus according to claim 6, is the vector quantization means, a plurality of vertical
Consists of multiple vector quantization means connected to a column
Multiple quantizing means, wherein the multiple quantizing means is
Is at least one of the above-mentioned multiple quantizing means.
The vector quantization means uses the above weighting coefficient to calculate
This is to quantize the residual signal .

[0032]

Also, the audio signal compression apparatus according to the present invention (claim 8) is any one of claims 5 to 7
2. The audio signal compression device according to
The torque envelope calculating means calculates, from the input audio signal,
Cut out the audio signal of a certain time length, and
Long audio signal through a multi-stage all-pass filter.
Then, a filter output signal for each stage is obtained, and
Audio signal, and the filter output signal of each stage
Of the human auditory sensitivity by multiply-accumulate (Equation 2)
On the mel frequency axis that expands and contracts the frequency axis corresponding to the
Calculate the autocorrelation function and calculate the autocorrelation function on the mel frequency axis.
The mel linear prediction coefficient is obtained from the
The spectrum envelope itself, or
A spectral envelope is obtained from a linear prediction coefficient .
However, (Equation 2) is

(Equation 2) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0034]

The audio signal compression apparatus according to the present invention (claim 9) provides the audio signal compression apparatus according to claim 8
In the compression device, the all-pass filter is a first-order
Filter .

[0036]

Further, the audio signal compression apparatus according to the present invention (Claim 10), in the audio signal compression apparatus according to claim 8 or claim 9, said all-pass Fi
Bark scale or Mel scale for filter coefficients
Weighting on the frequency corresponding to the human auditory sensitivity characteristics
To perform

[0038]

Further, the audio signal according to the present invention (Claim 11)
The signal compression method encodes the input audio signal.
Audio signal compression method that compresses the amount of information
And the input voice signal and the input voice signal.
The frequency axis extension corresponding to the human auditory sensitivity characteristics
Using the compressed audio signal, the self
Find the correlation function and determine if it is an autocorrelation function on the mel frequency axis.
The mel linear prediction coefficient is obtained from
The spectral envelope or the mel-linear
Calculate the spectral envelope from the measured coefficients and calculate the spectral envelope
Is used to smooth the input audio signal .

[0040]

Further, the audio signal according to the present invention (Claim 12)
The signal compression method encodes the input audio signal.
Audio signal compression method that compresses the amount of information
From the input audio signal,
The signal is cut out, and the audio signal having the predetermined time length is
Pass the filter output signal of each stage through an all-pass filter.
Signal and the input audio signal and the
By the product sum (Equation 3) performed finitely with the filter output signal,
Expansion and contraction of frequency axis corresponding to human auditory sensitivity characteristics
Find an autocorrelation function on the mel frequency axis,
The mel linear prediction coefficient is obtained from the above autocorrelation function,
Whether the linear prediction coefficient itself is the spectral envelope, or
Or obtain a spectral envelope from the mel linear prediction coefficient,
Using the spectral envelope, the input speech signal is
It is to lubricate . However, (Equation 3) is

(Equation 3) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0042]

[0043]

[0044]

The voice signal according to the present invention (claim 13)
The signal compression method according to claim 12, wherein the all-pass filter is a first-order all-pass filter.
Filter.

[0046]

Further, the audio signal according to the present invention (Claim 14)
The signal compression method according to claim 12 or claim 13.
In the voice signal compression method, the filter of the all-pass filter is used.
Using the Bark scale or the Mel scale for the filter coefficient,
Weighting on the frequency corresponding to the auditory sensitivity characteristics between
Things .

[0048]

[0049] The audio signal according to the present invention (Claim 15)
The signal compression device encodes the input audio signal.
Audio signal compressor that compresses the amount of information
And the input voice signal and the input voice signal.
The frequency axis extension corresponding to the human auditory sensitivity characteristics
Using the compressed audio signal, the self
Find the correlation function, and from the autocorrelation function on the mel frequency axis
Express the resulting mel formation prediction coefficients as a spectral envelope
A feature value calculating means for converting the feature value into
Normalize the audio signal by inverse filtering with the above features
And an envelope normalizing means for obtaining a residual signal;
Normalize based on the maximum or average power and
Power normalizing means for obtaining a normalized residual signal;
Vector quantization of residual signal by residual codebook
And a vector quantizing means for converting into a residual code .

[0050]

Further, the audio signal compression apparatus according to the present invention (claim 16) is the same as the audio signal compression apparatus according to claim 15,
In the above, the feature amount calculating means may include the input audio signal.
The audio signal of a fixed time length is cut out from the
Pass long audio signals through multiple stages of all-pass filters
To obtain the filter output signal of each stage,
Finite difference between the audio signal and the filter output signal of each stage
Multiply-accumulate (Equation 4) to respond to human auditory sensitivity characteristics
On the Mel Frequency Axis with Expansion and Contraction of the Changing Frequency Axis
Function and calculate the function from the autocorrelation function on the mel frequency axis.
The linear prediction coefficient is calculated, and the mel linear prediction coefficient is
It is converted into a feature quantity expressing a vector envelope . However, (Equation 4 ) is

(Equation 4) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0052]

[0053]

[0054]

[0055] The audio signal compression apparatus according to the present invention (Claim 17), <br/> Oite to the audio signal compression apparatus of claim 16, said all-pass filters, first order all-pass
Filter .

[0056]

Further, the audio signal compression apparatus according to the present invention (claim 18) provides the sound signal compression apparatus according to claim 16 or claim 17.
In the voice signal compression device, the filter of the all-pass filter is used.
Using the Bark scale or the Mel scale for the filter coefficient,
Weighting on the frequency corresponding to the auditory sensitivity characteristic between the two .

[0058]

Further, the voice recognition according to the present invention (claim 19).
The recognition method is a sound that recognizes voice from the input voice signal.
In the voice recognition method, the input speech signal and the input
Respond to human auditory sensitivity characteristics
Using the audio signal with the frequency axis expanded and contracted
Find the autocorrelation function on the wave number axis, and
Find the mel linear prediction coefficient from the autocorrelation function,
Of features expressing spectral envelope from shape prediction coefficients
It is those that.

[0060]

[0061] The voice certification according to the present invention (Claim 20)
The recognition method is a sound that recognizes voice from the input voice signal.
In the voice recognition method, one of
The audio signal of a fixed time length is cut out and the audio signal of the fixed time length is cut out.
Signal through a multi-stage all-pass filter,
A filter output signal is obtained, and the input audio signal is
A finite number of product sums with the filter output signal of each stage
From equation (5), the frequency corresponding to the human auditory sensitivity characteristic
Calculate the autocorrelation function on the mel frequency axis that has been stretched.
Mel linear prediction from the autocorrelation function on the mel frequency axis
The coefficient is obtained, and the spectral envelope is obtained from the mel linear prediction coefficient.
This is for obtaining a feature to be expressed . Where (Equation 5) is

(Equation 5) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0062]

The voice recognition according to the present invention (claim 21).
The speech recognition method according to claim 20, wherein
The all-pass filter is a first-order all-pass filter .

[0064]

[0065]

[0066]

Further, the voice recognition method according to the present invention (claim 22) provides the voice recognition method according to claim 20 or claim 21.
In the recognition method, the filter
Use the Bark scale or Mel scale for the number, and
The weighting on the frequency corresponding to the sensitivity characteristic is performed .

[0068]

The voice recognition device according to the present invention (claim 23) provides a sound recognition device for recognizing voice from an input voice signal.
In the voice recognition device, the input speech signal and the input
Corresponds to human auditory sensitivity characteristics for the input audio signal
Using the audio signal with the frequency axis
Find the autocorrelation function on the frequency axis, and on the mel frequency axis
Mel linear prediction to find mel formation prediction coefficients from autocorrelation function
Cepstrum from the mel linear prediction coefficient
Cepstrum coefficient calculating means for calculating a coefficient;
Multiple frames of strum coefficients and multiple standard models
Is calculated, and the shortest distance is calculated as
Is recognized as having the highest similarity among standard models
Voice recognition means .

[0070]

[0071] The speech recognition apparatus according to the present invention (Claim 24), in the speech recognition apparatus according to claim 23,
The mel linear prediction analysis means is configured to output the input speech signal
From the audio signal of a fixed time length,
Through the multi-stage all-pass filter,
Obtain the filter output signal for each stage, and
Finite round trip of the signal and the filter output signal of each stage
The product sum (Equation 6) corresponds to the human auditory sensitivity characteristics
Autocorrelation function on mel frequency axis with expansion and contraction of frequency axis
From the autocorrelation function on the mel frequency axis.
This is for obtaining a shape prediction coefficient . Where (Equation 6) is

(Equation 6) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0072]

Further, the speech recognition apparatus according to the present invention (claim 25) provides the speech signal compression method according to claim 24.
Therefore, the all-pass filter is a first-order all-pass filter.
Ruta .

[0074]

Further, the voice recognition apparatus according to the present invention (claim 26) provides a voice pressure control device according to claim 24 or claim 25.
In the compression method, a filter
Use the Bark scale or Mel scale for the number, and
The weighting on the frequency corresponding to the sensitivity characteristic is performed .

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

(Embodiment 1) FIG. 1 is a block diagram showing a configuration of an audio signal compression apparatus according to a first embodiment of the present invention. In FIG. 1, reference numeral 1 denotes a time-frequency conversion unit that converts a time series of a digital audio signal or an audio signal input by, for example, MDCT or FFT into a frequency characteristic signal sequence for each fixed period length (frame). It is. In addition, 2 uses the mel linear prediction analysis that incorporates the frequency stretching function into the prediction model,
This is a spectrum envelope calculation unit that obtains, for each frame, a spectrum envelope in which the analysis accuracy is changed for each frequency from the input audio signal. Reference numeral 3 denotes a normalization unit that flattens the frequency characteristics by dividing the frequency characteristic signal sequence calculated by the time-frequency conversion unit 1 by the spectrum envelope obtained by the spectrum envelope calculation unit 2 and normalizing the result. For the frequency characteristic signal sequence flattened by the unit 3, the maximum value of the power,
Alternatively, it is a power normalizing unit that normalizes power based on an average value or the like. Reference numeral 5 denotes a multi-stage quantization unit that vector-quantizes the frequency characteristic signal sequence flattened by the normalization unit 3 and the power normalization unit 4. The multi-stage quantization unit 5 includes a first stage connected in cascade with each other. , A second-stage quantizer 52,..., An N-th stage quantizer 53. 6
Receives the frequency characteristic signal sequence output from the time-frequency conversion unit 1 and the spectrum envelope calculated by the spectrum envelope calculation unit 2, and based on human auditory sensitivity characteristics, performs quantization at the quantization unit 5. It is an auditory weighting calculator for determining a weighting coefficient to be used.

Next, the operation will be described. A time series of an input digital audio signal (hereinafter also referred to as an input signal) is converted into a frequency characteristic signal series by the MDCT, FFT, or the like in the time-frequency converter 1 for each fixed period length (frame).

Further, the spectrum envelope of the input signal is obtained for each frame by the spectrum envelope calculation unit 2 by using the mel-linear prediction analysis in which the frequency expansion and contraction is incorporated in the prediction model and the analysis accuracy is changed for each frequency. FIG. 2 is a diagram illustrating the spectrum envelope calculation unit 2 that obtains a spectrum envelope in which the analysis accuracy is changed for each frequency from the input signal using the mel linear prediction analysis. In the figure, a spectrum envelope calculation unit 2 uses a mel linear prediction analysis to change the analysis accuracy for each frequency, that is, a mel-coefficient calculation unit 21 that obtains a mel-formed linear prediction coefficient, and a And an envelope calculating unit 22 for calculating the spectral envelope of the linear frequency used for the calculation. Hereinafter, each of the melding coefficient calculation unit 21 and the envelope calculation unit 22 will be described.

First, an outline of the processing in the melting coefficient calculation section 21 is shown in FIG. In FIG. 3, reference numeral 211 denotes an all-pass filter that expands and contracts the frequency axis of an input signal, and 212 generates a linear combination of the output signal of the all-pass filter 211 and a prediction coefficient, and outputs a predicted value of the input signal of the all-pass filter 211. The linear combination unit 213 includes a prediction value output from the linear combination unit 212 and the all-pass filter 2.
A least-squares method operation unit that applies the least-squares method to the 11 output signals and outputs a melded linear prediction coefficient. next,
A method of estimating a linear prediction coefficient in which the analysis accuracy is changed for each frequency, that is, a melded linear prediction coefficient will be described with reference to FIG. First, the input signal x [n] is converted to an i-stage all-pass filter 211.

[0101]

(Equation 13)

The output signal yi [n] passed through and the prediction coefficient

[0103]

[Equation 14]

Predicted value of x [n] by linear combination with

[0105]

(Equation 15)

Is represented by (Equation 16).

[0107]

(Equation 16)

Here, [] indicates a sequence on the time axis.
Here, the all-pass filter (Equation 13) is represented by (Equation 17)
It is represented by The output signal yi [n] is described later (Equation 2).
1) and (Equation 29).

[0109]

[Equation 17]

Here, z represents a z-conversion operator.

The frequency characteristics of this all-pass filter are as follows:
As shown in FIG. In FIG. 5, the horizontal axis represents the frequency axis before conversion, and the vertical axis represents the frequency axis after conversion. In the figure, α = -0.5
The state when the value of α is changed from 0.1 to α = 0.8 in increments of 0.1 is displayed. From the figure, it can be seen that when the value of α is positive, the low frequency band is extended and the high frequency band is contracted. When the value of α is negative, the reverse is true.

In the present invention, since audio signals and audio signals having different sampling frequencies, that is, bandwidths, are assumed as input signals, the value of α is determined according to each sampling signal in accordance with the sampling frequency. ,
By obtaining a frequency resolution that matches human auditory characteristics when obtaining a spectral envelope, it is possible to obtain a frequency resolution that matches human auditory characteristics when obtaining a spectral envelope. For example, the Bark scale is generally known as a scale derived from observation of a critical bandwidth with respect to the frequency resolution of hearing, and it is also possible to determine the value of α based on this characteristic.

The Bark scale is a measure obtained from the concept of the auditory filter proposed by Fletcher. The auditory filter described by Fletcher is a band-pass filter whose center frequency changes continuously. A band filter having a frequency analyzes the frequency of the signal sound, and a noise component that affects the masking of the sound is a filter that is limited to the frequency component in the band filter. Fletcher names the bandwidth of this bandpass filter the critical band. The Mel scale is generally known as a psychological scale that directly quantifies pitch sensation based on human subjectivity, and the value of α can be determined based on this characteristic.

For example, when the Mel scale is used as a weighting on the frequency corresponding to the auditory sensitivity characteristics, we set α = 0.31 at a sampling frequency of 8 kHz, α = 0.35 at 10 kHz, α = 0.41 at 12 kHz, and α = 0.41 at 12 kHz. α = 0.45,
At 44.1 kHz, α was set to 0.6 to 0.7. When the Bark scale is used as a weight on a frequency corresponding to the auditory sensitivity characteristic, α may be appropriately changed from these values. For example, in the case of the Bark scale, at 12 kHz, we have adopted α = 0.51.

Next,

[0116]

(Equation 18)

The output signal of the all-pass filter represented by
The least-squares-method calculating unit 213 uses the least-squares method to minimize the total square error ε between yi [n] and the predicted value (Equation 15).

[0118]

[Equation 19]

Can be obtained. Here, p is the order of the prediction coefficient, and p may be set in advance by a preliminary experiment in consideration of the amount of signal compression calculation. When the input signal is a speech signal, for example, 8 to 14 For example, when the input signal is an audio signal, for example, 10 to 20
And so on. However,

[0120]

(Equation 20)

[0121]

(Equation 21)

Is as follows.

The mel-formed linear prediction coefficient for minimizing the total square error ε of (Equation 18) is given by the following normal equation.

[0124]

(Equation 22)

Where the coefficient

[0126]

(Equation 23)

Is an autocorrelation function (mel autocorrelation function) on the mel frequency axis (mel frequency domain), and is given by the following equation.

[0128]

(Equation 24)

Here, (Equation 23) is expressed by the spectrum on the linear frequency axis according to Parseval's theorem.

[0130]

(Equation 25)

And (Equation 26). Here, () represents a sequence in the frequency domain.

[0132]

(Equation 26)

Further, when (Equation 26) is rewritten into the form of an expression on the mel frequency axis,

[0134]

[Equation 27]

Is as follows. However,

[0136]

[Equation 28]

Is as follows. This (Equation 28) is obtained by Fourier-transforming the all-pass filter represented by (Equation 17). (Equation 27) is the mel autocorrelation function (Equation 2)
3) is equivalent to the inverse Fourier transform of the power spectrum on the mel frequency axis. Therefore, (Equation 2)
The coefficient matrix of 2) is a Toeplitz-type autocorrelation matrix, and it is possible to obtain a linear prediction coefficient that has been melted by a simple recurrence formula. Hereinafter, the procedure of the actual calculation for obtaining the melded linear prediction coefficient will be described, and the flow thereof is shown in FIG. (Step 1) In step S1, the input signal x [n
Is obtained, and in step S2, the signal is passed through an i-stage all-pass filter, whereby the output signal yi [n] obtained in step S3 is obtained by the following equation.

[0138]

(Equation 29)

Is as follows. However, it is (Equation 21). (Step 2) In step S4, the input signal x [n
] And the filter output signal yi [n] of each stage are calculated by the following equation, so that in step S5,
Obtain the autocorrelation function on the mel frequency axis. At this time, the mel autocorrelation function (Equation 23) is obtained by calculating the difference in the number of stages of the all-pass filter from the relationship of (Equation 27).

[0140]

[Equation 30]

Since it depends only on (Equation 31)
, The calculation can be performed by the product-sum operation of N terms, and there is no need to perform approximation by terminating the operation. Note that this (Equation 31) is obtained by transforming (Equation 24) using (Equation 21) and (Equation 29).

[0142]

(Equation 31)

That is, as can be seen from this (Equation 31),
This calculation, which should normally require an infinite number of calculations in the case of a normal calculation method shown in (Equation 24), is completed by a finite number of calculations, so that a huge amount of calculation is not required. In addition, there is no need to perform any approximation such as truncation of a waveform, which is necessary when the computation is terminated by a finite number of computations instead of performing an infinite number of computations, and no error occurs due to the truncation of the waveform. Moreover, the amount of calculation is about twice as large as the normal autocorrelation coefficient, so that it can be obtained directly from the waveform. This point is an important point that is decisively different from the conventional calculation method shown in (Equation 24).

(Step 3) In step S6, the normal equation of (Equation 22) is solved by a well-known algorithm, for example, the method of Durbin, using the mel autocorrelation function (Equation 23). A mel-formed linear prediction coefficient (mel linear prediction coefficient) is obtained.

Next, an outline of the envelope calculating section 22 is shown in FIG. In FIG. 6, reference numeral 221 denotes an inverse mel transform unit which performs inverse mel transform on the mel-formed linear prediction coefficient and outputs a linear frequency linear prediction coefficient, and 222 denotes a Fourier transform of the linear frequency linear prediction coefficient and outputs a spectrum envelope. It is an FFT unit. Next, using FIG. 6, a method of obtaining a spectral envelope of a linear frequency used for spectrum flattening from a linear prediction coefficient obtained by changing the analysis accuracy for each frequency, that is, a melted linear prediction coefficient (Equation 19). Will be described. First, in the inverse mel transform unit 221, from the linear prediction coefficient (Equation 19)

[0146]

(Equation 32)

By the inverse mel transform, the linear prediction coefficient of the linear frequency

[0148]

[Equation 33]

Is obtained. To actually solve (Equation 32),
It can be solved by calculating the well-known Oppenheim recurrence equation. Where the allpass filter

[0150]

(Equation 34)

In (Equation 17), it is necessary to use the all-pass filter of (Equation 35) in which α is replaced by −α.

[0152]

(Equation 35)

As a result, it is possible to obtain a prediction coefficient converted from a mel frequency to a linear frequency. Further, in the FFT section 222, the spectrum envelope S (ejα) of the linear frequency used for spectrum flattening can be obtained from the linear prediction coefficient of the linear frequency (Equation 33) using FFT (Equation 36). .

[0154]

[Equation 36]

Next, the normalizing section 3 flattens the frequency characteristic signal sequence by dividing the frequency characteristic signal sequence calculated above by the spectrum envelope and normalizing the divided frequency characteristic signal sequence. The power normalization unit 4 further normalizes the power of the frequency characteristic signal sequence flattened by the normalization unit 3 based on the maximum value or average value of the power.

In the audio signal compression, normalization is performed by the same spectral envelope as in the normalization unit 3. In other words, the time series of the input audio signal is obtained by performing linear prediction analysis (LPC analysis) for each frame to obtain L
PC coefficient (linear prediction coefficient) and LSP coefficient (line spectru
m pair coefficient) or LPC spectrum envelope component such as a PARCOR coefficient (partial autocorrelation coefficient) and a residual signal whose frequency characteristic is flattened. The processing is equivalent to the division on the frequency by the spectral envelope component, and the linear prediction coefficient or LS obtained by the linear prediction analysis.
This is equivalent to performing inverse filtering processing on the time axis using a spectral envelope component such as a P coefficient or a PARCOR coefficient. Therefore, the mellated linear prediction coefficients obtained from the input speech as in the present invention, or the PACs obtained from the melted linear prediction coefficients by a known method similar to the method for obtaining the PARCOR coefficient from the normal linear prediction coefficients, are obtained. By using the melded LSP coefficient obtained from the melded linear prediction coefficient by a known method similar to that for obtaining the LSP coefficient from the melded PARCOR coefficient or the normal linear prediction coefficient, It is possible to perform audio signal compression by performing an on-axis inverse filtering process, or by separating a spectral envelope component and a residual signal.

On the other hand, the frequency characteristic signal sequence output from the time-frequency conversion unit 1 and the spectrum envelope obtained by the spectrum envelope calculation unit 2 are input to the auditory weighting calculation unit 6. For the spectrum of the frequency characteristic signal sequence, based on the auditory sensitivity characteristics that are human auditory characteristics such as the minimum audibility characteristics and the auditory masking characteristics, a characteristic signal in consideration of the auditory sensitivity characteristics is calculated. A weighting coefficient used for quantization is obtained based on the characteristic signal and the spectral envelope.

The residual signal output from the power normalizing section 4 is quantized by the first-stage quantizing section 51 of the multi-stage quantizing section 5 using the weighting coefficient obtained by the auditory weighting calculating section 6. The quantization error component due to the quantization in the first-stage quantization unit 51 is converted to the second-stage quantization unit 5 of the multi-stage quantization unit 5.
In step 2, the quantization is performed using the weighting coefficient obtained by the auditory weighting calculation unit 6, and similarly, in each of the plurality of stages of quantization units, the quantization error component of the quantization in the preceding stage quantization unit is calculated. Quantization is performed. Each of these quantization units outputs a code as a quantization result. Then, the quantization error component by the quantization in the (N-1) th stage quantization unit is quantized by the Nth stage quantization unit 53 using the weighting coefficient obtained by the auditory weighting calculation unit 6. As a result, the compression encoding of the audio signal is completed.

As described above, according to the audio signal compression method and the audio signal compression apparatus according to the first embodiment, the frequency characteristic signal sequence calculated from the input audio signal by the normalization unit 3 is converted into a human auditory sense. Since the normalization is performed using the spectral envelope whose analysis accuracy is changed for each frequency according to the auditory sensitivity characteristic, which is the property, the frequency characteristic signal sequence can be accurately flattened and efficient quantization performed. be able to.

In addition, the burden on vector quantization by the multi-stage quantization unit 5 is reduced, and efficient quantization can be performed. In the vector quantization, a frequency characteristic signal sequence is represented by limited information (code). Therefore, the simpler the shape of the frequency characteristic signal sequence, the smaller the number of codes that can be represented. Thus, in the present invention, in order to simplify the shape of the frequency characteristic signal sequence, the frequency characteristic signal sequence is normalized using a spectral envelope expressing the schematic shape thereof. By using the spectral envelope with the changed accuracy, the shape of the frequency characteristic signal sequence can be simplified more accurately, and efficient quantization can be performed.

A plurality of stages of vector quantizers 51 to 53 of the multistage quantizer 5 provide an auditory weighting calculator 6 with which the spectrum of the input audio signal, the auditory sensitivity characteristic which is a human auditory characteristic, and the human auditory sensitivity characteristic are obtained. Vector quantization is performed by using a weighting coefficient on a frequency calculated based on a spectral envelope in which the analysis accuracy is changed for each frequency in accordance with the auditory sensitivity characteristic, which is an auditory property, as a weight at the time of quantization. With this configuration, efficient quantization can be performed by utilizing the auditory characteristics of humans.

It should be noted that the melding coefficient calculating section 21 is a section for obtaining, from the input signal, a linear prediction coefficient in which the analysis precision is changed for each frequency using mel linear prediction analysis, that is, a melted linear prediction coefficient. This may be obtained using the following method. That is, the input signal is subjected to expansion and contraction of the frequency axis using an all-pass filter to obtain a frequency expansion and contraction signal, and normal linear prediction analysis is performed on the frequency expansion and contraction signal to change the analysis accuracy for each frequency. This is a method for obtaining the spectral envelope. Hereinafter, a method of estimating a linear prediction coefficient whose analysis accuracy is changed for each frequency, that is, a melded linear prediction coefficient will be described. First, the input signal x [n] is

[0163]

(37)

Output signal obtained by converting the frequency axis to the mel frequency

[0165]

(38)

Is obtained. Where the allpass filter

[0167]

[Equation 39]

Is represented by (Equation 17).

Next, a linear prediction coefficient obtained by performing a normal linear prediction analysis on the output signal (Equation 38), that is, a linear prediction coefficient obtained by changing the analysis accuracy for each frequency

[0170]

(Equation 40)

Can be obtained. Actually (Equation 37)
Can be solved by calculating the well-known Oppenheim recurrence formula. Melting coefficient calculator 21
Then, a linear prediction coefficient obtained by changing the analysis accuracy for each frequency obtained by such a method may be used.

Further, the spectrum envelope calculation unit 2 obtains a frequency expanded / contracted signal by directly expanding / contracting the frequency axis from the input signal using an all-pass filter, thereby obtaining a spectrum envelope in which the analysis accuracy is changed for each frequency. In addition to the method, by resampling the power spectrum of the input signal on the frequency axis, that is, by performing interpolation processing, the power spectrum that has been expanded / contracted in the frequency axis, that is, mel-transformed, is obtained, and the inverse DFT is performed on the power spectrum. It is also possible to obtain a spectral envelope in which the analysis accuracy is changed for each frequency.

Further, the spectrum envelope calculator 2 calculates an autocorrelation function obtained by expanding and contracting the frequency axis of the autocorrelation function obtained from the input signal through an m-stage all-pass filter. It is also possible to obtain a spectral envelope in which is changed.

In the audio signal compression apparatus shown in FIG. 1, the auditory weighting calculator 6 uses a spectral envelope to calculate the weighting coefficient. The weighting coefficient may be calculated using only the sensitivity characteristics.

In the audio signal compression apparatus of FIG. 1, all of the vector quantizers of the multi-stage quantizer 5 use a weighting coefficient based on the auditory sensitivity characteristic obtained by the auditory weight calculator 6 to quantize the quantized signal. If any one of the vector quantizers of the multiple stages of the multi-stage quantization unit 5 performs quantization using a weighting coefficient based on the auditory sensitivity characteristic, such an auditory sensitivity is used. Efficient quantization can be performed as compared with a case where a weighting coefficient based on characteristics is not used. Further, in the audio signal compression apparatus shown in FIG. 1, the signal to be compressed is described as an audio band signal. However, this may be an audio band signal. In this case, the audio signal compression apparatus shown in FIG. It becomes a compression device. Further, in the audio signal compression apparatus of FIG. 1, the mel scale is used as the frequency weighting corresponding to the auditory sensitivity characteristic which is a human auditory characteristic, but the value of α of the all-pass filter is appropriately changed. Thus, the apparatus can be changed to an audio signal compression apparatus that performs signal compression based on the Bark scale without changing the block configuration in FIG.

(Embodiment 2) FIG. 7 is a block diagram showing a configuration of a speech recognition apparatus according to a second embodiment of the present invention. In the figure, reference numeral 7 denotes a mel linear prediction analysis unit that calculates, for each frame, a mel linear prediction coefficient in which the resolution is changed for each frequency from the input speech using mel linear prediction analysis incorporating frequency expansion and contraction in the prediction model. is there. Reference numeral 8 denotes a cepstrum coefficient calculation unit that converts the mel linear prediction coefficients calculated by the mel linear prediction analysis unit 7 into cepstrum coefficients. 9 calculates the similarity between the time series of the cepstrum coefficient calculated by the cepstrum coefficient calculator 8 and a plurality of standard models such as words and phonemes prepared in advance, and calculates the word or phoneme having the highest similarity. It is a voice recognition unit for recognition. The voice recognition unit 9 may perform specific speaker recognition or may perform unspecified speaker recognition.

Next, a detailed operation will be described. First,
The time series of the input digital voice (hereinafter, also referred to as “input signal”) is obtained by performing a mel linear prediction analysis in which frequency expansion and contraction is incorporated in the prediction model by the mel linear prediction analysis unit 7 for each fixed period length (frame). The mel linear prediction coefficient corresponding to the spectrum envelope of which the resolution is changed for each frequency is calculated using the mel linear prediction coefficient. Hereinafter, the operation of the mel linear prediction analysis unit 7 will be described.

[0178] First, FIG. 8 a schematic Mel linear predictive analysis section 7
Shown in Figure 8 linear prediction coefficients of varying resolution for each frequency by using, i.e. the method of calculating the mel-LPC coefficients will be described.

[0179]

[Equation 41]

Model replaced by

[0181]

(Equation 42)

Is used. However,

[0183]

[Equation 43]

Is a mel linear prediction coefficient, and α is an expansion / contraction coefficient for changing the resolution of the linear prediction analysis for each frequency. The frequency characteristics of the all-pass filter have already been shown in FIG. For example, as the expansion coefficient, the sampling frequency is α = 0.31 at 8 kHz, α = 0.35 at 10 kHz, and 12 kHz.
α = 0.41 at z, α = 0.45 at 16kHz, α at 44.1kHz
A value such as = 0.6 to 0.7 may be used. Here, the prediction error for a finite-length waveform x [n] (n = 0, ..., N-1) of length N is

[0185]

[Equation 44]

Evaluation is made based on the total squared prediction error over the infinite section as described above. At this time,

[0187]

[Equation 45]

And yi [n] is converted to the input signal x
Assuming that [n] is the output waveform that has passed through the i-stage all-pass filter, the predicted value of yi [n]

[0189]

[Equation 46]

Is represented by the following linear combination.

[0191]

[Equation 47]

Thus, the coefficient (Equation 43) that minimizes the prediction error is given by the following simultaneous equation.

[0193]

[Equation 48]

Here, φij is an infinite-length waveform yi [n] and yj
The covariance of [n], but Parseval's theorem and allpass filter

[0195]

[Equation 49]

By using the Fourier-transformed representation on the frequency axis, φij is given by a finite number of product-sum operations as in the following equation.

[0197]

[Equation 50]

Further,

[0199]

(Equation 51)

In other words, it can be shown that r [m] has a property as an autocorrelation function,

[0201]

(Equation 52)

The stability of the above is also guaranteed. (Equation 50)
As can be seen from this calculation, if the normal calculation method shown in the middle side of (Equation 50) should originally require an infinite number of calculations, the calculation is a finite number of times shown in the right side of (Equation 50). Since it ends with the calculation of, a huge amount of calculation is not required. Also,
There is no need to perform any approximation, such as truncation of the waveform, which is necessary when the computation is terminated by finite computations instead of performing infinite computations, and no error occurs due to the waveform truncation. In addition, the amount of calculation is several times the amount of a normal autocorrelation coefficient, and thus can be obtained directly from the waveform. This is an important point that is crucially different from the conventional calculation method.

FIG. 8 shows an actual calculation procedure for obtaining the mel linear prediction coefficient. This part is the first embodiment.
8 is the same as FIG. 3. In FIG. 8, reference numeral 71 denotes an all-pass filter that expands and contracts the frequency axis of the input signal, and 72 creates a linear combination of the output signal of the all-pass filter 71 and the prediction coefficient. A linear combination unit 73 that outputs a predicted value of the input signal, and a least square method operation unit 73 that applies the least square method to the predicted value output from the linear combination unit 72 and the input signal to output a melded linear prediction coefficient It is. Next, a method of estimating a linear prediction coefficient in which the analysis accuracy is changed for each frequency, that is, a melded linear prediction coefficient will be described with reference to FIG.

(Step 1) Output signal yi [n] obtained by passing input signal x [n] through i-stage all-pass filter 71 is
It is calculated by the following equation.

[0205]

(Equation 53)

The following is obtained. However, it is (Equation 21). (Step 2) In the linear combination unit 72, the input signal x
An autocorrelation function on the mel frequency axis is obtained by the product sum of [n] and the filter output signal yi [n] of each stage as in the following equation. At this time, the Mel autocorrelation function (Equation 23) is expressed by (Equation 2)
7) Difference in the number of stages of the all-pass filter

[0207]

(Equation 54)

Since it depends only on the above equation, it can be calculated by the product-sum operation of N terms without approximating truncation as in the following equation.

[0209]

[Equation 55]

(Step 3) The least squares method operation section 73 uses the mel autocorrelation function (Equation 23) to obtain (Equation 22)
Of the normal equation of a known algorithm, such as Du
A mel-formed linear prediction coefficient (mel linear prediction coefficient) is obtained by solving with the rbin method or the like.

From the mel linear prediction coefficient (Equation 43) obtained as described above, the cepstrum coefficient calculator 8 calculates
Convert to cepstrum coefficients. The method of conversion to cepstrum coefficients is already known, and is described in, for example, literatures (Kiyoshi Hiroshi,
This is described in detail in Satoshi Nakamura and Shiro Ise, "Digital Signal Processing of Speech and Sound Information", Shokodo, pp. 10-16, and converts mel linear prediction coefficients in the same way as ordinary linear prediction coefficients. Just do it. As a result, the cepstrum coefficient on the mel frequency axis can be obtained.

The time series of the cepstrum coefficients (hereinafter referred to as mel LPC cepstrum coefficients) calculated in this manner is obtained by calculating the similarity between a plurality of standard models such as words and phonemes prepared in advance in the speech recognition unit 9. Is calculated,
Recognize words and phonemes with the highest similarity.

As a standard model, there is a method called a Hidden Markov Model (HMM) that expresses a time series of feature amounts for each of a plurality of vocabularies to be recognized as stochastic transitions, and has already been widely used and known ( For example, Seiichi Nakagawa:
“Speech Recognition by Probabilistic Model,” edited by IEICE. The HMM is a method in which a time series of phonemes and word feature values due to individual differences is learned in advance by an HMM model, and how close the input voice is to the model as a probability value is recognized and recognized. In the present embodiment, the time series of the mel LPC cepstrum coefficients described above is used as the time series of the feature amount.

Further, as the standard model, a time series of a representative feature amount in a time series of feature amounts for each of a plurality of recognition target vocabularies may be used as a model. A normalized time series of feature amounts obtained by normalizing (expanding or contracting) in frequency may be used. For example, there is DP matching (dynamic programming) as a method of normalizing to an arbitrary length on the time axis, and normalizes a time series of temporal feature values according to a predetermined association rule. It is possible. In this embodiment, there is no problem in using the standard model in any case, since the time series of the mel LPC cepstrum coefficients described above may be used as the time series of the feature amount.

In the present embodiment, the recognition is performed using the mel LPC cepstrum coefficient as the time series of the feature amount obtained from the input speech. However, the same as the case of obtaining the PARCOR coefficient from the normal linear prediction coefficient. Mel PA that can be determined from the Mel linear prediction coefficient by a known method
LS from RCOR coefficient or ordinary linear prediction coefficient
It is also possible to use mel LSP coefficients, which can be obtained from mel linear prediction coefficients by a known method similar to that for obtaining the P coefficient, for speech recognition. The mel linear prediction coefficient obtained from the mel linear prediction coefficient, mel PARC
OR coefficient, Mel LSP coefficient, Mel LPC cepstrum coefficient, etc. are used in a wide range of fields such as speech synthesis as well as speech recognition, linear prediction coefficients, PARCOR coefficients, LSP coefficients, LPC coefficients obtained from conventional linear prediction analysis. It can be used in place of a cepstrum coefficient or the like.

In the present embodiment, the mel linear prediction analysis unit 7 converts a linear prediction coefficient whose resolution is changed for each frequency using the mel linear prediction analysis from an input signal, that is, a mel-formed linear prediction coefficient. The first was
May be obtained by using the same method as in the embodiment. That is, the frequency expansion / contraction signal is obtained by expanding / contracting the frequency axis of the input signal using an all-pass filter, and a normal linear prediction analysis is performed on the frequency expansion / contraction signal, thereby changing the spectrum for each frequency. This is a method to find the envelope.

As described above, by the mel-linear prediction analysis based on the weighting on the frequency corresponding to the auditory sensitivity characteristic, which is a human auditory characteristic, the spectral envelope in which the resolution is changed for each frequency in accordance with the auditory sensitivity characteristic. By obtaining the feature amount corresponding to, it is possible to efficiently capture the features of the spectral envelope even with a small amount of feature, and by using this feature amount for speech recognition, it is possible to achieve high recognition with less processing amount than before. Performance can be realized.

(Embodiment 3) FIG. 9 is a block diagram showing a configuration of an audio signal compression apparatus according to Embodiment 3 of the present invention. The audio signal compression device according to the present embodiment is an audio signal compression device mainly used in narrow-band signal compression of audio and the like.
In the figure, reference numeral 11 denotes a mel parameter for obtaining, for each frame, a mel linear prediction coefficient expressing a spectral envelope in which the analysis accuracy is changed for each frequency from an input audio signal by mel linear prediction analysis incorporating frequency expansion and contraction into a prediction model. It is a calculation unit. 12 is a mel parameter calculator 1
A parameter conversion unit that converts the mel linear prediction coefficient on the mel frequency axis obtained in the above into a feature quantity expressing a spectral envelope such as a linear prediction coefficient on a linear frequency axis. Reference numeral 13 denotes an envelope normalizing unit that calculates a residual signal by inverse filtering and normalizing the input audio signal with the feature amount obtained by the parameter converting unit 2, and 14 denotes a residual signal calculated by the envelope normalizing unit 13. A power normalizing unit that normalizes the power of the signal based on the maximum value or the average value of the power. Reference numeral 15 denotes a vector quantization unit that performs vector quantization on the normalized residual signal normalized by the power normalization unit 14 using the residual codebook 16 and converts the signal into a residual code.

Next, the operation will be described. The time series of a digital audio signal such as an input voice (hereinafter, also referred to as an input signal or an input voice) is obtained by incorporating frequency expansion / contraction into a prediction model by the mel parameter calculation unit 11 for each fixed period length (frame). The mel-linear predictive analysis determines a mel-linear predictive coefficient that expresses a spectral envelope in which the analysis accuracy is changed for each frequency from the input signal.
The part for calculating the mel linear prediction coefficient expressing the spectrum envelope is the same as the method described in the melding coefficient calculation unit 21 of the first embodiment, and calculating the feature amount expressing the spectrum envelope in the same procedure. Can be.

Next, the parameter conversion unit 12 converts the mel linear prediction coefficient on the mel frequency axis calculated by the mel parameter calculation unit 11 into a feature quantity expressing a spectral envelope such as a linear prediction coefficient on the linear frequency axis. . This part also
This is the same as the method described in the first embodiment, and can be realized by a method similar to that of the envelope calculation unit 22. By the way, mainly in the compression of the audio signal, the time series of the input audio signal is obtained by performing a linear prediction analysis (LPC analysis) for each frame.
LPC coefficients (linear prediction coefficients) and LSP coefficients (line spect
A characteristic amount representing an LPC spectrum envelope component such as a rum pair coefficient or a PARCOR coefficient (partial autocorrelation coefficient) is obtained, and a residual signal is calculated by inverse filtering and normalizing the characteristic amount. Therefore, a mel-formed linear prediction coefficient obtained from the input speech as in the present embodiment is used as a feature amount for normalization,
Alternatively, a melded PARCOR coefficient obtained from a mellated linear prediction coefficient by a known method similar to a method for obtaining a PARCOR coefficient from a normal linear prediction coefficient, or an LSP coefficient obtained from a normal linear prediction coefficient. Using the melded LSP coefficients obtained from the melformed linear prediction coefficients by a similar known method, inverse filtering on the time axis, or separation into a spectral envelope component and a residual signal, Thus, more accurate normalization and separation can be performed.

Similarly, the envelope normalizing section 13 of the present embodiment
In the above, inverse filtering is performed using a feature amount expressing a spectral envelope such as a linear prediction coefficient of a linear frequency axis converted by the parameter converting unit 12 to normalize a spectral envelope component, thereby calculating a residual signal. . Further, the power normalizing section 14 normalizes the power of the residual signal obtained by the envelope normalizing section 3 based on the maximum value or average value of the power. Then, in the vector quantization unit 15, the residual signal output from the power normalization unit 14 is vector-quantized using a residual codebook 16 obtained in advance. As a result, the vector quantization unit 15
Outputs the code as the quantization result, thereby completing the compression encoding of the input signal.

As described above, according to the audio signal compression method and the audio signal compression apparatus according to the present embodiment, the mel parameter calculation unit 1 converts the frequency characteristic signal sequence calculated from the input audio signal into a human auditory sense. A mel linear prediction coefficient expressing a spectrum envelope in which the analysis accuracy is changed for each frequency according to the auditory sensitivity characteristic which is a property is obtained, and the mel linear prediction coefficient is converted into a linear prediction coefficient of a linear frequency axis by the parameter conversion unit 2. Is converted into a feature quantity expressing the spectral envelope of
Since the residual signal is normalized by inverse filtering and normalization with the feature amount obtained by the parameter conversion unit 2, the frequency characteristic signal sequence can be accurately flattened, and the quantum efficiency can be improved efficiently. Can be performed. Also,
In vector quantization, a residual signal is represented by limited information (code). Therefore, the simpler the shape of the residual signal, the smaller the number of codes that can be represented.
Therefore, in the present invention, in order to simplify the shape of the residual signal, by using a spectral envelope in which the analysis accuracy is changed for each frequency, it is possible to more accurately simplify the shape of the residual signal, Efficient quantization can be performed.

(Embodiment 4) FIG. 10 is a block diagram showing a configuration of a mobile phone according to a fourth embodiment of the present invention. The mobile phone according to the present embodiment is similar to the mobile phone according to the third embodiment.
In the above description, the signal compression is performed by using an audio signal compression device mainly used in compression of narrow band signals such as audio. In the figure, reference numeral 11 denotes a mel parameter for obtaining, for each frame, a mel linear prediction coefficient expressing a spectral envelope in which the analysis accuracy is changed for each frequency from an input audio signal by mel linear prediction analysis incorporating frequency expansion and contraction into a prediction model. It is a calculation unit. Reference numeral 12 denotes a parameter conversion unit that converts the mel linear prediction coefficient on the mel frequency axis obtained by the mel parameter calculation unit 1 into a feature quantity expressing a spectral envelope such as a linear prediction coefficient on a linear frequency axis. Reference numeral 13 denotes an envelope normalizing unit that calculates a residual signal by inverse filtering and normalizing the input audio signal with the feature amount obtained by the parameter converting unit 2, and 14 denotes a residual signal calculated by the envelope normalizing unit 13. A power normalizing unit that normalizes the power of the signal based on the maximum value or the average value of the power. Fifteen
Is a vector quantization unit that performs vector quantization on the normalized residual signal normalized by the power normalization unit 14 using the residual codebook 16 and converts it into a residual code. Reference numeral 10 denotes a mel parameter calculator 11, a parameter converter 12,
It comprises an envelope normalizing unit 13, a power normalizing unit 14, a vector quantizing unit 15, and a residual codebook 16, and adapts an input audio signal input from a microphone or the like to an auditory sensitivity characteristic which is a human auditory characteristic. This is a voice compression unit that compresses information based on the weighted frequency. 31
Represents a code whose information has been compressed by the audio compression unit 10,
A transmitting unit 32 modulates and transmits a high-frequency signal of a frequency and a modulation method according to the specifications of the mobile phone.
This is an antenna that transmits a high-frequency signal from the antenna.

Next, the operation will be described. Audio compression unit 1
The operation of 0 is the same as that of the audio signal compression device according to the third embodiment. That is, the time series of a digital audio signal such as an input voice (hereinafter also referred to as an input signal or an input voice) is subjected to frequency expansion and contraction by the mel parameter calculation unit 11 for each predetermined period length (frame). Is obtained from the input signal, a mel linear prediction coefficient expressing a spectrum envelope in which the analysis accuracy is changed for each frequency. The part for calculating the mel linear prediction coefficient expressing the spectrum envelope is the same as the method described in the melding coefficient calculation unit 21 of the first embodiment, and calculating the feature amount expressing the spectrum envelope in the same procedure. Can be.

Next, the parameter conversion unit 12 converts the mel linear prediction coefficient on the mel frequency axis calculated by the mel parameter calculation unit 11 into a feature quantity expressing a spectral envelope such as a linear prediction coefficient on the linear frequency axis. . This part also
This is the same as the method described in the first embodiment, and can be realized by a method similar to that of the envelope calculation unit 22. By the way, mainly in the compression of the audio signal, the time series of the input audio signal is obtained by performing a linear prediction analysis (LPC analysis) for each frame.
LPC coefficients (linear prediction coefficients) and LSP coefficients (line spect
A characteristic amount representing an LPC spectrum envelope component such as a rum pair coefficient or a PARCOR coefficient (partial autocorrelation coefficient) is obtained, and a residual signal is calculated by inverse filtering and normalizing the characteristic amount. Therefore, a mel-formed linear prediction coefficient obtained from the input speech as in the present embodiment is used as a feature amount for normalization,
Alternatively, a melded PARCOR coefficient obtained from a mellated linear prediction coefficient by a known method similar to a method for obtaining a PARCOR coefficient from a normal linear prediction coefficient, or an LSP coefficient obtained from a normal linear prediction coefficient. Using the melded LSP coefficients obtained from the melformed linear prediction coefficients by a similar known method, inverse filtering on the time axis, or separation into a spectral envelope component and a residual signal, Thus, more accurate normalization and separation can be performed.

Similarly, the envelope normalizing section 13 of the present embodiment
In the above, inverse filtering is performed using a feature amount expressing a spectral envelope such as a linear prediction coefficient of a linear frequency axis converted by the parameter converting unit 12 to normalize a spectral envelope component, thereby calculating a residual signal. . Further, the power normalizing section 14 normalizes the power of the residual signal obtained by the envelope normalizing section 3 based on the maximum value or average value of the power. Then, in the vector quantization unit 15, the residual signal output from the power normalization unit 14 is vector-quantized using a residual codebook 16 obtained in advance. As a result, the vector quantization unit 15
Outputs a code as a quantization result, thereby completing the compression encoding of the audio signal. Then, the code of the audio signal that has been compression-encoded in the audio compression unit 10 is input to the transmission unit 31, and the transmission unit 31
The frequency is converted into a frequency in accordance with the specification adopted by the mobile phone and a high frequency of a modulation method, and transmitted to the base station via the antenna 32.

As described above, according to the mobile phone of the present embodiment, mel parameter calculating section 1 converts the frequency characteristic signal sequence calculated from the input audio signal according to the auditory sensitivity characteristic which is a human auditory characteristic. To obtain a mel linear prediction coefficient representing a spectral envelope in which the analysis precision is changed for each frequency, and the parameter conversion unit 2 converts the mel linear prediction coefficient into a feature quantity representing a spectral envelope such as a linear prediction coefficient on a linear frequency axis. , And the envelope normalization unit 3 normalizes the residual signal by performing inverse filtering and normalization with the feature amount obtained by the parameter conversion unit 2, so that the frequency characteristic signal can be accurately calculated. The sequence can be flattened, and efficient quantization can be performed.
In vector quantization, a residual signal is represented by limited information (code). Therefore, the simpler the shape of the residual signal, the smaller the number of codes. Therefore, in the present invention, in order to simplify the shape of the residual signal, by using a spectral envelope in which the analysis accuracy is changed for each frequency, it is possible to more accurately simplify the shape of the residual signal, Efficient quantization can be performed. So if you use the same band,
The communication quality can be improved as compared with the conventional one, and the number of channels can be further increased if the communication quality is the same as the conventional one. Note that the present embodiment can be applied to mobile communication such as an automobile telephone in addition to a mobile telephone.

(Embodiment 5) FIG. 11 is a block diagram showing a configuration of a network device according to a fifth embodiment of the present invention. The network device according to the present embodiment
An Internet telephone which performs signal compression using an audio signal compression device mainly used in narrow-band signal compression of audio and the like in the third embodiment, and sends this to another network device via a network such as the Internet. Etc. are assumed. In the figure, 1
Reference numeral 1 denotes a mel parameter calculation unit that obtains, for each frame, a mel linear prediction coefficient that expresses a spectral envelope in which the analysis accuracy is changed for each frequency from an input audio signal by mel linear prediction analysis incorporating frequency expansion and contraction into a prediction model. . Reference numeral 12 denotes a parameter conversion unit that converts the mel linear prediction coefficient on the mel frequency axis obtained by the mel parameter calculation unit 1 into a feature quantity expressing a spectral envelope such as a linear prediction coefficient on a linear frequency axis. Reference numeral 13 denotes an envelope normalizing unit that calculates a residual signal by inverse filtering and normalizing the input audio signal with the feature amount obtained by the parameter converting unit 2, and 14 denotes a residual signal calculated by the envelope normalizing unit 13. A power normalizing unit that normalizes the power of the signal based on the maximum value or the average value of the power. Reference numeral 15 denotes a vector quantization unit that performs vector quantization on the normalized residual signal normalized by the power normalization unit 14 using the residual codebook 16 and converts the signal into a residual code. Reference numeral 10 denotes a mel parameter calculation unit 11, a parameter conversion unit 12, an envelope normalization unit 13, a power normalization unit 14, and a vector quantization unit 15.
And a residual codebook 16, which is an audio compression unit for compressing information on an input audio signal input from a microphone or the like based on weighting on a frequency corresponding to the auditory sensitivity characteristic which is a human auditory characteristic. 40 converts the code compressed by the voice compression unit 10 into a code for transmitting voice data over a network,
This is a network interface unit that transmits data in accordance with a protocol such as an IP protocol according to network specifications.

Next, the operation will be described. Audio compression unit 1
The operation of 0 is the same as that of the audio signal compression device according to the third embodiment. That is, the time series of a digital audio signal such as an input voice (hereinafter, also referred to as an input signal) is
For each fixed period length (frame), the mel parameter calculation unit 11 expresses a spectral envelope in which the analysis accuracy is changed for each frequency from an input audio signal by mel linear prediction analysis incorporating frequency expansion and contraction into a prediction model. A mel linear prediction coefficient is determined. The part for calculating the mel linear prediction coefficient expressing the spectrum envelope is the same as the method described in the melding coefficient calculation unit 21 of the first embodiment, and calculating the feature amount expressing the spectrum envelope in the same procedure. Can be.

Next, the parameter conversion unit 12 converts the mel linear prediction coefficient on the mel frequency axis calculated by the mel parameter calculation unit 11 into a feature quantity expressing a spectral envelope such as a linear prediction coefficient on the linear frequency axis. . This part also
This is the same as the method described in the first embodiment, and can be realized by a method similar to that of the envelope calculation unit 22. By the way, mainly in the compression of the audio signal, the time series of the input audio signal is obtained by performing a linear prediction analysis (LPC analysis) for each frame.
LPC coefficients (linear prediction coefficients) and LSP coefficients (line spect
A characteristic amount representing an LPC spectrum envelope component such as a rum pair coefficient or a PARCOR coefficient (partial autocorrelation coefficient) is obtained, and a residual signal is calculated by inverse filtering and normalizing the characteristic amount. Therefore, a mel-formed linear prediction coefficient obtained from the input speech as in the present embodiment is used as a feature amount for normalization,
Alternatively, a melded PARCOR coefficient obtained from a mellated linear prediction coefficient by a known method similar to a method for obtaining a PARCOR coefficient from a normal linear prediction coefficient, or an LSP coefficient obtained from a normal linear prediction coefficient. Using the melded LSP coefficients obtained from the melformed linear prediction coefficients by a similar known method, inverse filtering on the time axis, or separation into a spectral envelope component and a residual signal, Thus, more accurate normalization and separation can be performed.

Similarly, the envelope normalization unit 13 of the present embodiment
In the above, inverse filtering is performed using a feature amount expressing a spectral envelope such as a linear prediction coefficient of a linear frequency axis converted by the parameter converting unit 12 to normalize a spectral envelope component, thereby calculating a residual signal. . Further, the power normalizing section 14 normalizes the power of the residual signal obtained by the envelope normalizing section 13 based on the maximum value or average value of the power. Then, in the vector quantization unit 15, the residual signal output from the power normalization unit 14 is vector-quantized using a residual codebook 16 obtained in advance. As a result, the vector quantization unit 15
Outputs a code as a quantization result, thereby completing the compression encoding of the audio signal. Then, the code of the audio signal compressed and encoded in the audio compression unit 10 is input to the network interface unit 40,
In the network interface unit 40, the code information compressed by the audio compression unit 10 is converted into a code for transmitting audio data over a network,
The packet is transmitted to the network according to a protocol such as an IP protocol according to the specifications of the network.

As described above, according to the network equipment of the present embodiment, mel parameter calculating section 11 converts the frequency characteristic signal sequence calculated from the input audio signal according to the auditory sensitivity characteristic which is a human auditory characteristic. To obtain a mel linear prediction coefficient expressing a spectrum envelope in which the analysis accuracy is changed for each frequency,
Then, this mel linear prediction coefficient is converted into a feature quantity expressing a spectral envelope such as a linear prediction coefficient on a linear frequency axis,
Further, the envelope normalization unit 13 normalizes the residual signal by performing inverse filtering and normalization using the feature amount obtained by the parameter conversion unit 12, so that the frequency characteristic signal sequence can be accurately flattened. This makes it possible to perform efficient quantization. In vector quantization, a residual signal is represented by limited information (code). Therefore, the simpler the shape of the residual signal, the smaller the number of codes. Therefore, in the present invention, in order to simplify the shape of the residual signal, by using a spectral envelope in which the analysis accuracy is changed for each frequency, it is possible to more accurately simplify the shape of the residual signal, Efficient quantization can be performed. For this reason, if the data transfer speed of the network is the same, the call quality can be improved as compared with the conventional one, and if the call quality is equivalent to the conventional one, the number of terminals that can be accommodated is increased. It becomes possible. Note that the present embodiment assumes Internet devices such as a personal computer, an Internet telephone, and an Internet TV, but can also be applied to terminals using protocols other than the Internet, such as personal computer communication.

(Embodiment 6) FIG. 12 is a block diagram showing a configuration of a network device according to a sixth embodiment of the present invention. The network device according to the present embodiment
In the first embodiment, an Internet device or the like that performs signal compression using an audio signal compression device mainly used in signal compression of an audio band and sends this to another network device via a network such as the Internet. It is assumed. In FIG. 1, reference numeral 1 denotes a time-frequency conversion unit that converts a time series of a digital audio signal or an audio signal input by, for example, MDCT or FFT into a frequency characteristic signal sequence for each fixed period length (frame). It is. Also, 2
This is a spectrum envelope calculation unit that obtains, for each frame, a spectrum envelope in which the analysis accuracy is changed for each frequency from the input audio signal using mel linear prediction analysis in which a frequency expansion function is incorporated in the prediction model. Reference numeral 3 denotes a normalization unit that flattens the frequency characteristics by dividing the frequency characteristic signal sequence calculated by the time-frequency conversion unit 1 by the spectrum envelope obtained by the spectrum envelope calculation unit 2 and normalizing the result. A power normalizing unit that normalizes the power of the frequency characteristic signal sequence flattened by the unit 3 based on the maximum value or the average value of the power. Reference numeral 5 denotes a multi-stage quantization unit that vector-quantizes the frequency characteristic signal sequence flattened by the normalization unit 3 and the power normalization unit 4. The multi-stage quantization unit 5 includes a first stage connected in cascade with each other. , A second-stage quantizer 52,..., An N-th stage quantizer 53. 6 receives the frequency characteristic signal sequence output from the time-frequency conversion unit 1 and the spectrum envelope obtained by the spectrum envelope calculation unit 2 as inputs, and performs quantization at the quantization unit 5 based on human auditory sensitivity characteristics. This is an auditory weighting calculation unit that obtains a weighting coefficient used for. Reference numeral 20 includes a time-frequency conversion unit 1, a spectrum envelope calculation unit 2, a normalization unit 3, a power normalization unit 4, a quantization unit 5, and an auditory weighting calculation unit 6. The audio signal compression unit compresses information based on frequency weighting corresponding to the auditory sensitivity characteristic that is a human auditory characteristic. 41 converts the code information compressed by the audio signal compression unit 20 into a code for transmitting audio data over a network,
A network interface unit that transmits data according to a protocol such as a P / IP protocol according to network specifications.

Next, the operation will be described. The operation of the audio signal compression unit 20 is the same as that of the audio signal compression device according to the first embodiment. That is, a time series of an input digital audio signal (hereinafter also referred to as an input signal) is converted into a frequency characteristic signal series by the MDCT, FFT, or the like by the time-frequency conversion unit 1 for each fixed period length (frame). .

Further, for the input signal, the spectrum envelope in which the analysis accuracy is changed for each frequency is obtained by the spectrum envelope calculation unit 2 for each frame by using the mel-linear prediction analysis in which frequency expansion and contraction are incorporated in the prediction model. Next, the normalizing unit 3 flattens the frequency characteristic signal sequence by dividing the frequency characteristic signal sequence calculated above by the spectrum envelope and normalizing the divided frequency characteristic signal sequence. The power normalization unit 4 further normalizes the power of the frequency characteristic signal sequence flattened by the normalization unit 3 based on the maximum value or average value of the power. On the other hand, the perceptual weighting calculator 6 receives the frequency characteristic signal sequence output from the time-frequency converter 1 and the spectrum envelope obtained by the spectrum envelope calculator 2, and outputs the frequency output from the time-frequency converter 1. For the spectrum of the characteristic signal series, based on the auditory sensitivity characteristics which are human auditory characteristics such as the minimum audibility characteristic and the auditory masking characteristic, calculate a characteristic signal in consideration of this auditory sensitivity characteristic, and further calculate this characteristic signal and A weighting coefficient used for quantization is obtained based on the spectral envelope.

The residual signal output from the power normalizing section 4 is quantized by the first-stage quantizing section 51 of the multi-stage quantizing section 5 using the weighting coefficient obtained by the auditory weighting calculating section 6. The quantization error component due to the quantization in the first-stage quantization unit 51 is converted to the second-stage quantization unit 5 of the multi-stage quantization unit 5.
In step 2, the quantization is performed using the weighting coefficient obtained by the auditory weighting calculation unit 6, and similarly, in each of the plurality of stages of quantization units, the quantization error component of the quantization in the preceding stage quantization unit is calculated. Quantization is performed. Each of these quantization units outputs a code as a quantization result. Then, the quantization error component by the quantization in the (N-1) th stage quantization unit is quantized by the Nth stage quantization unit 53 using the weighting coefficient obtained by the auditory weighting calculation unit 6. As a result, the compression encoding of the audio signal is completed. Then, the code of the audio signal compressed and encoded in the audio signal compression unit 20 is input to the network interface unit 41,
In the network interface unit 40, the code information compressed by the audio signal compression unit 20 is converted into a code for transmitting audio data in a network, and the network is transmitted in accordance with a protocol such as a TCP / IP protocol according to the network specification. Send to.

As described above, according to the network device of the sixth embodiment, the frequency characteristic signal sequence calculated from the input audio signal by the normalizing unit 3 is converted into the auditory sensitivity characteristic which is a human auditory characteristic. Since it is configured to normalize using the spectral envelope with the analysis accuracy changed for each frequency, the frequency characteristic signal sequence can be accurately flattened,
Efficient quantization can be performed. In addition, the burden when performing vector quantization in the multi-stage quantization unit 5 is reduced,
Efficient quantization can be performed. In the vector quantization, a frequency characteristic signal sequence is represented by limited information (code). Therefore, the simpler the shape of the frequency characteristic signal sequence, the smaller the number of codes that can be represented. Thus, in the present invention, in order to simplify the shape of the frequency characteristic signal sequence, the frequency characteristic signal sequence is normalized using a spectral envelope expressing the schematic shape thereof. By using the spectral envelope with the changed accuracy, the shape of the frequency characteristic signal sequence can be simplified more accurately, and efficient quantization can be performed.

In the multiple-stage vector quantizers 51 to 53 of the multi-stage quantizer 5, the auditory weighting calculator 6 calculates the spectrum of the input audio signal, the auditory sensitivity characteristics which are human auditory characteristics, and the human auditory sensitivity characteristics. Vector quantization is performed by using a weighting coefficient on a frequency calculated based on a spectral envelope in which the analysis accuracy is changed for each frequency in accordance with the auditory sensitivity characteristic, which is an auditory property, as a weight at the time of quantization. With this configuration, efficient quantization can be performed by utilizing the auditory characteristics of humans. As described above, since the audio signal is efficiently quantized, if the data transfer rate of the network is the same, the audio quality can be improved as compared with the conventional one, and the same audio quality as the conventional one can be obtained. If it is good, it is possible to increase the number of terminals that can be accommodated. In addition,
In this embodiment, Internet devices such as a personal computer and an Internet TV are assumed.
The present invention can be applied to a terminal using a protocol other than the Internet.

[0239]

As described above, according to the audio signal compression method according to the present invention (claim 1), an audio signal for encoding an input audio signal and compressing the information amount is encoded. In the compression method, the above input
Audio signal and the input audio signal
Expands and contracts the frequency axis corresponding to human auditory sensitivity characteristics
Self-phase on the Mel frequency axis
Calculate the function and calculate from the autocorrelation function on the mel frequency axis.
The mel linear prediction coefficient is obtained, and the mel linear prediction coefficient
Is the spectral envelope or the mel linear prediction
Calculate the spectral envelope from the coefficients and calculate the spectral envelope
In this case, the input audio signal is smoothed for each frame, so that an audio signal compression method capable of performing efficient signal compression utilizing human auditory characteristics is obtained. There is.

Further, according to the audio signal compression method according to the present invention (claim 2), the input audio signal
On the other hand, perform encoding and compress the amount of information.
In the video signal compression method, the input audio
From the signal, cut out the audio signal of a certain time length,
An audio signal with a fixed length of time can be
Filter to determine the filter output signal for each stage,
The input audio signal and the filter for each stage
By the product sum (Equation 1) performed finitely with the output signal, human
Mel circumference with frequency axis expansion / contraction corresponding to hearing sensitivity characteristics
Find the autocorrelation function on the wave number axis, and
Find the mel linear prediction coefficient from the autocorrelation function,
The shape prediction coefficient itself as the spectral envelope, or
Finds the spectral envelope from the mel linear prediction coefficient,
Using the spectral envelope, the input audio signal
Since the signal is smoothed for each frame, there is an effect that an audio signal compression method capable of performing efficient signal compression using human auditory characteristics is obtained. In addition, Mel linear which originally required infinite number of operations
Calculation of prediction coefficients does not require any approximation calculation.
Since it is obtained by a finite number of calculations set in advance,
Improves compression performance of input audio signals and recognizes them
Performance can be improved. However, (Equation 1) is

(Equation 1) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

According to the audio signal compression method according to the present invention (claim 3), in the audio signal compression method according to claim 2, the all-pass filter has one
Since the next all-pass filter is used,
What required a limited number of operations is now feasible
By using a first-order all-pass filter, approximate calculation can be performed.
Preset finite number of operations without any need
This means that efficient signal compression can be performed.
Wear.

Further, according to the audio signal compression method according to the present invention (claim 4), the audio signal compression method according to claim 2 or 3 is described.
In the audio signal compression method described above,
Bark scale or Mel scale is used for the filter coefficient of the filter.
Frequency, the frequency weight corresponding to the human auditory sensitivity
The Bark scale or Mel
Using a scale, low frequency bands that are important to human hearing
Analysis with higher frequency resolution than the higher frequency band
This makes it possible to obtain an audio signal compression method capable of performing efficient signal compression by utilizing human auditory characteristics.

[0243] Further, according to the audio signal compression apparatus according to the present invention (Claim 5), the input audio signals, performs encoding, and, in the audio signal compression apparatus for compressing the information amount, the The entered audio
E time signal that converts the signal into a frequency domain signal and outputs it
Number conversion means, the input audio signal,
For human auditory sensitivity characteristics to the input audio signal
Use the audio signal that has expanded and contracted the corresponding frequency axis.
The autocorrelation function on the mel frequency axis
Mel linear prediction coefficient obtained from autocorrelation function on wavenumber axis
Is the spectral envelope, or the mel linear prediction
A spectral envelope calculator that calculates the spectral envelope from the coefficients
And normalizing the frequency domain signal with the spectral envelope
Normalizing means for obtaining a residual signal;
Normalize based on the maximum or average power and
A power normalizing means for obtaining a normalized residual signal;
Vector quantization of the difference signal by the residual codebook,
Vector quantization means for converting to a residual code.
Therefore, there is an effect that an audio signal compression device capable of efficiently compressing a signal by utilizing the auditory characteristics of human beings is obtained.

According to the audio signal compression apparatus of the present invention (claim 6), the audio signal compression apparatus of claim 5
In the signal compression device, for the spectrum envelope,
Performs weighting on frequencies corresponding to human auditory sensitivity characteristics.
Perceptual weighting calculation to output as perceptual weighting coefficient
Means, wherein said vector quantization means comprises:
The quantization of the above normal residual signal is performed using the
In this case, the auditory weighting coefficient is
Instead, it can be obtained by a predetermined finite number of calculations.
From spectral envelope, obtained from possible mel linear prediction coefficients
Since sought, audio signal compression instrumentation which utilizes the auditory nature of human can do a good signal compression of efficiency
There is an effect that can be obtained.

[0245]

[0246]

[0247] Further, according to the audio signal compression apparatus according to the present invention (Claim 7), Oite to the audio signal compression apparatus according to claim 6, is the vector quantization means, double
A number of such vector quantizers connected in a number column
Multiplex quantization means comprising:
The means comprises at least one of the multiple quantizing means.
The vector quantization means uses the auditory weighting coefficient.
And quantize the residual signal.
Therefore, the lower frequency band that is important for human hearing
Analysis with higher frequency resolution than higher frequency band
And each of the plurality of quantization means can be used.
Used to calculate the individual auditory weighting factors
The vector envelope is not an infinite number of calculations, but a preset finite
It is possible to determine by calculation times, there is an effect that an audio signal compression apparatus utilizing auditory nature of human can do a good signal compression of efficiency is obtained.

Further, according to the audio signal compression device according to the present invention (claim 8), any one of claims 5 to 7 can be used.
In the audio signal compression device described in any of the above,
The spectral envelope calculating means calculates the input audio signal
Audio signal of a fixed time length from
Multi-stage all-pass fill for fixed time audio signals
To obtain the filter output signal for each stage,
Input audio signal and filter output for each stage
Hearing by humans by multiply-accumulate (Equation 2) with signals
Mel frequency with expanded and contracted frequency axis corresponding to sensitivity characteristics
Calculate the autocorrelation function on the axis
The mel linear prediction coefficient is obtained from the correlation function, and the above mel linear prediction coefficient is obtained.
Using the measured coefficient itself as the spectral envelope, or
A spectral envelope is obtained from the mel linear prediction coefficient.
Approximate calculation when compressing audio signals
No processing is required, and processing can be performed with a preset finite number of calculations
Thus, there is an effect that an audio signal compression apparatus capable of performing efficient signal compression utilizing human auditory characteristics can be obtained. However, (Equation 2) is

(Equation 2) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0249] Further, according to the audio signal compression apparatus according to the present invention (Claim 9), the audio according to claim 8
In the signal compression device, the all-pass filter includes 1
Since the next all-pass filter is used,
What required a limited number of operations is now feasible
By using a first-order all-pass filter, approximate calculation can be performed.
Preset finite number of operations without any need
This means that efficient signal compression can be performed.
Wear.

[0250] Further, according to the audio signal compression apparatus according to the present invention (Claim 10), in the audio signal compression apparatus according to claim 8 or claim 9, said Orupa
The filter coefficients of the filter
Using the scale, the frequency corresponding to the human auditory sensitivity characteristics
Since weighting is performed, the bark scale or
Low frequency band that is important to human hearing
Frequency range with higher frequency resolution than the higher frequency band side.
This makes it possible to obtain an audio signal compression method capable of performing efficient signal compression utilizing human auditory characteristics.

The audio signal according to the present invention (claim 11)
According to the signal compression method , a code
Audio signal compression method that compresses the amount of information
The input audio signal and the input audio signal.
Frequency corresponding to human auditory sensitivity characteristics for audio signal
Using the audio signal that has undergone axis expansion and contraction,
Of the mel frequency axis
Find the mel linear prediction coefficient from the function,
The number itself as the spectral envelope, or
Calculate the spectral envelope from the linear prediction coefficients and calculate the spectrum
Smoothing the input audio signal by using the envelope
Thus, there is an effect that an audio signal compression method capable of performing efficient signal compression utilizing human auditory characteristics is obtained.

The audio signal according to the present invention (Claim 12).
According to the signal compression method , a code
Audio signal compression method that compresses the amount of information
In the method, a fixed time
Audio signal of a predetermined time length, and
Pass through several stages of all-pass filter, filter each stage
An output signal is obtained, and the input audio signal is
To the product sum (Equation 3) performed finite times with the filter output signal of each
The expansion and contraction of the frequency axis corresponding to human hearing sensitivity characteristics
The autocorrelation function on the mel frequency axis is determined, and the
Find the mel linear prediction coefficient from the autocorrelation function on the wave number axis,
Use the Mel linear prediction coefficient itself as the spectral envelope
Or the spectral envelope from the mel linear prediction coefficients
And using the spectral envelope, the input speech
Since the signal is smoothed, the mel linear prediction
To get the number, you do not need any approximation
Thus, the processing can be performed with a finite number of calculations set, and there is an effect that an audio signal compression method capable of performing efficient signal compression using human auditory characteristics is obtained. However,
(Equation 3) is

(Equation 3) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0253]

[0254]

The audio signal according to the present invention (claim 13)
The audio signal compression method according to claim 12, according to the signal compression method.
In the method , the all-pass filter comprises :
Filter so that the performance is essentially infinite
What needed to be calculated,
The use of a multi-pass filter allows
Only a finite number of calculations can be done without need
Thus, efficient signal compression can be performed.

The audio signal according to the present invention (claim 14)
According to Patent compression method, the audio signal compression method of the serial <br/> mounting in claim 12 or claim 13, said all-pass fill
Use Bark scale or Mel scale for filter coefficients
Weighting on frequencies corresponding to human auditory sensitivity characteristics
So that the Bark scale or Mel scale
Using the low frequency band side that is important for human hearing,
Analyze with higher frequency resolution than the higher frequency band.
This makes it possible to obtain an audio signal compression method capable of performing efficient signal compression utilizing human auditory characteristics.

[0257] Further, according to the speech signal <br/> No. compression apparatus according to the present invention (Claim 15), the input audio signal, code
Signal compression device that compresses and compresses the amount of information
The input audio signal and the input audio signal.
Frequency corresponding to human auditory sensitivity characteristics for audio signal
Using the audio signal that has undergone axis expansion and contraction,
Of the mel frequency axis.
The mel formation prediction coefficient obtained from the
A feature value calculating means for converting into a feature value to be expressed;
Inverse filtering of the audio signal
Envelope normalizing means for normalizing and obtaining a residual signal;
Normalize signal based on maximum or average power
Power normalizing means for obtaining a normalized residual signal;
The vector value of the normalized residual signal is calculated using the residual codebook.
And a vector quantization means for converting to a residual code.
As a result, there is an effect that an audio signal compression apparatus capable of performing efficient signal compression utilizing human auditory characteristics can be obtained.

Further, according to the audio signal compression apparatus according to the present invention (claim 16), the audio signal compression apparatus according to claim 15 is provided.
In the apparatus, the feature amount calculating means may include the input
Cut out the audio signal of a fixed time length from the audio signal, and
Converting fixed-length audio signals to multi-stage all-pass filters
To obtain the filter output signal for each stage,
Of the output audio signal and the filter output signal of each of the above stages,
The harmful sum of products (Equation 4) is performed finite times to improve human auditory sensitivity characteristics.
Self on the mel frequency axis that has expanded and contracted the corresponding frequency axis
Find the correlation function and determine if it is an autocorrelation function on the mel frequency axis.
The mel linear prediction coefficient is obtained from
Converted to spectral feature
Therefore, when compressing the audio signal, infinitely many operations are originally required.
What was necessary was no pre-defined
Boss was made possible treatment with a finite number of operations, the speech signal compression apparatus auditory nature of human can perform efficient signal compression to take advantage there is the effect obtained. However, (number
4 ) is

(Equation 4) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0259]

[0260]

According to the audio signal compression apparatus of the present invention (claim 17), the audio signal compression apparatus of claim 16 is provided.
In the apparatus , the all-pass filter includes :
Filter so that the performance is essentially infinite
What needed to be calculated,
The use of a multi-pass filter allows
Only a finite number of calculations can be done without need
And efficient compression of the audio signal can be performed.
You.

Further, according to the audio signal compression apparatus according to the present invention (claim 18), the sound signal compression apparatus according to claim 16 or claim 17 is described.
The audio signal compression apparatus described above,
Use Bark scale or Mel scale for filter coefficients
Weighting on frequencies corresponding to human auditory sensitivity characteristics
So that the Bark scale or Mel scale
Using the low frequency band side that is important for human hearing,
Analyze with higher frequency resolution than the higher frequency band.
Sorted enabling, human voice signal compression apparatus auditory nature can be take advantage perform efficient signal compression has the effect to be obtained.

The voice recognition according to the present invention (Claim 19).
According to the recognition method , the voice is recognized from the input voice signal.
The input speech signal
And a human auditory sensitivity characteristic with respect to the input audio signal.
Using an audio signal with the frequency axis corresponding to
To find the autocorrelation function on the mel frequency axis,
Find the mel linear prediction coefficient from the autocorrelation function on the wave number axis,
A characteristic that expresses the spectral envelope from the above mel linear prediction coefficients
It is important for human hearing because we have been asked to collect
Frequency analysis of lower frequency band side than higher frequency band side
This makes it possible to perform analysis with higher performance, and has an effect of obtaining a voice recognition method capable of performing highly accurate voice recognition using human auditory characteristics.

The voice recognition according to the present invention (Claim 20).
According to the recognition method , the voice is recognized from the input voice signal.
In the voice recognition method,
From the audio signal of a certain time length,
Pass the audio signal through a multi-stage all-pass filter,
Find the filter output signal for each stage, and
Signal and the filter output signal of each stage described above for a finite number of times.
By the product sum (Equation 5), the frequency corresponding to the human auditory sensitivity characteristic
The autocorrelation function on the mel frequency axis with the expansion and contraction of the wavenumber axis is
From the autocorrelation function on the mel frequency axis.
The spectral envelope from the mel linear prediction coefficient
Since the feature value that expresses
When performing compression, an inherently infinite number of calculations was required.
However, no approximation calculation is required, and a preset finite number of
It can be processed by arithmetic and utilizes human auditory properties
A speech recognition method that can perform more accurate speech recognition
There is an effect that a law can be obtained. Where (Equation 5) is

(Equation 5) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

[0265] The voice certification according to the present invention (Claim 21)
According to 識方method, your speech recognition method according to claim 20
And the all-pass filter has a first-order all-pass filter.
Filter, so infinitely many operations are required
What was the first-order all-pass that is actually feasible
Filters allow you to predict without any approximation
The finite number of calculations that have been set is sufficient, and the processing amount is small.
And can perform highly accurate speech recognition.
You.

[0266]

[0267]

The voice recognition according to the present invention (Claim 22).
According to 識方method, in the speech recognition method according to claim 20 or claim 21, the all-pass filter Fi
Using the Bark scale or the Mel scale for the Luta coefficient,
Weight on the frequency corresponding to the auditory sensitivity characteristics of
So, using Bark scale or Mel scale,
The lower frequency band, which is important for human hearing, is
Enables analysis with higher frequency resolution than several bands
And more accurate sound utilizing the auditory properties of humans
There is an effect that a voice recognition method capable of performing voice recognition is obtained.

According to the speech recognition apparatus of the present invention (claim 23), speech is recognized from an input speech signal.
The input speech signal
And a human auditory sensitivity characteristic for the input audio signal.
Using an audio signal that has expanded and contracted the frequency axis corresponding to the
The autocorrelation function on the mel frequency axis
A method for determining the mel formation prediction coefficient from the autocorrelation function on several axes
Linear prediction analysis means, and
Cepstrum coefficient calculating means for calculating a strum coefficient,
Multiple frames of the above cepstrum coefficients and multiple standards
Calculate the distance between the model and the one with the shortest distance
Is the highest similarity among the above standard models.
And voice recognition means for recognizing that
The lower frequency band, which is important for human hearing, is
Enables analysis with higher frequency resolution than several bands
Thus, there is an effect that a speech recognition device capable of performing speech recognition with high accuracy by utilizing human auditory characteristics is obtained.

[0270] Further, according to the speech recognition apparatus according to the present invention (Claim 24), and have your <br/> to the speech recognition apparatus of claim 23, said mel linear predictive analysis means, being said input sound
From the voice signal, cut out a voice signal of a certain time length and
Time-length audio signals are passed through multiple stages of all-pass filters.
Then, a filter output signal for each stage is obtained, and
Of the audio signal and the filter output signal of each stage
The product-sum operation (Equation 6) is performed for the limited number of times,
Self-phase on mel frequency axis with corresponding expansion and contraction of frequency axis
Calculate the function and calculate from the autocorrelation function on the mel frequency axis.
Since the mel linear prediction coefficient is determined, the
When performing compression, an inherently infinite number of calculations was required.
For a finite number of pre-defined performances without any need for approximation calculations.
Thus, there is an effect that a speech recognition device capable of performing speech recognition with high accuracy by making use of the auditory properties of humans that can be processed by calculation is obtained. Where (Equation 6) is

(Equation 6) Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No.

According to the speech recognition apparatus of the present invention (claim 25), in the speech signal compression method according to claim 24, the all-pass filter includes a primary all-pass filter.
Filter, so that infinitely many operations can be performed.
What we needed was the first order
Approximate calculation is absolutely necessary by using a path filter
Finite number of operations set in advance without
And the amount of processing is reduced, and more accurate speech recognition
It can be performed.

According to the voice recognition device of the present invention (claim 26), the speech recognition device according to claim 24 or claim 25 is provided.
In the audio compression method, the filter of the all-pass filter is used.
Using the Bark scale or the Mel scale for the Luta coefficient,
Weight on the frequency corresponding to the auditory sensitivity characteristics of
So, using Bark scale or Mel scale,
The lower frequency band, which is important for human hearing, is
Enables analysis with higher frequency resolution than several bands
Thus, there is an effect that a speech recognition device capable of performing speech recognition with high accuracy by utilizing human auditory characteristics is obtained.

[0273]

[0274]

[0275]

[0276]

[0277]

[0278]

[0279]

[0280]

[0281]

[0282]

[0283]

[0284]

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an audio signal compression device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing an example of a detailed configuration of a spectrum envelope calculation unit according to the audio signal compression device according to the first embodiment of the present invention.

FIG. 3 is a block diagram showing an example of a detailed configuration of a melding coefficient calculator according to the audio signal compression device according to the first embodiment of the present invention.

FIG. 4 is a block diagram showing an example of a detailed calculation procedure of a melding coefficient calculation unit in the audio signal compression device according to the first embodiment of the present invention.

FIG. 5 is a diagram showing characteristics of a frequency axis expansion / contraction function (all-pass filter).

FIG. 6 is a block diagram showing an example of a detailed configuration of an envelope calculation unit according to the audio signal compression device according to the first embodiment of the present invention.

FIG. 7 is a block diagram showing a configuration of a speech recognition device according to a second embodiment of the present invention.

FIG. 8 is a block diagram illustrating an example of a detailed configuration of a mel linear prediction analysis unit according to the speech recognition device according to the second embodiment of the present invention.

FIG. 9 is a block diagram showing a configuration of an audio signal compression device according to a third embodiment of the present invention.

FIG. 10 is a block diagram showing a configuration of a mobile phone according to a fourth embodiment of the present invention.

FIG. 11 is a block diagram showing a configuration of a network device according to a fifth embodiment of the present invention.

FIG. 12 is a block diagram showing a configuration of a network device according to a sixth embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Time frequency conversion part 2 Spectrum envelope calculation part 3 Normalization part 4 Power normalization part 5 Multistage quantization part 6 Auditory weight calculation part 7 Mel linear prediction analysis part 8 Cepstrum coefficient calculation part 9 Voice recognition part 51 First stage quantum Quantizer 52 Second-stage quantizer 53 Third-stage quantizer

──────────────────────────────────────────────────の Continuing on the front page (51) Int.Cl. ⁷ Identification code FIG10L 9/08 301A 3/00 515Z (72) Inventor Tomokazu Ishikawa 1006 Odakadoma, Kadoma-shi, Osaka Matsushita Electric Industrial Co., Ltd. 72) Inventor Mitsuhiko Serikawa 1006 Kadoma Kadoma, Kadoma City, Osaka Prefecture, Japan Matsushita Electric Industrial Co., Ltd. (72) Inventor Dairo Katayama 1006 Odaka Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (72) Inventor Nakahashi Junichi 1006 Kadoma, Kazuma, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (72) Inventor Junko Yagi 1006, Kadoma, Kazuma, Kadoma, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (56) References JP-A-3-138700 ( JP, A) JP-A-5-313695 (JP, A) JP-A-7-160297 (JP, A) JP-A-7-191696 (JP, A) JP-A-8-115095 (JP, A A) JP-A-9-244698 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 15/00-15/02 G10L 19/00-19/14

Claims (26)

(57) [Claims] 1. An audio signal compression method for encoding an input audio signal and compressing the amount of information, said input audio signal and said input audio signal
Frequency corresponding to human auditory sensitivity characteristics for audio signals
Using the audio signal that has undergone axis expansion and contraction,
Find the autocorrelation function on the number axis, and use the mel linear prediction function from the autocorrelation function on the mel frequency axis.
Find the number and use the mel linear prediction coefficient itself as the spectral envelope
Or the spectral envelope from the mel linear prediction coefficients
Determined, the upper kiss spectrum with envelope, the audio <br/> Oh signal the input smoothes for each frame, the audio signal compression method, characterized in that. 2. A coding method for an input audio signal.
Signal pressure that compresses and compresses the amount of information
In the compression method, a fixed time length of audio is input from the input audio signal.
Audio signal of a certain length of time , and
A filter output signal for each stage is obtained through a filter, and the input audio signal and the filter for each stage are obtained.
The product sum (Equation 1) performed finitely with the output signal
Melted and expanded frequency axis corresponding to the auditory sensitivity characteristics of
The autocorrelation function on the frequency axis is obtained, and the mel linear prediction
Find the number and use the mel linear prediction coefficient itself as the spectral envelope
Or the spectral envelope from the mel linear prediction coefficients
Using the above-mentioned spectrum envelope,
An audio signal compression method characterized by smoothing an e-signal for each frame . However, (number
1) [number 1] Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No. 3. The audio signal compression method according to claim 2, wherein said all-pass filter is a first-order all-pass filter.
In it, the audio signal compression method, characterized in that. 4. The audio signal compression method according to claim 2 , wherein the filter coefficient of the all-pass filter includes a bark scale.
Corresponds to human hearing sensitivity characteristics using degree or mel scale
A method for compressing an audio signal, wherein weighting is performed on a frequency to be processed . 5. An audio signal compression apparatus which encodes an input audio signal and compresses the information amount , transforms the input audio signal into a frequency domain signal.
Time-frequency converting means for converting and outputting, the input audio signal, and the input audio signal.
Frequency corresponding to human auditory sensitivity characteristics for audio signals
Using the audio signal that has undergone axis expansion and contraction,
Find the autocorrelation function on the number axis and calculate the autocorrelation function on the mel frequency axis.
The mel linear prediction coefficient obtained from the correlation function
Or from the mel linear prediction coefficient
A spectrum envelope calculating means for obtaining a torque envelope, and normalizing the frequency domain signal with the spectrum envelope.
Normalizing means for obtaining a residual signal, and the residual signal based on a maximum value or an average value of power.
Power normalizing means for normalizing and obtaining a normalized residual signal
And the above-mentioned normal residual signal is vectorized by the residual codebook.
Vector quantization means for quantizing and converting to a residual code,
The provided, the audio signal compression apparatus characterized by. 6. The audio signal compression apparatus according to claim 5,
In location, with respect to the spectrum envelope, against the human auditory sensitivity characteristics
Weighting on the corresponding frequency, and
And a perceptual weight calculating means for outputting the perceptual weight, wherein the vector quantizing means uses the perceptual weighting coefficient.
An audio signal compression apparatus for quantizing the normal residual signal . 7. The audio signal compression device according to claim 6, wherein said vector quantization means comprises a plurality of columns connected in a plurality of columns.
Multiple quantization consisting of the vector quantization means of the number
Means , wherein the multiple quantizing means comprises a plurality of components constituting the multiple quantizing means.
At least one of the vector quantization means is configured to
The quantization of the residual signal is performed using the
That, the audio signal compression apparatus characterized by. 8. The method according to claim 5, wherein
In the audio signal compression apparatus described above, the spectrum envelope calculating means converts an audio signal having a predetermined time length from an input audio signal.
Excised O signals, the all-pass multiple stages audio signal of the predetermined time length
A filter output signal for each stage is obtained through a filter, and the input audio signal and the filter for each stage are obtained.
The product sum (Equation 2) performed finite times with the output signal
Melted and expanded frequency axis corresponding to the auditory sensitivity characteristics of
The autocorrelation function on the frequency axis is obtained, and the mel linear prediction
Find the number and use the mel linear prediction coefficient itself as the spectral envelope
Or the spectral envelope from the mel linear prediction coefficients
And requests the audio signal compression apparatus characterized by. However, (number
2) is [Equation 2] Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No. 9. The audio signal compression apparatus according to claim 8, wherein said all-pass filter is a first-order all-pass filter.
In it, the audio signal compression apparatus characterized by. 10. The audio signal compression device according to claim 8, wherein a filter coefficient of the all-pass filter includes a bark scale.
Corresponds to human hearing sensitivity characteristics using degree or mel scale
An audio signal compression apparatus , which performs weighting on a frequency to be performed . 11. An input audio signal is encoded.
Audio signal compression method to compress the amount of information
Oite, the audio signal the input pair to the input audio signal
To expand and contract the frequency axis corresponding to the human auditory sensitivity characteristics.
Auto-correlation on the mel frequency axis using the
The mel linear prediction coefficient is calculated from the autocorrelation function on the mel frequency axis.
Find the number and use the mel linear prediction coefficient itself as the spectral envelope
Or the spectral envelope from the mel linear prediction coefficients
And using the spectral envelope, the input speech signal
The smoothing, audio signal compression method, characterized in that. 12. An input audio signal is encoded.
Audio signal compression method to compress the amount of information
In the above, an audio signal of a fixed time length is
Cut out the audio signal of a certain length of time , and
To obtain a filter output signal for each stage, and the input audio signal and the filter output signal for each stage are obtained.
Hearing of humans by multiply-accumulate (Equation 3) with signals
Mel frequency with expanded and contracted frequency axis corresponding to sensitivity characteristics
The autocorrelation function on the axis is obtained, and the mel linear prediction coefficient is obtained from the autocorrelation function on the mel frequency axis.
And use the mel linear prediction coefficient itself as the spectral envelope
Or the spectral envelope from the mel linear prediction coefficients
Determined by using the spectrum envelope, the sound signal the input
An audio signal compression method characterized by smoothing . However, (Equation 3 ) is [Equation 3] Where φ (i, j) is an autocorrelation function and x [n]
Is an input signal, and y _(ij) [n] is a filter output signal for each stage. 13. The audio signal compression method according to claim 12, wherein the all-pass filter is a first-order all-pass filter.
In it, the audio signal compression method, characterized in that. 14. The method according to claim 12 or claim 13.
In the audio signal compression method, the filter coefficient of the all-pass filter includes a bark scale.
Corresponds to human hearing sensitivity characteristics using degree or mel scale
A sound signal compression method, wherein weighting is performed on a frequency to be performed . 15. An encoding method for an input audio signal.
And an audio signal compression device that compresses the amount of information
Oite, the audio signal the input pair to the input audio signal
To expand and contract the frequency axis corresponding to the human auditory sensitivity characteristics.
Auto-correlation on the mel frequency axis using the
The number is obtained from the autocorrelation function on the mel frequency axis.
Mel formation prediction coefficient
A characteristic amount calculating means for converting the input audio signal into an amount,
Envelope normalizing means for obtaining a residual signal by performing a normalization process, and the residual signal based on the maximum value or average value of the power.
Power normalizing means for normalizing and obtaining a normalized residual signal
And the normalized residual signal is vectorized using the residual codebook.
Vector quantization means for quantizing and converting to a residual code
Comprises, when it audio signal compressor according to claim. 16. An audio signal compression apparatus according to claim 15,
In the above, the feature amount calculating means cuts off the audio signal of a fixed time length from the input audio signal.
And outputs the audio signal of the fixed time length to a multi-stage all-pass filter.
Through the filter to obtain a filter output signal for each stage, the input audio signal and the filter output for each stage
Hearing by humans by multiply-accumulate (Equation 4) with signals
Mel frequency with expanded and contracted frequency axis corresponding to sensitivity characteristics
The autocorrelation function on the axis is calculated, and the mel linear prediction coefficient is calculated from the autocorrelation function on the mel frequency axis.
Patent which obtains the number, the mel linear predictive coefficients, representing spectral envelope
An audio signal compression device , which converts the audio signal into a characteristic amount . However, (Equation 4 ) is [Equation 4] Where φ (i, j) is an autocorrelation function and x [n]
Is an input signal, and y _(ij) [n] is a filter output signal for each stage. 17. The audio signal compression device according to claim 16, wherein the all-pass filter is a first-order all-pass filter.
In it, the speech signal compression apparatus characterized by. (18) The method according to (16) or (17),
Oite to the audio signal compression apparatus, the filter coefficients of the all-pass filter, bark scale
Corresponds to human hearing sensitivity characteristics using degree or mel scale
An audio signal compression apparatus for performing weighting on a frequency to be performed . 19. Recognizing a voice from an input voice signal.
In speech recognition method for the speech signal the input pair to the input audio signal
To expand and contract the frequency axis corresponding to the human auditory sensitivity characteristics.
Auto-correlation on the mel frequency axis using the
The mel linear prediction coefficient is calculated from the autocorrelation function on the mel frequency axis.
Number, and express the spectral envelope from the mel linear prediction coefficient.
A voice recognition method for determining a collection amount . 20. Recognizing a voice from an input voice signal
In the voice recognition method, a voice signal having a predetermined time length is converted from the input voice signal.
Cut out the audio signal of a certain length of time , and
To obtain a filter output signal for each stage, and the input audio signal and the filter output signal for each stage are obtained.
Hearing by humans by multiply-accumulate (Equation 5) with signals
Mel frequency with expanded and contracted frequency axis corresponding to sensitivity characteristics
The autocorrelation function on the axis is obtained, and the mel linear prediction coefficient is obtained from the autocorrelation function on the mel frequency axis.
The calculated feature representing the spectral envelope from the mel linear predictive coefficients
A speech recognition method characterized by determining a quantity . However, (Equation 5) is [Equation 5] Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [ n] is the filter output signal for each stage.
No. 21. The speech recognition method according to claim 20 , wherein said all-pass filter is a first-order all-pass filter.
In a speech recognition method characterized by. 22. The speech recognition method according to claim 20 , wherein a bark scale is added to a filter coefficient of the all-pass filter.
Corresponds to human hearing sensitivity characteristics using degree or mel scale
A voice recognition method for performing weighting on a frequency to be performed . 23. Recognizing a voice from an input voice signal
A voice recognition device that performs the above-described processing on the input voice signal and the input voice signal.
To expand and contract the frequency axis corresponding to the human auditory sensitivity characteristics.
Auto-correlation on the mel frequency axis using the speech signal
A function is calculated, and the mel
A mel linear prediction analysis means for determining a formation prediction coefficient, and calculating a cepstrum coefficient from the mel linear prediction coefficient
Cepstrum coefficient calculation means, a plurality of frames of the cepstrum coefficient,
Calculate the distance between the model and the one with the shortest distance
Is the highest similarity among the above standard models.
And a speech recognition means for recognizing as speech recognition apparatus characterized by. 24. The speech recognition apparatus according to claim 23 , wherein the mel linear prediction analysis means converts a speech signal having a predetermined time length from the input speech signal.
Cut out the audio signal of a certain length of time , and
To obtain a filter output signal for each stage, and the input audio signal and the filter output signal for each stage are obtained.
By the sum of products (Equation 6) performed finite times with the signal, human hearing
Mel frequency with expanded and contracted frequency axis corresponding to sensitivity characteristics
The autocorrelation function on the axis is calculated, and the mel linear prediction coefficient is calculated from the autocorrelation function on the mel frequency axis.
A voice recognition device for determining a number . However, (Equation 6) is [Equation 6] Where φ (i, j) is an autocorrelation function and x [n]
Is the input signal, and y _(ij) [n] is the filter output signal for each stage.
No. 25. The audio signal compression method according to claim 24, wherein said all-pass filter is a first-order all-pass filter.
In it, the speech recognition apparatus characterized by. (26) The method according to (24) or (25),
In the audio compression method, the filter coefficient of the all-pass filter includes a bark scale.
Corresponds to human hearing sensitivity characteristics using degree or mel scale
A voice recognition device for performing weighting on a frequency to be performed .