JP2001273000A - Adaptive noise suppressing speech encoder - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an adaptive noise suppressing speech encoder which detects a noise level during current talk, makes it possible to eliminate the noise or reduce an influence thereof, and also prevents a signal processing time from increasing by integrating the speech encoder and a noise suppressor in one body and further simplifying the software. SOLUTION: An adaptive noise suppressing speech encoder without a malfunction due to noise can be obtained by paying attention to a feature of an auto- correlation function between a speech signal and noise, subtracting a noise component from the auto-correlation calculated by an auro-correlation function calculating circuit 1303, obtaining an auto-correlation function of only the speech signal, creating a speech encoding parameter by a prediction coefficient calculating circuit 102 using the value, and further eliminating the noise by an internal operation of the speech encoder by a block A for eliminating the noise during talking.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive noise suppression speech coding apparatus applicable to communication technology fields such as wired and wireless telephones (mobile telephones and automobile telephones). To reduce or reduce the effect thereof, and to integrate the speech encoder and the noise suppressor to simplify the hardware and prevent an increase in signal processing time. It is composed.

[0002]

2. Description of the Related Art In recent years, devices using voice as input information have been used in various environments, and it has become important to be able to use the devices even in a noisy environment. Cellular phones and car phones are also examples. Advances in IC technology have made it possible to reduce the ambient noise superimposed on the audio signal by using a DSP (Digital Signal Processor) in such a device, and by using advanced digital signal processing technology. Have been. Hereinafter, a conventional noise suppression device will be described with reference to the drawings.

[0003] Regarding the conventional noise suppressor, the method is described in "IEEE Transaction on A" published in April 1979.
coustics, speech, and Signal Processing, vol. ASSP
-27, NO.2, pp, 113-120 "" Supression of Acoustic N
oise Speech Using Spectral Subtraction "as a spectrum subtraction method. This spectrum subtraction method will be briefly described using a conventional first noise suppression device shown in FIG. Fig. 25 is a block diagram showing a configuration of a noise suppression device of Fig. 25. In Fig. 25, reference numeral 901 denotes a cutout unit, 902 denotes a Fourier analysis unit, 903 denotes a noise spectrum average calculation unit, and 904.
Is a noise spectrum memory, 905 is a noise signal determination unit, 906 is an amplitude subtraction unit, 907 is a half-wave rectification unit, 908 is an inverse Fourier transform unit, and 909 is a waveform reproduction unit. The operation of the conventional noise suppressor composed of the above components is shown in FIG.
This will be described below using the flowchart shown in FIG.

In step 1001, the input speech is cut by a window for obtaining a frequency resolution accuracy required for Fourier analysis later in the cutout unit 901 for 10 ms.
A waveform in a fixed-length frame of a section of about 20 ms is cut out at a cycle of about. As a condition of the window for cutting out a waveform of a certain section from a continuous waveform in this manner, a discontinuous waveform does not occur when combined and connected in a later combining process, and the accuracy of Fourier analysis can be secured. Select an analysis window that overlaps by 50% so that the waveform of The following Hanning window is used as this window function. Here, assuming that the number of data of the Fourier analysis is N, the weighting W (n) for the n-th data after windowing is expressed by Equation (1).
It is expressed as shown below.

W (n) = 0.5−0.5 * cos [2nπ / (N−1)] (1)

In step 1002, the Fourier section 902 performs frequency analysis (FFT analysis) on the speech waveform of the frame to be analyzed by Fourier analysis, and the analysis result is composed of a real part and an imaginary part. Decompose into an amplitude spectrum and a phase spectrum. The amplitude spectrum is output to noise spectrum memory 904, noise signal determination section 905, and noise subtraction section 906. On the other hand, the phase spectrum is output to inverse Fourier transform section 908 for resynthesizing the speech waveform.

In step 1003, a noise signal determination unit 905
In, it is determined whether the input signal is a signal in which a noise signal is superimposed on a voice signal or a signal including only noise. This determination is made based on the power information of the input signal. A high-power portion is determined to be a voiced voice section, and a low-power portion is determined to be a no-voice noise section. The information “present” and “absent” of this voice is output to the noise spectrum memory 904.

In step 1004, the noise spectrum in the past several tens of ms is stored in the noise spectrum memory. In the noise section, the amplitude spectrum from Fourier analysis section 902 is stored. In the voice section, the noise spectrum memory 904 does not store the amplitude spectrum from the Fourier analysis unit 902.

For this reason, the noise spectrum memory 904 is always used.
Means that the latest noise spectrum is stored. In step 1005, the noise spectrum average calculating unit 903 calculates the past number 1 stored in the noise spectrum memory 904.
The average spectrum of the amplitude spectrum of the 0 frame is calculated as the amplitude information of the estimated amount of noise, and is set as the amplitude spectrum of the noise immediately before speech utterance, that is, the noise spectrum.
Thus, the amount of memory used to calculate the noise spectrum is the number of window lengths * the number of storage frames analyzed by Fourier analysis.

At step 1006, the amplitude subtraction unit 906 calculates the noise amplitude for each frequency of the noise spectrum average calculation unit 903 from the amplitude corresponding to each frequency of the voice signal on which the noise from the Fourier analysis unit 902 is superimposed. , The amplitude of the voice only is calculated by subtraction for each frequency.

In step 1007, the half-wave rectification unit 907 sets the amplitude of the frequency at which the amplitude is negative as a result of the subtraction in the amplitude subtraction unit 906 to zero. Step 10
As 08, the inverse Fourier transform unit 908 synthesizes a voice waveform from which noise is removed from the amplitude corresponding to each frequency from the half-wave rectification unit 907 and the phase corresponding to each signal from the Fourier analysis unit 902. By doing so, a noise-free speech waveform for one frame is generated.

In step 1009, the waveform reproducing unit 909 performs a reverse process of the clipping process in the clipping unit on the one-frame noise-free speech waveform from the inverse Fourier transform unit 908, and performs frame-by-frame processing for one frame. Are overlapped and added to output the original continuous waveform.

The above-described apparatus (or method) digitizes an audio signal on which noise is superimposed, applies a window function to extract the audio signal at a constant frame rate, and performs FFT (Fast Fourier Transform) to discretize the signal in the frequency domain. On the other hand, the noise level during the period when there is no voice is previously
In this configuration, a signal is stored in a storage device in advance as a signal in the frequency domain that has been discretized by T (Fast Fourier Transform), and the value is subtracted for each frequency. However, another device (method) is also a conventional example. Is considered as.

FIG. 27 is a block diagram showing the configuration of a second conventional noise suppression device using such a method. 27, illustrated numbers 901 to 905, and illustrated numbers 908 and 909
Are the same as those shown in FIG. Drawing number 11
01 is an S / N measurement unit for each channel, 1102 is a gain control unit for each channel, and 1103 is a variable gain amplifier for each channel.

[0015] The second conventional system comprising the above components is described.
The operation of the noise suppressor described above will be described below with reference to the flowchart shown in FIG. Steps 1001 to 1005 perform the operation already described with reference to FIG.

Next, in step 1104, each channel S / N
In the measuring section, for each frequency converted to the frequency domain (the frequency takes a discrete value), the level of the current speech signal including noise is stored in the noise spectrum memory in step 1005.
The noise level of each frequency stored in 904 is taken out, and a signal-to-noise ratio (S / N ratio) is calculated. In step 1105, each channel gain control unit 1102 uses the value to obtain a signal-to-noise ratio (S / N ratio). Instead of using the S / N ratio as it is, a certain converted value may be used.) The gain control value is passed to the variable gain control unit 1103 for each channel.

In step 1106, the audio signal containing noise is amplified by changing the gain for each channel using the value from the variable gain control unit for each channel. That is, it is possible to suppress noise by increasing the gain in a channel having a good ratio and decreasing the gain in a channel having a poor ratio. Actually, since controlling the gain by changing the frequency for each frequency increases the number, the frequency is divided into groups by several groups and the channels are controlled. The operations of steps 1008 and 1009 have been described with reference to FIG.

The conventional noise suppressor described above is
The noise level of a frame determined to have no voice in the past is stored for each frequency band, and a voice signal including noise of the current frame determined to have voice is converted to a frequency domain, and each frequency band is For each, subtract the stored noise level, or calculate the S / N ratio for each frequency and control the gain of the amplifier for each frequency,
It does not reduce noise by adapting to the noise level in the frame currently being talked to. Further, the conventional noise suppressor is not intended to simplify the hardware by integrating the noise suppressor with a speech encoder currently used in mobile phones.

According to the present invention, to estimate the amount of noise contained in a speech signal, an autocorrelation function of an input signal containing noise is calculated, and a characteristic of the current speech is utilized by utilizing characteristics of the autocorrelation function. Estimate the amount of noise contained in the signal. However, the digital method using this (TDMA method or CD
(MA system) In a mobile phone, a speech coding device is connected after the noise suppression device. In the speech coding apparatus, the autocorrelation function of the input signal is calculated and used in a coding parameter calculation system.

Focusing on this, the noise suppression device and the speech coding device should be able to be integrated, whereby
The adoption of the noise suppression device should prevent the circuit from becoming complicated and increasing power consumption. For that purpose, it is considered good to fully understand the contents of the voice encoding device currently used in the digital mobile phone, and then to devise an integrated device.

There are various types of speech coding devices used in current digital cellular phones, but most of them are
This is called CELP (Code-Excited Linear Prediction). Therefore, ACELP (Algebraic CEL), which is most widely adopted as the prior art, is described below.
P; algebra CELP) is described below.
For the explanation, the following materials which are the standard documents of the GSM mobile phone are used. `` GSM Global System For Moble Communications '' Euro
pean Telecommunication Standard ETS 300 726 March 1
997 “Digital cellular telecommunications system; En
hanced Full Rate speech transcoding "(GSM 06.06)"

FIG. 29 is a simplified block diagram of the ACELP speech coding apparatus. In FIG. 29, the reference numbers 1201 are a fixed codebook, 1202 is an amplifier, 1203 is an adaptive codebook,
1204 is an amplifier, 1205 is an adder, 1206 is a linear prediction synthesis circuit, and 1207 is an auditory correction circuit.

The general operation of the ACELP speech coding apparatus will be described below with reference to FIG. Fixed codebook 1201
Is a positive / negative random impulse train having an amplitude of 1. The adaptive codebook 1203 is a repetitive pulse train having a constant period. They are multiplied by different gains in the amplifiers 1202 and 1204, respectively. So far, it corresponds to the sound source part of the human voice generation mechanism.

Then, both of them are added to the adder 1205, and they are input together to the linear predictive synthesis circuit 1206, and the output is a voice signal. So far, it corresponds to the sound path portion of the human voice generation mechanism. That is, the sound is output to the outside via the throat and the mouth.

The output of the linear prediction synthesis circuit 1206 is input to the auditory correction circuit 1207. The output is an audio signal modified in consideration of the human voice sensing characteristics. Note that what is sent from the speech encoding device to the decoding device on the receiving side is
Fixed codebook 1201, amplifier 1202, adaptive codebook
1203, amplifier 1204, and linear prediction synthesis circuit 1206.

Next, FIG. 30 shows a detailed functional block diagram of the ACELP speech coding apparatus. Referring to FIG.
Is a preprocessing circuit, 1302 is a windowing circuit, 1303 is an autocorrelation function calculation circuit, 1304 is a prediction coefficient calculation circuit, 1305 is a coefficient conversion circuit, 1306 is a coefficient quantization circuit, 1307 is a subframe coefficient interpolation circuit, and 1308A is Subframe coefficient interpolation circuit (before quantization), 1308B is an impulse response calculation circuit, 1309 is a 4-subframe weighted speech calculation circuit, 1310A is an open loop pitch delay detection circuit, 1310B is an open loop pitch gain detection circuit, 1311A Is an application codebook, 1311B is a pitch gain adjustment circuit, 1311C is an impulse response circuit, 1311D is an error calculation circuit, 1311E is an error minimization circuit, 1312 is an optimum delay and gain finding circuit, 1313 is a pitch gain quantization circuit, 1314 is Applied codebook contribution calculation circuit, 1316 is a fixed codebook, 1317A is a fixed codebook gain adjustment circuit, 1317B is an impulse response circuit, 1317C is an error calculation circuit, and 1317D is an error minimization circuit. 1313E, a fixed codebook optimum value finding circuit, 1318, a fixed codebook gain quantization circuit, 1319, an excitation calculation circuit, and 1320, an optimum filter storage circuit.

Although the name "Circuit" is used in FIGS. 29 and 30, it is actually a one-chip semiconductor integrated circuit which is now called a digital signal processor (DSP). Although it is constituted by a storage device, the above names are used for convenience of explanation.

The general operation of the ACELP speech coding apparatus will be described below with reference to FIG.

[Pre-processing circuit 1301] The pre-processing circuit 1301 amplifies the output of the audio signal of the microphone of the portable telephone as an electric signal, passes through a low-pass filter, and is twice the cutoff frequency of the low-pass filter. It works by sampling at a frequency and supplying it as a digital signal to an audio encoder through an A / D converter. Taking specific values as an example, the cutoff frequency of the low-pass filter is 4 kHz, the sampling frequency is 8 kHz, and the A /
The D converter converts one sample into a 13-bit digital signal.
This digital signal is input to the windowing circuit 1302.

[Windowing circuit 1302] FIG. 31 shows a situation where the windowing circuit 1302 multiplies a digital signal by a windowing function and extracts the digital signal. That is, one frame of the audio signal is 20 ms,
It is divided into four equal parts and is composed of four subframes of 5 ms per subframe. The data actually taken in as the data are the second subframe and the fourth subframe. The prediction coefficient of the signal is calculated for these subframes, and the calculation is performed for the first and third subframes. Using the results, it is determined by interpolation. The shapes of the window functions applied to the second subframe and the fourth subframe have different shapes as shown in FIG. The data to be fetched is in the period of 30 ms, and it is the same that there are 240 samples.

[Autocorrelation Function Calculation Circuit 1303] The autocorrelation function calculation circuit 1303 predicts a current sample from a past sample by utilizing the fact that the taken audio signal has redundancy. However, since it cannot be done completely, an error remains. Assuming that the windowed audio signal is s' (n), a predicted value s "(n)
Is represented by the following equation (2). s ″ (n) = − Σ _{k = 0} ^{k = 10} a _k s ′ (nk) (2) That is, prediction is performed using data at the past 10 sample points. The prediction error e (n) is expressed by the following equation (3). . represented by e (n) = s '( n) -s "(n) = Σ k = 0 k = 10 a k s'(nk); k = 0, -, -, 10; a 0 = 1 ( 3) Here, a ₁ , a ₂ , −, −, and a ₁₀ may be selected so that the following equations (4) and (5) are minimized. e (n) * e (n ) = Σ k = 0 k = 10 a k s '(nk) * Σ j = 0 j = 10 a j s' (n-j) (4) e (n) * e (n) = Σ _{k = 0} ^{k = 10} Σ _{j = 0} ^{j = 10} a _k a _j s '(nk) * s' (n-j) (5) Here, the audio signal was captured through a window , S '(nk) * s' (n
-j) depends only on (n-j)-(nk) = (k-j),
It is represented by an autocorrelation function according to the following equation (6). e (n) * e (n) = Σ _{k = 0} ^{k = 10} Σ _{j = 0} ^{j = 10} a _k a _j R (k-j) (6) Here, the following equation (7) holds. . R (kj) = s '(nk) * s' (n-j) (7) Strictly speaking, it is calculated with all sample values in which n has been taken in and averaged.

[Prediction coefficient calculation circuit 1304] Prediction coefficient calculation circuit
1304 minimizes equation (6), a₁, A_Two,-,-, A_Ten 
The prediction coefficient is calculated by selecting. The method is
A in equation (6)₁, A_Two,-,-, A _Ten And the partial differential
Then, a value that makes the result zero may be selected. The calculation process is
Although omitted, the result is given by the following equation (8).

(Equation 1)

Since the above equation (8) can be solved by the Levinson-Durbin algorithm, the optimum prediction coefficients a ₁ , a ₂ ,-,-, and a ₁₀ are obtained using the autocorrelation function.

[Coefficient conversion circuit 1305] The coefficient conversion circuit 1305
The optimal prediction coefficients a ₁ , a ₂ ,-,
-, but a ₁₀ is a parameter in the time domain (the coefficient of the impulse response), which is a parameter of the frequency domain, the following manner to another factor, called line spectrum pair (LSP), is converted . As described above, the prediction error was expressed by equation (3).

When this is z-transformed, the following equation (9) is obtained. E (z) = S (z) + Σ _{k = 1} ^{k = 10} a _k S (z) z ^-k ; k = 1,2,3,-,-, 10 (9) A (z) is defined by the following equation (10) as a function. A (z) = E (z) / S (z) = 1 + Σ _{k = 1} ^{k = 10} a _k z ^-k ; k = 1,2,3,-,-, 10 (10) A (z) Are used to define the following functions (11) and (12). P (z) = A (z) −z ⁻¹¹ A (z ⁻¹ ) (11) Q (z) = A (z) + z ⁻¹¹ A (z ⁻¹ ) (12) P (z), Q ( z) is divisible by (1−z ⁻¹ ) and (1 + z ⁻¹ ), respectively. The remainder is expressed as a polynomial in z. And this polynomial has a root on the unit circle of z. Accordance connexion, at ^{z = e sT = e σ +} jωT = e jωT (13) the above equation is obtained T = 1 far and the following expression (14). (z + z ⁻¹ ) / 2 = (e ^jω + e ^−jω ) / 2 = cosω (14), and this polynomial has five roots related to cosω,
It is expressed by the following equations (15) and (16). P (z) = (1−z ⁻¹ ) Π _{i = 1} ^{i = 5} (1-2 z ⁻¹ cos (ωi) + z ⁻² ); i = 1, −, −, 5 (15) Q ( z) = (1 + z ^-1 ) Π _{i = 1} ^{i = 5} (1-2 z ^-1 cos (βi) + z ^-2 ); i = 1,-,-, 5 (16) This is the sum of ωi and βi 10 roots are line spectrum
This is called a pair (LSP).

This line spectrum pair (LS
P) is known to have the following advantages. (a) The stability of the system can be determined only by comparing the parameters. (b) Since it is a parameter on the frequency axis, the movement of the parameter is similar to the movement of the spectrum envelope. That is, the parameter interpolation characteristics are good. (c) The sensitivity of the parameter to spectral distortion is almost constant.

[Coefficient quantization circuit 1306] Coefficient quantization circuit 1306
Converts the LSP thus obtained into an appropriate number of bits in order to transmit the LSP to the receiving side in a wireless system.

[Subframe coefficient interpolation circuit 1307] The subframe coefficient interpolation circuit 1307 calculates the LSPs of the second and fourth subframes calculated as described above.
Are used to determine the LSP parameters of the first and third subframes by the following equations (17) and (18). LSP (frame: m; subframe: 1) = 0.5 * LSP (frame: (m-1); subframe: 4) + 0.5 * LSP (frame: (m); subframe: 2) (17) LSP (frame: m; subframe: 3) = 0.5 * LSP frame: (m); subframe: 2) {+0.5 * LSP frame: (m); subframe: 4) (18) Using these values, A '(z) which is a transfer function of the linear prediction circuit is obtained. The reason why “′” is added to A is to distinguish it from A (z) obtained by using the LSP before quantization described below.

[Subframe coefficient interpolation circuit (before quantization) 1
308A] The subframe coefficient interpolation circuit (before quantization) 1308 similarly obtains the LSP parameters of the first and third subframes using the LSP before performing quantization for transmission to the wireless system, Using it, the transfer function A (z) of the linear prediction circuit is obtained.

[Impulse Response Calculation Circuit 1308B] The impulse response calculation circuit 1308B includes A (z), which is a transfer function of the linear prediction circuit, obtained by the subframe coefficient interpolation circuit (before quantization) 1308A and the subframe coefficient Using the transfer function A '(z) of the quantized linear prediction circuit obtained by the interpolation circuit 1307, h (n) which is the impulse response of the linear prediction synthesis circuit corrected for human hearing. ). The z-transform of h (n) is represented by the following equation (19). H (z) = (1 / A ′ (z)) · (W (z)) = {A (z / γ ₁ )} / {A ′ (z) · A (z / γ ₂ )} (19) Usually, γ ₁ = 0.9 and γ ₂ = 0.6 are selected. H (z) and A '(z)
Is in a reciprocal relationship. A ′ (z) subtracts a predictable amount from the input audio signal and outputs a residual. The residual becomes a random pulse when prediction is properly performed. H
(z) indicates that a residual, that is, a random pulse is input, an audio signal is output, and the output is obtained by correcting human auditory characteristics. Although a detailed method of obtaining h (n) is omitted, h (n) is obtained by the above-described inverse z-transform of H (z). This h (n) is necessary for searching the adaptive codebook and the fixed codebook.

[4 Subframe Weighted Speech Calculation Circuit 130
9] The 4-subframe weighted audio calculation circuit 1309 passes the audio of the 4 subframes (that is, one audio signal) of the digital audio signal before windowing through a filter of a transfer function in consideration of human auditory characteristics. The filter of the transfer function considering human auditory characteristics is expressed by the following equation (20). W (z) = {A (z / γ ₁ )} / {A (z / γ ₂ )} (20) Here, γ ₁ = 0.9 and γ ₂ = 0.6. That is, since A (z) is found by the subframe coefficient interpolation circuit (before quantization) 1308A, W (z) is also obtained.

[Open loop pitch delay finding circuit 1310A], [Open loop pitch gain finding circuit 1310B] Open loop pitch delay finding circuit 1310A, open loop pitch gain finding circuit 1310B
Estimates the pitch cycle, ie, the delay time and the gain, once every 10 ms, that is, twice every 20 ms, which is one frame of the audio signal. The method of finding pitch delay and gain by open loop is represented by the following relational expression. Rlag (open) = Σ _{n = 0} ^{n = 79} sw (n) sw (nk); n = 0, −, −, 79 (21) This value is calculated in the following three ranges. (3rd range; i = 3): k = 18,-,-,-, 35 (22) (2nd range; i = 2): k = 36,-,-,-, 71 (23) (23rd) 1 range; i = 1): k = 72,-,-,-, 143 (24)

As a result, a pitch delay and a pitch gain optimum value are selected as values for maximizing Rlag (open) from among three values in three ranges for maximizing Rlag (open),
The delay time is passed to the adaptive codebook 1311A, and the pitch gain is passed to the pitch gain adjustment circuit 1311B.

[Adaptive Codebook 1311A], [Pitch Gain Adjustment Circuit 1311B] The adaptive codebook 1311A is a codebook of impulses representing a pitch component of a voice signal, and adaptively stores a pitch delay time and generates a pulse train. The pitch gain adjustment circuit 1311B has a function of storing and adjusting the pitch gain. First, the initial values of the adaptive codebook 1311A and the pitch gain adjustment circuit 1311B are the pitch delay time and the pitch gain obtained by the open loop pitch delay finding circuit 1310A and the open loop pitch gain finding circuit 1310B, respectively, as the adaptive codebook 1311A. Are used in the pitch gain adjustment circuit 1311B.

[Impulse response circuit 1311C] The impulse response circuit 1311C obtains h obtained by the impulse response calculation circuit 1308B.
(n), adaptive codebook 1311A, pitch gain adjustment circuit 1
The pitch gain pulse synthesized by 311B is convoluted and integrated to convert it into a synthesized signal corresponding to a weighted audio signal, which is input to an error calculation circuit 1311D.

[Error Calculation Circuit 1311D] Error Calculation Circuit 1311D
Is a weighted voice signal calculated by the 4-subframe weighted voice calculation circuit 1309 and the impulse response circuit 1311
The difference of the synthesized signal corresponding to the weighted audio signal which is the output of C, that is, the error is calculated and output to the error minimizing circuit 1311E.

[Error minimizing circuit 1311E] Error minimizing circuit 131
1E finds the mean square value of the error calculated by the error calculation circuit 1311D, and adapts the adaptive code book 13 so as to minimize the value.
11A and the pitch gain adjustment circuit 1311B. Although the details are omitted, the open loop pitch delay finding circuit 1310 described above is used.
A and the periphery of the pitch delay time and pitch gain obtained by the open loop pitch gain finding circuit 1310B are precisely searched.

[Optimal delay and gain finding circuit 1312] The optimal delay and gain finding circuit 1312 finds the optimum delay time and gain value as a result of the control of the error minimizing circuit 1311E.

The value of the adaptive codebook and the pitch gain thus obtained are stored respectively, and the optimal values of the next adaptive codebook and the pitch gain are searched by using them, so that the operation can be performed efficiently. The delay time required for the search is suppressed to 40 subframes (200 ms) or less.

[Pitch Gain Quantizing Circuit 1313] The pitch gain quantizing circuit 1313 quantizes the optimum pitch gain found by the optimum delay and gain finding circuit 1312 with an appropriate number of bits in order to transmit the optimum pitch gain through a radio system. Become

[Applied Codebook Contribution Calculation Circuit 1314] The applied codebook contribution calculation circuit 1314 converts the error signal obtained by using the impulse train based on the value obtained by the optimum delay and gain finding circuit 1312 into the error calculation circuit 1311D. It outputs to the error calculation circuit 1317C as an output.

[Fixed Code Book 1316] Fixed Code Book
1316, those pulses made of a combination of a total of 10 single pulses divided into 5 tracks,
It can take the values of "+1" and "-1". There are 40 subframes as sample points, so each track is 8
This is a sample point, and the position of each pulse is represented by 3 bits for each track. The situation is shown in FIG.

[Fixed Gain Adjustment Circuit 1317A] The fixed gain adjustment circuit 1317A is a circuit for adjusting the amplitude of eight pulses per subframe generated by the fixed codebook 1316, and outputs its output to the impulse response circuit 1317B.

[Impulse Response Circuit 1317B] The impulse response circuit 1317B obtains h obtained by the impulse response calculation circuit 1308B.
(n), fixed codebook 1316, fixed gain adjustment circuit 1317
The fixed code impulse synthesized by A is convoluted and integrated to convert it into a synthesized signal corresponding to a weighted audio signal, which is input to an error calculation circuit 1317C.

[Error Calculation Circuit 1317C] Error Calculation Circuit 1317C
Is the error minimizing circuit 13 by the impulse response circuit 1315.
A signal obtained by converting the error signal minimized in 11E into a weighted audio signal, and a fixed code impulse synthesized by a fixed codebook 1316 and a fixed gain adjustment circuit 1317A by an impulse response circuit 1317B, a weighted audio signal Is calculated, and the difference is output to the error minimizing circuit 1317D.

[Error minimizing circuit 1317D] Error minimizing circuit 131
7D calculates the mean square value of the error calculated by the error calculation circuit 1317C, and fixes the fixed code book 13 so as to minimize the value.
16 and the fixed gain adjustment circuit 1317A. The details are omitted.

[Fixed codebook optimum value finding circuit 1317E]
The fixed codebook optimum value finding circuit 1317E controls the fixed codebook 1316 and the fixed gain adjustment circuit 1317A so that the error minimizing circuit 1317D minimizes the mean square value of the error. To find the optimal value of.

[Fixed codebook gain quantization circuit 1318]
The fixed codebook gain quantization circuit 1318 quantizes the gain of the fixed codebook with an appropriate number of bits necessary for wireless transmission.

[Excitation Calculation Circuit 1319] The excitation calculation circuit 1319 calculates the optimum values of the adaptive codebook and the pitch gain and the optimum values of the fixed codebook and the fixed gain for each sub-frame. The excitation pulse signal of the synthesized speech is created by the following equation (25). u (n) = g _p v (n) + g _a C (n); n = 0,1,2,-,-,-, 39 (25) where, u (n) is an excitation vector, v (n) is the excitation vector of the adaptive codebook, g _p is its gain, C (n) is the excitation vector of the fixed codebook, g _a
Is the gain.

[Optimal Filter Storage Circuit 1320] The optimum filter storage circuit 1320 stores the parameters of the linear prediction synthesis circuit of FIG. 29 and its initial conditions in preparation for the next operation. The data transmitted to the receiving side as encoded data by the ACELP audio encoding device is as follows. (1) Line spectrum pair (2) Delay time of pitch (3) Gain of pitch (4) Address of fixed codebook (5) Gain of fixed codebook

[0061]

However, in the above-described noise suppressor, the noise removed from the speech is past noise, not noise during the current call, so that the noise is reduced during the call. There was a problem that was not able to respond to.

Further, a mobile phone uses a speech coding apparatus, and performs considerably complicated signal processing. However, adding a noise suppression apparatus further increases the hardware complexity. Increase current consumption,
There were problems such as an increase in delay time due to signal processing.

The present invention solves such a problem, detects a noise level during a current call, removes the noise level, or reduces the influence of the noise level. It is a first object of the present invention to provide an adaptive noise-suppressed speech coding apparatus that integrates a suppression apparatus, simplifies hardware, and prevents an increase in signal processing time.

Further, by transmitting coded information from which noise has been removed to the receiving side, the receiving side can not only reproduce the noise-reduced sound, but also reproduce the noise inside the voice coder. Since noise is also removed during operation, even when the noise level becomes large, the speech encoding apparatus does not malfunction. That is, it is a second object of the present invention to provide an adaptive noise suppression speech coding apparatus which does not malfunction even when excessive noise is added to the speech coding apparatus.

[0065]

Means for Solving the Problems According to the first aspect of the present invention.
Of the present invention converts an audio signal on which ambient noise is superimposed into an analog signal.
A means for converting to a digital signal with a digital converter,
Means to extract as data of fixed length section, extracted signal
Calculate the autocorrelation function R (0), R (1), R (2),-,-, R (M) of the signal
Means, the extracted signal is a section with only noise,
Judging means for judging whether the audio signal section is folded,
If the determination means determines that the section is only noise,
Let the autocorrelation function R (0), R (1), R (2),-,-, R (M) be the value of Rn_m(0), Rn
_m(1), Rn _m(2),-,-, Rn_mNoise self-phase stored as (M)
Function storage means, autocorrelation function R (0), R (1), R (2),-,-, R
Means for determining a constant of the first prediction means based on the value of (M);
Measuring the impulse response of the reciprocal of the transfer function of the first predicting means
First impulse response calculating means for calculating
Book, first fixed code gain adjusting means,
The first fixed via the fixed code gain adjusting means of
A codebook output pulse train and the first impulse response
Convolution with the impulse response calculated by the answer calculation means
First convolution means for obtaining an output, said first convolution means
First error for calculating the error between the output of the integration means and the noise
Calculating means for calculating a mean square value of the error and calculating the error
The output code of the first fixed book and the previous
The gain of the first fixed code gain adjusting means is controlled.
First control means, said first impulse after minimization is achieved
A first noise state storage means for storing a response constant and an initial state
A step, wherein the determination means further includes a sound on which noise is superimposed.
If it is determined that the voice signal section, the noise autocorrelation
Rn stored in function storage means_m(0), Rn_m(1), Rn_m(2),-,
-, Rn_mRead the values of (M) and those values and the autocorrelation
R (0), R (1), R (2),-,-, R (M) calculated by the function calculation means
Calculates the value K determined as a function of
A noise-free audio signal using the value of K
Only the autocorrelation functions Rs (0), Rs (1), Rs (2),-,-, Rs (M)
And R (0) −KRn_m(0), R (1) −KRn_m(1), R (2) −KRn_m(2),-,-,
R (M) −KRn_mMeans for calculating (M), said Rs (0), Rs (1), Rs
(2) Determine the constant of the second prediction means using-,-, Rs (M)
Means, the impulse response of the inverse of the transfer function of the second predictor
Second impulse response calculation means for calculating the answer,
Constant code book, second fixed code gain adjustment means, before
The second fixed code gain adjusting means;
Output pulse train of the fixed codebook and the second impulse.
Second convolution product to obtain convolution integral output with Lus response
The output of the second convolution integrator and the weight of the noise.
Error between the demultiplexed audio signal and the denoised signal
Second error calculating means for calculating the mean square value of the error
And the second fixed block so as to minimize the error.
Output code and the second fixed code gain adjusting means.
Second control means to control stage gain, achieve minimization
After that, the constant and the initial state of the second impulse response are stored.
A second audio signal state storage means for
Is done.

By constructing and operating the speech encoding apparatus as described above, noise is eliminated in the internal operation of the speech encoding apparatus, and operation with an improved signal-to-noise ratio is realized. This has the effect of preventing a malfunction due to noise of the digitizing device.

The invention described in claim 2 is the same as that in claim 1.
In the above-described speech encoding apparatus, when the decision unit decides that the speech section has noise superimposed thereon, the output of the second error calculation unit is used as an input to the control unit that minimizes the first error. Is used to stabilize the operation of the first error calculating means.

The third aspect of the present invention is the first aspect of the present invention.
Or the adaptive noise-suppressed speech coding apparatus according to claim 2.
The Rn stored in the first storage means_m(0), Rn
_m(1), Rn _m(2),-,-, Rn_mRead the values of (M) and read those values
And the sound on which noise is superimposed by the autocorrelation function calculating means.
R (0), R (1), R (2),-,-, R (M)
As a time variation value K of the noise determined as a certain function, R
(k), Rn_m(k) (k = 1,2,-,-, M) is a positive or negative value, respectively.
And take the positive and negative and R (k), Rn_mConsider the magnitude relation of (k)
And R (k), Rn_m(k) (k = 1,2,-,-, M) divided into multiple states
And calculated for each state using a different K formula for each state
From the minimum value of K, select the minimum value of K
Value.

By selecting K in this way, it is possible to remove noise in accordance with the noise during the current call.

The invention described in claim 4 is the same as the invention described in claim 3.
In the speech coding apparatus described above, the time variation rate value K of the noise
R (k), Rn _m (k) (k = 1, 2,-,-, M) take positive or negative values, respectively, and the positive and negative values of R (k), Rn _m (k) taking into account the magnitude _{relation, R (k), Rn m} (k) (k = 1,2, -, -, M) when _{classifying, R (k), Rn m} (k) (k = 1 , 2, -, -, and classified into the four states by positive and negative relationship M), further _{R (k), Rn m (} k) (k = 1,2,
-,-, M) have the same sign, R (k), Rn _m (k) are classified into two categories according to the magnitude relationship, and classified into a total of six states. From the minimum value of K calculated for each state,
Further, the minimum value of K is selected and set as the value of K.

By selecting K in this manner, it is possible to remove noise in accordance with the noise during the current call.

The invention described in claim 5 is the same as the invention in claim 4.
In the adaptive noise-suppressed speech coding apparatus as described, R (k),
Rn _m (k) (k = 1,2,-,-, M) is classified into six states, and the calculation formula of K different in each state is K shown in Table 1 below.
5. The adaptive noise-suppressed speech coding apparatus according to claim 4, wherein

[Table 1]

By selecting K in this way, it becomes possible to remove noise in accordance with the noise during the current call.

The invention described in claim 6 is the first invention.
Or the adaptive noise-suppressed speech coding apparatus according to claim 2.
The Rn stored in the first storage means_m(0), Rn
_m(1), Rn _m(2),-,-, Rn_mRead the values of (M) and read those values
And the sound on which noise is superimposed by the autocorrelation function calculating means.
R (0), R (1), R (2),-,-, R (M)
As a time variation value K of noise determined as a certain function,
Calculate K for k = 1,2,-,-, M by the following formula
And the minimum value is selected as K.
You. K = {b / a} [1- {1- (ca / b^Two)}^1/2(58) Here, a, b, and c have the following values. a = {R^Twon_m(0) −R^Twon_m(k)} (54) b = {R (0) * Rn_m(0) −R (k) * Rn_m(k)} (55) c = {R^Two (0) −R^Two (k)} (56) By selecting K in this way, the
Accordingly, noise can be removed.

The invention described in claim 7 is the first invention.
3. The adaptive noise suppression speech coding apparatus according to claim 2, wherein the autocorrelation function of the speech signal at a pitch period L is provided.
Means for calculating R (L), further comprising means for storing the value, Rn read from the noise autocorrelation function storage means
_m (0), Rn _m (1), Rn _m (2),-,-, the value of Rn _m (M) and the value of the autocorrelation function R (L) of the pitch period read from the storage means R (0), R (1), R calculated as the autocorrelation function of the voice signal on which the noise is superimposed by the autocorrelation function calculation means.
(2) A value K determined as a certain function with the values of-,-, and R (M) is defined as a time variation rate of noise.

As described above, by adding the estimation based on the autocorrelation function R (L) using the pitch period L to the selection of K, it is possible to adapt to the noise during the current call and accurately remove the noise. Become.

The invention according to claim 8 is the same as the invention according to claim 7.
In adaptive noise reduced speech coding apparatus according as the time variation rate K of the _{noise, R (k), Rn m} (k) (k = 1,2, -, -, M) is respectively positive or negative value And the sign plus and minus R (k), Rn
Considering the magnitude relation of _m (k), R (k), Rn _m (k) (k = 1, 2,-,
-, M) is classified into a plurality of states, and the autocorrelation of the pitch period and the value obtained by further selecting the minimum K value from the minimum value of K calculated for each state using a different K formula for each state The smaller value of the function R (L) and a value determined as a certain function is the noise time variation rate K.

As described above, by adding the estimation based on the autocorrelation function R (L) using the pitch period L to the selection of K, it is possible to remove the noise with high accuracy by adapting to the noise during the current call. Become.

The invention described in claim 9 is the same as the invention described in claim 8
In adaptive noise reduced speech coding apparatus according as the time variation rate K of the _{noise, R (k), Rn m} (k) (k = 1,2, -, -, M) is respectively positive or negative value And the sign plus and minus R (k), Rn
Considering the magnitude relation of _m (k), R (k), Rn _m (k) (k = 1, 2,-,
-, M) are classified into multiple states, R (k), Rn _m (k)
(K = 1,2, -, - , M) are classified into the four states by positive and negative relationship, further _{R (k), Rn m (} k) (k = 1,2, -, -, M) is In the case of the same code, two classifications are made according to the magnitude relation of R (k) and Rn _m (k), the classification is made into a total of six states, and K calculated for each state by a different K formula for each state. The smaller of the value of K and the value determined by the autocorrelation function R (L) of the pitch period, whichever is smaller, is defined as the time variation rate K of the noise.

As described above, by adding the estimation based on the autocorrelation function R (L) using the pitch period L to the selection of K, it is possible to remove the noise with high accuracy by adapting to the noise during the current call. Become.

[0081] Further, an invention according to claim 10, in the adaptive noise suppression speech coding apparatus according to claim 9, as a time variation rate K of the _{noise, R (k), Rn m} (k) (k = 1,2,-,-, M)
Takes positive or negative values, respectively, and the positive and negative values and R (k), Rn
Considering the magnitude relation of _m (k), R (k), Rn _m (k) (k = 1, 2,-,
-, M) are classified into a plurality of states, and are classified into six states, and the calculation formula of K different in each state uses a formula described in Table 1 shown in claim 5, and a pitch period. As a value determined by the autocorrelation function R (L) of, the minimum value is calculated after calculating K _L by the following equation, and K and K _L
Is set as the time variation rate K of the noise. _{K L = [R (0)} - R (L)] / Rn m (0) (67)

As described above, by adding the estimation based on the autocorrelation function R (L) using the pitch period L to the selection of K, it is possible to remove the noise with high accuracy by adapting to the noise during the current call. Become.

According to an eleventh aspect of the present invention, in the adaptive noise-suppressed speech coding apparatus according to the sixth or seventh aspect, the value of the time variation rate K of the noise is calculated according to the equation described in the sixth aspect. The value of K determined and the value of K _L determined as a function of the autocorrelation function R (L) of the pitch period according to claim 7.
Is selected as the value of K, whichever is smaller.

As described above, by adding the estimation based on the autocorrelation function R (L) using the pitch period L to the selection of K, it is possible to remove the noise with high accuracy by adapting to the noise during the current call. Become.

According to a twelfth aspect of the present invention, in the adaptive noise suppression speech coder according to the eleventh aspect, the value determined by the autocorrelation function R (L) of the pitch period is expressed by the following expression. Is used.

As described above, by adding the estimation based on the autocorrelation function R (L) using the pitch period L to the selection of K, it is possible to remove the noise with high accuracy by adapting to the noise during the current call. Become.

According to a thirteenth aspect of the present invention, in the adaptive noise-suppressed speech coding apparatus according to the first or second aspect, the determining means determines that the section is a speech signal section in which noise is superimposed. In this case, the value determined using the final value of the gain of the first fixed code gain adjusting means controlled by the data of the immediately preceding fixed section length is changed to the noise time variation rate K.
, Rs (0), Rs (1), Rs
(2), -, -, Rs (M) and _{R (0) -KRn m (0} ), R (1) -KRn m (1), R (2) -K
_{Rn m (2), -,} -, it is obtained as defined as _{R (M) -KRn m (M} ).

The speech coding apparatus is constructed and operated as described above, thereby eliminating noise in the internal operation of the speech coding apparatus, realizing an operation with an improved signal-to-noise ratio, and implementing the speech coding apparatus. This has the effect of preventing a malfunction due to noise of the digitizing device.

Further, according to a fourteenth aspect of the present invention, in the adaptive noise suppression speech coder according to the thirteenth aspect, the time variation rate K of the noise is controlled by data of a preceding fixed section length. 14. The adaptive noise-suppressed speech coding apparatus according to claim 13, wherein the value of K is obtained by using the following equation, assuming that the final value of the gain of the first fixed code gain adjusting means is g _a '. K = {g _a '} ² Y / {R _nm (0) + Σ _{k = 1} ^{k = M} _ak R _nm (k)} (80) where Y is the actual number of samples in the fixed codebook. , And a _k is a prediction coefficient of the linear prediction circuit.

The speech coding apparatus is constructed and operated as described above, thereby eliminating noise in the internal operation of the speech coding apparatus, realizing an operation with an improved signal-to-noise ratio, and implementing the speech coding apparatus. This has the effect of preventing a malfunction due to noise of the digitizing device.

[0091]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below with reference to FIGS.

(First Embodiment) The first embodiment of the present invention corresponds to the first, third, fourth, fifth and sixth aspects of the present invention. The present invention will be described with reference to a type, but the present invention can be applied to all speech coding apparatuses that use a signal autocorrelation function for short-term prediction of a speech signal.

FIG. 1 is a block diagram for explaining the first aspect of the present invention, and is a diagram showing functional blocks of an adaptive noise suppression speech coding apparatus or method according to the first embodiment.
In FIG. 1, the reference numerals 1301 to 1320 correspond to FIG.
This is the same as that shown in FIG. Illustrated numbers 101A, 101B, 102A to 102F, 103, 105, 106, 11
Items 6,117A to 117E and 118 to 120 are added to realize the present invention. Illustration number 101A is voice /
Noise section judgment circuit, 101B is voice / noise switch SW control circuit, 1
02A to 102F are voice / noise switching switches, 103 is an autocorrelation function storage circuit, 105 is a noise change rate estimation circuit, and 106 is an autocorrelation function estimation circuit.

The illustrated numbers 116, 117A to 117E and 118 to 120
30 are replaced with 1316, 1317A to 1317E and 1318 to 1320, respectively, and have the same names and functions as those of the same numbers in FIG. 30 described above, but have been added in the present invention in order to function differently from FIG. It was done. That is, the illustrated number 116 is a fixed codebook, 117A is a fixed codebook gain adjustment circuit, 117B is an impulse response circuit, 117C is an error calculation circuit, 117D is an error minimization circuit, 117E is a fixed codebook optimal value finding circuit, and 118 is fixed. A codebook gain quantization circuit, 119 is an excitation calculation circuit, and 120 is an optimum filter storage circuit.

The operation of the adaptive noise suppression speech coder according to the first embodiment configured as described above will be described below.

[Preprocessing circuit 1301], [Windowing circuit 1302],
[Autocorrelation Function Calculation Circuit 1303] Conventional Example As already described with reference to FIG. 30, the input signal including the ambient noise from the microphone is converted from an analog signal to a digital signal by these circuits and windowed. The autocorrelation function is calculated.

[Speech / Noise Section Judging Circuit 101A] A circuit for discriminating whether a signal of a fetched section is only a noise section or a section having speech, there are various known methods. In the present invention, the autocorrelation function calculation circuit 1303 calculates the autocorrelation function of the acquired signal.If these are R (0), R (1),-,-, R (M), R Since (0) represents the power of the captured signal, using this value, if R (0) exceeds a certain value, the period during which the audio signal is present,
If the value does not exceed a certain value, it is determined that the period is only a noise without a sound signal.

[Audio / Noise Switching SW Control Circuit 101B], [Audio / Noise Switching SW Control Circuit 101B]
Noise switch SW102A] to [voice / noise switch SW102F] The voice / noise switch SW control circuit 101B performs voice / noise determination based on a determination signal indicating whether a voice signal exists or is only a noise by the voice / noise interval determination circuit 101A. / Noise switch SW102A to voice / noise switch SW102F are controlled. In the present invention, the role of the operation of each means is different depending on whether a section where a voice signal exists or a section where only noise is present.
Switching is performed by W102A to voice / noise switch SW102F, and the switching control signal is transmitted to the voice / noise switch SW control circuit 1.
Supplied by 01B.

As described above, switching the same circuit for use has an effect of preventing an increase in circuit scale. Voice / noise switch SW102A to voice / noise switch SW102F are common contacts C and S
A common contact C is connected to an S contact in a section where an audio signal is present, and a common contact C is connected to a V contact in a section where only noise is present. The following description will proceed with a case where it is initially determined that the section is only a noise section. In this case, each of the common contacts C of the voice / noise switch SW102A to the voice / noise switch SW102F in FIGS. 1 and 2 is connected to the V contact. FIG. 3 shows a portion that is not used in the operation in the section including only noise. Hereinafter, description will be made with reference to FIG.

[Speech / Noise Switching SW102A], [Autocorrelation Function Storage Circuit 103] Conventional Example The autocorrelation functions R (0), R (1) in the noise section calculated by the autocorrelation function calculation circuit 1303 described in FIG. ), R (2),-,-, R
(M) to the storage device via the voice / noise switch SW102A, and Rn
_{m (0), R (1} ) n m, R (2) n m, -, -, it is stored as Rn _{m (M).}

[Speech / Noise Switching SW 102B], [Predictive Coefficient Calculation Circuit 1304] Conventional Example The noise / auto-correlation function calculated by the auto-correlation function calculation circuit 1303 described with reference to FIG.
Is input to the prediction coefficient calculation circuit 1304. The prediction coefficient calculation circuit 1304 obtains a noise prediction coefficient by using the Levinson-Durbin algorithm described above.

[Coefficient conversion circuit 1305] The coefficient conversion circuit 1305
The optimal prediction coefficients a ₁ , a ₂ ,
-, -, it converts the a ₁₀ is a frequency domain parameter, the coefficient called a line spectrum pair (LSP).

[Subframe coefficient interpolation circuit (before quantization) 1
308A] The subframe coefficient interpolation circuit (before quantization) 1308A obtains the LSPs of the first and third subframes by using the LSPs of the second and fourth subframes before quantization, and uses them. A (z), which is the transfer function of the linear prediction circuit for each subframe
Get.

[4 Subframe Weighted Speech Calculation Circuit 130
9] The four-subframe weighted speech calculation circuit 1309 passes a noise of four subframes (that is, one frame) before windowing through a filter of a transfer function in consideration of human auditory characteristics. The filter of the transfer function in consideration of the human auditory characteristics is represented by the above-described equation (20). W (z) = {A (z / γ ₁ )} / {A (z / γ ₂ )} (20) Here, γ ₁ = 0.9 and γ ₂ = 0.6. That is, since A (z) has been found by the subframe coefficient interpolation circuit (before quantization) 1308A, W (z) is also obtained.

[Coefficient quantization circuit 1306] Coefficient quantization circuit 1306
Converts the LSP thus obtained into an appropriate number of bits.

[Subframe coefficient interpolation circuit 1307] The subframe coefficient interpolation circuit 1307 calculates the LSP of the second subframe and the fourth subframe calculated as described above.
, LSP parameters of the first and third subframes are obtained. Further, using these values, a transfer function) A ′ (z) of the linear prediction circuit is obtained. A '(z) is used to distinguish it from A (z) which is a transfer function of a linear prediction circuit for a frame made by a subframe coefficient interpolation circuit (before quantization) 1308A.

[Impulse response calculation circuit 1308B], [voice /
Noise switch SW102D], [Speech / noise switch SW102F] The impulse response calculation circuit 1308B includes the transfer function A (z) of the linear prediction synthesis circuit obtained by the subframe coefficient interpolation circuit (before quantization) 1308A. Subframe coefficient interpolation circuit 13
Using A ′ (z) obtained in 07, h (n), which is an impulse response of the linear prediction synthesis circuit subjected to human auditory correction, is obtained, and via voice / noise switching SW102D and SW102F, h (n)
Is represented by equation (19) described above. H (z) = (1 / A ′ (z)) · (W (z)) = {A (z / γ ₁ )} / {A ′ (z) · A (z / γ ₂ )} (19) H (z) and A ′ (z) have an inverse relationship. A ′ (z) subtracts a predictable amount from the input audio signal and outputs a residual. The residual becomes a random pulse when prediction is properly performed. On the other hand, H (z) indicates that a residual, that is, a random pulse is input, an audio signal is output, and the output is obtained by correcting human auditory characteristics. Although a detailed method of obtaining h (n) is omitted, h (n) is obtained by the above-described inverse z-transform of H (z).

[Voice / Noise Switching SW102C], [Voice / Noise Switching S
W102E], [Fixed codebook 116] Fixed codebook 116
Is composed of a total of 10 single pulses divided into 5 tracks. The pulses are "+
It can take the values of "1" and "-1". The fixed codebook 116 generates a pulse train corresponding to noise.

[Fixed Gain Adjustment Circuit 117A] The fixed gain adjustment circuit 117A is a circuit for adjusting the amplitude of 10 pulses per subframe generated by the fixed codebook 1316, and passes its output to the impulse response circuit 117B. .

[Impulse Response Circuit 117B] The impulse response circuit 117B obtains h obtained by the impulse response calculation circuit 1308B.
The output pulse train of (n) and the fixed codebook 1316 via the fixed gain adjustment circuit 1317A is convoluted and integrated to be converted into a synthesized signal corresponding to a weighted audio signal, and is input to the error calculation circuit 117C.

[Error Calculation Circuit 117C] Impulse Response Circuit 11
The output of 7B is input to the error calculation circuit 117C, and the error calculation circuit
On the other side of 117C, the noise weighted by the 4-subframe weighted voice calculation circuit 1309 is also the voice / noise switch S.
Input via W102C and SW102E. The difference between the two is calculated, and the value is input to the error minimizing circuit 117D.

[Error minimizing circuit 117D] Error minimizing circuit 117D
Calculates the root-mean-square value of the error calculated by the error calculation circuit 117C, and controls the fixed codebook 116 and the fixed gain adjustment circuit 117A to minimize the value.

[Fixed Codebook Optimal Value Finding Circuit 117E] The fixed codebook optimal value finding circuit 117E is an error minimizing circuit.
As a result of controlling the fixed codebook 116 and the fixed gain adjustment circuit 117A so that 117D minimizes the mean square value of the error, the optimum value of the fixed codebook and the optimum value of the fixed gain are found.

[Fixed Codebook Gain Quantization Circuit 118]
The fixed codebook gain quantization circuit 118 quantizes the gain of the fixed codebook with an appropriate number of bits.

[Excitation Calculation Circuit 119] The excitation calculation circuit 119 converts the noise excitation pulse signal for each sub-frame into the following equation (26) by using the fixed codebook and the optimum value of the fixed gain created by the above-described matters. Create by u (n) = g _a C (n); n = 0,1,2,-,-,-, 39 (26) where u (n) is an excitation vector and C (n) is a fixed code The book's excitation vector, g _a is its gain.

[Optimal filter storage circuit 120] The optimal filter storage circuit 120 is h (n) which is the impulse response of the linear prediction synthesis circuit obtained by the impulse response calculation circuit 1308B.
And its initial conditions are stored in preparation for the next operation.

The operation of the apparatus of the present invention in the section where only the noise is present has been described above, and the operation in the section where the voice signal is present will be described. In this case, voice / noise switch SW102A ~ SW102F
Is connected to the common contact C and S contact. The state in this case is shown in FIGS. Hereinafter, an operation in a section in which an audio signal exists will be described with reference to FIGS.

[Autocorrelation Function Calculation Circuit 1303] Conventional Example FIG.
As described above, the autocorrelation function in the section where the speech signal exists calculated by the autocorrelation function calculation circuit 1303 is input to the noise change rate estimation circuit 105 and the autocorrelation function estimation circuit 106. Autocorrelation function calculation circuit 13
The autocorrelation function calculated in 03 is the noise change rate estimation circuit 10
5, and input to the autocorrelation function estimating circuit 106, but in the section of only noise, the noise change rate estimating circuit 105 and the autocorrelation function estimating circuit 106 are inactive.

[Speech / Noise Switching SW102A], [Autocorrelation Function Storage Circuit 103] The autocorrelation function Rn _m (0), R (1) of the noise stored in the noise section
_{n m, R (2) n} m, -, -, noise change ratio estimating circuit 105 Rn _m (M) are read out, and are input to the autocorrelation function estimation circuit 106.

[Noise change rate estimation circuit 105] As described above, the autocorrelation function R (0), R (1), R (2),-,-, R (M) is input, and the autocorrelation function R
n _m (0), R (1) _nm , R (2) _nm ,-,-, R _nm (M) are input.

The autocorrelation function of only the speech signal in the signal section in which noise is superimposed on the speech signal and the noise only, Rs (0),
Rn (0), Rs (k), Rn (k); k = 1, 2,-,-,-, M are unknown, but the following equations (27), (28) and (29) hold. R (0) = Rs (0) + Rn (0); (however, Rs (0) and Rn (0) are positive) (27) Rs (0)> | Rs (k) |; k = 1,2 ,-,-,-, M (28) Rn (0)> | Rn (k) |; k = 1,2,-,-,-, M (29)

Using this property, Rs (0) can be estimated. Below are some examples. <Example 1> Rs (0) = function ｛| R (1) |, | R (2) |,-,-,-, | R (M) |｝ (30) function ｛｝ is in {｛ Is defined as a function with the value of | R (1) | means taking an absolute value. Specifically, the following method is conceivable. <Example 1-a> Rs (0) = Max ｛| R (1) |, | R (2) |,-,-,-, | R (M) |｝ (31) Therefore, Rs (0) Is estimated, Rn (0) can be estimated by the following equation (32). Rn (0) = R (0) -Rs (0) (32) The noise change rate is expressed by the following equation (33). K = Rn (0) / Rn _m (0) (33)

It is assumed that the statistical properties of the noise included in the audio signal are considered to be the same as in the section where no audio signal exists. Then, the following equation (34) is established. Rn (k) = K * Rn m (k); k = 1,2 -, -, -, M (34)

If the value of K is not carefully estimated, not only can noise not be effectively removed, but also the circuit may become unstable due to oscillation. The following method is considered as a more accurate estimation. <Example 2> Rs (0) = function {R (1), R (2), -, -, R (M), Rn m (0), R (1) n m, -, -, Rn m ( M)} (35) <Example 2-a> As a specific example, the value of K is represented by the following equation (36). K = Min (K ₁ , K ₂ , K _3a , K _3b , K _4a , K _4b ) (36) Min means taking the minimum value, and K ₁ , K ₂ , K _3a , K _3b ,
K _4a and K _4b are determined by the following method. _{<K 1, K 2, K} 3a, K 3b, K 4a, how to determine the _{K 4b> R (k),} Rn
_m (k); k = 1, 2,-,-, and M are known and take positive or negative values, respectively. Therefore, there are four types of positive and negative, negative and positive, positive and positive, and negative and negative. There are combinations. For a more positive and positive, if negative and negative, | R (k) |> | Rn m (k) | with, | R (k) | < | Rn
_m (k) | is divided into the following six types in total. (1) R (k) is positive, Rn _m (k) is negative; k = 1,2, -, - , M (37) (2) R (k) is negative, Rn _m (k) is positive; k = 1,2, -, -, M (38) (3a) R (k) is positive, Rn _m (k) is _{positive: R (k)> Rn m} (k); k = 1,2, - , -, M (39) ( 3b) R (k) is positive, Rn _m (k) is _{positive: R (k) <Rn m} (k); k = 1,2, -, -, M (40) (4a) R (k) is negative, Rn _m (k) is negative: | R (k) |> | Rn m (k) |; k = 1,2, -, -, M (41) (4b) R (k) is negative, Rn _m (k) is negative: | R (k) | < | Rn m (k) |; k = 1,2, -, -, M (42)

First, (1) R (k) is positive, R _nm (k) is negative; k = 1,
Consider the case of 2,-,-, M. Unknown Rs (0), Rs (k)
Is given by the following equation (43) if K ₁ is correctly selected. R (k) = Rs (k) −K ₁ | R _nm (k) | (43) R (0) = Rs (0) + K ₁ R _nm (0) (44) Equations (28) and (29) By using the relationship, the following equation (45) is established. _{R (0) -K 1 Rn m} (0)> = R (k) + K 1 | Rn m (k) | (45)

In the above,> = means greater or equal. The following equation (46) is satisfied when the equality holds the maximum value of K _{1. R (0) -K 1 Rn m} (0) = R (k) + K 1 | Rn m (k) | (46)

[0127] accordance connexion K ₁ is obtained by the following equation _{(47) K 1 = {R} (0) -R (k)} / {Rn m (0) + | Rn m (k) |} (47) actually Is (1) R (k) is positive, R _nm (k) is negative; k = 1,2,-,-, M
Is calculated for all k that satisfies and the minimum value is K ₁
And

When the same calculation is performed for the other five cases, Table (1) is obtained.

[Table 2] <Example 2-b> The following example is described as another specific example. R (0) = Rs (0 ) + KRn m (0) (48) R (k) = Rs (k) + KRn m (k) k = 1,2, -, -, -, M (49) Rs (0)> = | Rs (k) | k = 1,2,-,-,-, M (50) R (0) −KRn _m (0)> = | R (k) −KRn _m (k) (51) By squaring both sides of equation (51), the following equation is obtained. {R (0) −KR _nm (0)} ² > = {R (k) −KR _nm (k)} ² (52) By rearranging assuming that the equality holds, the following equation is obtained. aK ² -2bK + c = 0 (53) Here, a, b, and c have the following values. a = {R ² _nm (0) −R ² _nm (k)} (54) b = {R (0) * R _nm (0) −R (k) * R _nm (k)} (55) c = {R ² (0) −R ² (k)} (56)

When K is obtained using the above equations (54) to (56), the following equation is obtained.
(57) and (58) are obtained. K ₁ = {b / a} [1+ {1− (ca / b ² )} ^1/2 ] (57) K = {b / a} [1− {1− (ca / b ² )} ^1/2 ] (58) as the value of K, selecting a is smaller than K ₁ (58).

The present invention also includes a method of multiplying the value of K obtained by using any of the various methods described above by an empirically obtained safety coefficient (smaller than 1) to obtain K. The value of K obtained by the above method is passed to the autocorrelation function estimation circuit 106.

[Autocorrelation function estimating circuit 106] The autocorrelation function estimating circuit 106 receives the K
Is used to estimate the autocorrelation function of the noise mixed in the current speech signal as follows. Rn (j) = K * Rn m (j); j = 1,2, -, -, M (59)

Next, the autocorrelation function of only the audio signal, Rs
(0), Rs (1), Rs (2),-,-, Rs (M) are calculated by the following equation (60). Rs (j) = R (j ) -K * Rn m (j); j = 1,2, -, -, Rs obtained by M (60) or (0), the value of Rs (j), The signal is passed to the prediction coefficient calculation circuit 1304 via the voice / noise switch SW102B.

[Speech / Noise Switching SW102B], [Prediction Coefficient Calculation Circuit 1304] The prediction coefficient calculation circuit 1304 uses the autocorrelation function of only the noise-free audio signal to obtain the optimum prediction coefficient a by the Levinson-Durbin algorithm. _{1, a 2, -, -} , seeking a _{1 0.}

[Coefficient conversion circuit 1305] The coefficient conversion circuit 1305
The optimal prediction coefficients a ₁ , a ₂ ,
-, -, a ₁₀ is a parameter in the time domain (the coefficient of the impulse response), which is a parameter of the frequency domain,
Convert to another coefficient called a line spectrum pair (LSP).

[Subframe coefficient interpolation circuit (before quantization) 1
308A] The subframe coefficient interpolation circuit (before quantization) 1308A obtains the LSP parameters of the first and third subframes using the LSP parameters of the second and fourth subframes before performing the quantization, Is used to obtain A (z) which is a transfer function of the linear prediction circuit.

[4 Subframe Weighted Speech Calculation Circuit 130
9] The 4-subframe weighted audio calculation circuit 1309 passes the audio of the 4 subframes (that is, one audio signal) of the digital audio signal before windowing through a filter of a transfer function in consideration of human auditory characteristics. The filter of the transfer function in consideration of the human auditory characteristics is represented by the above-described equation (20). W (z) = {A (z / γ ₁ )} / {A (z / γ ₂ )} (20)

[Coefficient quantization circuit 1306] Coefficient quantization circuit 1306
Converts the LSP thus obtained into an appropriate number of bits.

[Subframe coefficient interpolation circuit 1307] The subframe coefficient interpolation circuit 1307 (after quantization) uses the LSP values of the second subframe and the fourth subframe to calculate the first and third subframes. Find the LSP parameter of. Further, using these values, A '(z) which is a transfer function of the linear prediction circuit is obtained.

[Voice / Noise Switching SW102C], [Open Loop Pitch Delay Finding Circuit 1310A], [Open Loop Pitch Gain Finding Circuit 1]
310B] The open loop pitch delay finding circuit 1310A and the open loop pitch gain finding circuit 1310B estimate the pitch period, that is, the delay time and the gain once every 10 ms, that is, twice every 20 ms, which is one frame of the audio signal. Pitch delay and pitch gain optimum values are selected as values that maximize Rlag (open),
The delay time is passed to the adaptive codebook 1311A, and the pitch gain is passed to the pitch gain adjustment circuit 1311B.

[Impulse Response Calculation Circuit 1308B] The impulse response calculation circuit 1308B includes A (z), which is the transfer function of the linear prediction circuit, obtained by the subframe coefficient interpolation circuit (before quantization) 1308A, and the subframe coefficient. Using A ′ (z) obtained by the interpolation circuit 1307 (after quantization), h (n), which is an impulse response of a linear prediction synthesis circuit corrected for human hearing,
Ask for. The z-transform of h (n) is represented by the above-described equation (19). H (z) = (1 / A ′ (z)) · (W (z)) = {A (z / γ ₁ )} / {A ′ (z) · A (z / γ ₂ )} (19) This h (n) is necessary for searching the adaptive codebook and the fixed codebook.

[Adaptive Codebook 1311A], [Pitch Gain Adjustment Circuit 1311B] The initial values of the adaptive codebook 1311A and pitch gain adjustment circuit 1311B are obtained by the open loop pitch delay finding circuit 1310A and open loop pitch gain finding circuit 1310B. The obtained pitch delay time and pitch gain are used in adaptive codebook 1311A and pitch gain adjustment circuit 1311B, respectively.

[Voice / Noise Switching SW102D], [Impulse Response Circuit 1311C] The impulse response circuit 1311C is an impulse response calculation circuit 130.
8 (b) and the output pulse of the adaptive codebook 1311A via the pitch gain adjustment circuit 1311B are convoluted and integrated to convert them into a synthesized signal corresponding to a weighted audio signal, which is then sent to an error calculation circuit 1311D. Output.

[Error Calculation Circuit 1311D] Error Calculation Circuit 1311D
Is a weighted voice signal calculated by the 4-subframe weighted voice calculation circuit 1309 and the impulse response circuit 1311
The difference of the synthesized signal corresponding to the weighted audio signal which is the output of C, that is, the error is calculated and output to the error minimizing circuit 1311E.

[Error minimizing circuit 1311E] Error minimizing circuit 131
1E finds the mean square value of the error calculated by the error calculation circuit 1311D, and adapts the adaptive code book 13 so as to minimize the value.
11A and the pitch gain adjustment circuit 1311B. The vicinity of the pitch delay time and the pitch gain obtained by the open loop pitch delay finding circuit 1310A and the open loop pitch gain finding circuit 1310B are precisely searched.

[Optimal delay and gain finding circuit 1312] The optimal delay and gain finding circuit 1312 finds the optimum delay time and gain value as a result of the control of the error minimizing circuit 1311E. The obtained values of the adaptive codebook and the pitch gain are respectively stored, and the optimum values of the next adaptive codebook and the pitch gain are searched using the stored values.

[Pitch gain quantization circuit 1313] The pitch gain quantization circuit 1313 quantizes the optimum pitch gain found by the optimum delay and gain finding circuit 1312 with an appropriate number of bits.

[Applied Codebook Contribution Calculation Circuit 1314] The applied codebook contribution calculation circuit 1314 is an error signal of the error calculation circuit 1311D obtained by using the impulse train based on the value obtained by the optimum delay and gain finding circuit 1312. Is output to the error calculation circuit 117C via the voice / noise switch SW102E.

[Fixed Code Book 116] Fixed Code Book 1
16 is made up of a combination of a total of 10 single pulses divided into 5 tracks, with the pulses being “+1” and “−”.
The initial value of these pulses is the optimum value set in the operation in the noise section.

[Fixed Gain Adjustment Circuit 117A] The fixed gain adjustment circuit 117A is a circuit for adjusting the amplitude of eight pulses per subframe generated by the fixed codebook 116, and passes its output to the impulse response circuit 117B. The initial value is the optimum value set in the operation in the noise section.

[Optimal filter storage circuit 120] The optimum filter storage circuit 120 stores the following impulse response (h (n)) of the linear prediction synthesis circuit corrected by the human auditory sense obtained by the impulse response calculation circuit 1308B. It is stored in preparation for the operation. The optimum filter storage circuit 120 outputs the impulse response h (n) of the linear prediction synthesis circuit stored in the operation in the noise section and its initial condition to the impulse response circuit 117B.

[Impulse Response Circuit 117B] The impulse response circuit 117B is an impulse response corresponding to h (n), which is the impulse response of the linear predictive synthesis circuit with human auditory correction output from the optimal filter storage circuit 120. A h
v (n) (where the suffix v represents noise) and the fixed codebook 116 and the fixed code pulse train synthesized by the fixed gain adjustment circuit 117A are convoluted and integrated to convert into a synthesized signal corresponding to weighted noise. Error calculation circuit 117C
To enter.

[Error Calculation Circuit 117C] The error calculation circuit 117C
The error signal from the error calculation circuit 1311D in a state where the root mean square error of the error is minimized by the error minimization circuit 1311E is input. The other input is a fixed codebook 11
6.Calculate the difference between the fixed code pulse train synthesized by the fixed gain adjustment circuit 117A and the signal converted to the synthesized signal of weighted noise by the impulse response circuit 117B, and output the value to the error minimization circuit 117D. .

[Error minimizing circuit 117D] Error minimizing circuit 117D
Calculates the root-mean-square value of the error calculated by the error calculation circuit 117C, and controls the fixed codebook 116 and the fixed gain adjustment circuit 117A to minimize the value. Although the details of the method are omitted, the noise component is largely removed by the operation here, while the audio signal component remains as it is, and the error calculation circuit
The residual component at 117C is input to error calculation circuit 1317C.

[Fixed Codebook Optimal Value Finding Circuit 117E] The fixed codebook optimal value finding circuit 117E is an error minimizing circuit 11.
As a result of controlling the fixed codebook 116 and the fixed gain adjustment circuit 117A so that 7D minimizes the root mean square value of the error, the optimum value of the fixed codebook and the optimum value of the fixed gain are found. Each of these values is a value corresponding to noise.

[Fixed Codebook Gain Quantization Circuit 118]
The fixed codebook gain quantization circuit 118 quantizes the gain of the fixed codebook with an appropriate number of bits.

[Excitation Calculation Circuit 119] The excitation calculation circuit 119 uses the fixed codebook and the optimum value of the fixed gain created as described above as the excitation pulse signal of the synthesized noise for each sub-frame as Created according to 61). u (n) = g _a C (n); n = 0,1,2,-,-,-, 39 (61) where u (n) is an excitation vector and C (n) is a fixed codebook , The excitation vector, g _a is its gain. All correspond to noise.

[Optimal filter storage circuit 120] The optimal filter storage circuit 120 stores h (n), which is the impulse response of the linear predictive synthesis circuit corresponding to noise, and its initial condition in preparation for the next operation.

[Fixed Code Book 1316] Fixed Code Book
1316, those pulses made of a combination of a total of 10 single pulses divided into 5 tracks,
It can take the values of "+1" and "-1".

[Fixed Gain Adjustment Circuit 1317A] The fixed gain adjustment circuit 1317A is a circuit for adjusting the amplitude of 10 pulses per subframe generated by the fixed codebook 1316, and outputs its output to the impulse response circuit 1317B.

[Impulse Response Circuit 1317B] The impulse response circuit 1317B obtains h obtained by the impulse response calculation circuit 1308B.
(n) and the output pulse train of the fixed codebook 1316 via the fixed gain adjustment circuit 1317A are converted into a synthesized signal corresponding to a weighted audio signal obtained by convolution and integration, and input to the error calculation circuit 1317C.

[Error Calculation Circuit 1317C] The difference between the error signal output of the error calculation circuit 117C and the output of the impulse response circuit 1317B is calculated, and the value is output to the error minimization circuit 1317D.

[Error minimizing circuit 1317D] Error minimizing circuit 131
7D calculates the mean square value of the error calculated by the error calculation circuit 1317C, and fixes the fixed code book 13 so as to minimize the value.
16 and the fixed gain adjustment circuit 1317A.

[Fixed codebook optimum value finding circuit 1317E]
The fixed codebook optimum value finding circuit 1317E controls the fixed codebook 1316 and the fixed gain adjustment circuit 1317A so that the error minimizing circuit 1317D minimizes the mean square value of the error. To find the optimal value of.

[Fixed codebook gain quantization circuit 1318]
The fixed codebook gain quantization circuit 1318 quantizes the gain of the fixed codebook with an appropriate number of bits necessary for wireless transmission.

[Excitation Calculation Circuit 1319] The excitation calculation circuit 1319 calculates the optimum values of the adaptive codebook and the pitch gain and the optimum values of the fixed codebook and the fixed gain for each sub-frame. The excitation pulse signal of the synthesized speech is generated by the above-described equation (25). u (n) = g _p v (n) + g _a C (n); n = 0,1,2,-,-,-, 39 (25) where, u (n) is an excitation vector, v (n) is the excitation vector of the adaptive codebook, g _p is its gain, C (n) is the excitation vector of the fixed codebook, g _a
Is the gain. C (n) and g _a correspond to speech from which noise has been removed.

[Optimal filter storage circuit 1320] The optimal filter storage circuit 1320 stores h (n), which is the impulse response of the linear prediction / synthesis circuit corresponding to the audio signal, and its initial condition in preparation for the next operation. .

As described above, according to the first embodiment of the present invention, the short-term prediction coefficient calculation circuit, coefficient conversion circuit, and coefficient quantization circuit of the speech signal existing in the speech coding apparatus are used to remove noise. Autocorrelation functions Rs (0), Rs (1), Rs
(2) By operating using-,-, Rs (M), a speech coding apparatus capable of reducing ambient noise is made possible. As described above, it has become clear that an adaptive speech coding apparatus or method can be realized that adapts to the amount of noise currently contained in the speech signal and removes it.

As described above, the first embodiment of the present invention
This is a faithful application to the ACELP speech coding device described in the GSM mobile phone technical standard. One of the objects of the present invention is that when a speech signal containing noise is input to a speech encoding device, the speech encoding device does not operate properly when the noise level increases, and the signal-to-noise ratio may be further deteriorated. Therefore, it is an object of the present invention to eliminate noise in the internal operation of the speech coding apparatus, to operate the speech coding apparatus stably, and to prevent deterioration of the signal-to-noise ratio.

To explain the principle of the present invention in an easy-to-understand manner, FIG.
FIG. 6 is a simplified diagram of FIG. 5 in which parts irrelevant to the claims are deleted. Hereinafter, description will be made with reference to FIG.

Each part is broadly grouped into blocks A, B, C and D. The output of each block is shown as output A, output B, output C, and output D, respectively.

The block A has a function of removing a noise component from an input signal, and a signal from which a noise component has been removed is output to an output A. The block B has a function of removing the pitch component of the audio signal from the input signal, and the output B outputs a signal from which the pitch component has been removed. The block C has a function of removing a predictable component of the audio signal from the input signal, and the output C outputs a signal from which the predictable component of the audio signal has been removed. The block D has a function of estimating noise included in the input signal and the like, and an output D outputs a sound signal including noise corrected in consideration of human hearing.

Since the conventional speech coding apparatus does not have the block A, it does not have a function of removing a noise component from the input signal. Therefore, the noise contained in the input signal is directly output to the output C of the block C. appear. Therefore, when the noise becomes large, there is a risk that the error minimizing circuit 1317D becomes inoperable.

According to the present invention, the block A has a function of removing a noise component from an input signal, and no noise component is output at the output A. As a result, no noise appears at the output C of the block C. . Therefore error minimizing circuit 1317D
Is not in danger of becoming inoperable, and the speech coding apparatus of the present invention can operate without being affected even if the input signal contains noise.

As is clear from FIG. 6, the blocks are arranged in the order of block B, block A and block C, but they can be arranged in the order of block A, block B and block C. FIG. 7 is a diagram for easily explaining the principle of the present invention in this case. The contents of the operation are self-evident from the already described matters and will not be described. An advantage of this case is that since an audio signal from which noise has been removed is input to the input signal of the block B having a function of removing the pitch component, the operation of searching for the optimum value of the pitch component can be performed without being disturbed by the noise. It is in.

The block B for removing the pitch component is important as a speech coding apparatus, but is not included in the first aspect. The reason for this is that if it is described in such a manner, the sentence will be too long and the meaning will be unclear, and it is publicly known and unnecessary that the speech coding apparatus has a part for predicting the pitch component. Therefore, it is only pointed out that the contents shown in FIG. 7 are also included in the present invention.

(Second Embodiment) A second embodiment of the present invention is for describing the second aspect of the present invention, and is shown in FIGS. 8, 11, 12, 13, and 14. This will be described with reference to FIG.

FIG. 8 is a diagram for explaining the principle of the second embodiment of the present invention in an easy-to-understand manner. FIG. 13 is a detailed diagram for explaining the operation in a section where only noise exists, and FIG. 14 and FIG. FIG. 11 and FIG. 12 are diagrams illustrating the operation of a section of an audio signal including noise, and FIGS. 11 and 12 are diagrams in which the operation of the section of an audio signal including noise and the operation of the section of an audio signal including noise are integrated using a voice / noise switch. In FIGS. 11 and 12, what is newly added to FIGS. 1 and 2 is only the voice / noise switch SW102G and the voice / noise switch SW102H.

As is clear from FIG. 6 or FIG. 7, the outputs of the blocks A and B are input to the next block and to the respective error minimizing circuits. In contrast to this, in FIG. 8, the outputs of blocks A and B are:
Each is only input to the next block, and is not input to each error minimizing circuit. The output C of the final block, block C, is input to each error minimizing circuit.

With this configuration, the same output C as the error minimizing circuit of the final block, block C, is input to the error minimizing circuits of blocks A and B, so that the error minimizing operation can be performed. Unnecessary components are removed, and extremely stable and quick operation becomes possible.
The details of the operation are obvious from the description of the first embodiment, and will not be repeated.

(Third Embodiment) A third embodiment of the present invention corresponds to claims 7, 8, 9, 10, 11, and 12, and corresponds to a pitch period L of an audio signal. The autocorrelation function R (L) of the delay time is used in addition to the estimation of the time variation rate of the noise.

FIG. 16, FIG. 17, FIG. 18, and FIG. 19 explain the invention according to claims 7, 8, 9, 10, 11, and 12, and show an adaptive type of the third embodiment of the present invention. It is a functional block diagram of a noise suppression speech coding device. 16 and FIG.
7 corresponds to FIGS. 1, 2, 4 and 5 of the first embodiment, respectively, and FIGS. 18 and 19 are for explaining the operation in a section where noise is superimposed on the voice.

In FIGS. 16 and 17, only the pitch autocorrelation function storage circuit 107 is newly added to FIGS. The operation of the adaptive noise suppression speech coding apparatus configured as described above will be described with reference to FIGS. 16, 17, 18, and 19. FIG.

The operation in the noise-only section is exactly the same as in the case of FIG. 3 described in the first embodiment, and will not be described.

The operation in a section where a voice signal is present will be described. In this case, the voice / noise switch SW102A is used in FIGS.
The common contact C and the S contact are connected to SW102F. The state in this case is shown in FIGS.

[Autocorrelation Function Storage Circuit 103], [Speech / Noise Switching SW 102A], [Autocorrelation Function Storage Circuit 103] The above operations are the same as those described in the first embodiment.

[Pitch autocorrelation function storage circuit 107] The pitch autocorrelation function storage circuit 107 stores the loop pitch in the operation of the open loop pitch delay finding circuit 1310A and the open loop pitch gain finding circuit 1310B in the immediately preceding subframe. The discovery of delay and pitch gain is based on the previously described relational expression, Rlag
(open) = Σ _{n = 0} ^{n = 79} sw (n) sw (nk); n = 0,-,-, 79
The maximum value is stored based on (21), and the value is passed to the noise change rate estimation circuit 105 and the autocorrelation function estimation circuit 106 as R (L). FIG. 20 shows an example of the pitch autocorrelation function R (L).

[Noise Change Rate Estimation Circuit 105] and [Autocorrelation Function Estimation Circuit 106] Conventional Example The autocorrelation function R (0), R (0) of the speech signal containing noise is obtained from the autocorrelation function calculation circuit 1303 described with reference to FIG. 1), R (2),-,
-, R (M) is input, and from the autocorrelation function storage circuit 103,
Autocorrelation function Rn _m (0), R (1) _nm , R (2) _nm ,-,-, Rn
_m (M) is input, and the pitch autocorrelation function storage circuit 107
, A pitch autocorrelation function R (L) is input.

The autocorrelation function of only the speech signal in the signal section in which noise is superimposed on the speech signal and the noise only, Rs (0),
Rn (0), Rs (k), Rn (k); k = 1, 2, 2,-,-, M is unknown. Rs (0)> R (L) (62) Strictly speaking, Rs (0) in equation (28) should be R (0), but the autocorrelation function of noise at delay time L may be considered to be zero. Therefore, there is no problem in practical use. Utilizing these properties,
Rs (0) can be estimated. Below are some examples.

<Example A> Rs (0) = function ｛| R (1) |, | R (2) |,-,-,-, | R (M) |, R (L)｝ (63) function ｛｝ means to be defined as a function with a value in｝ ｛. | R (1) | means taking an absolute value. Specifically, the following method is conceivable. <Example A-a> Rs (0) = Max ｛| R (1) |, | R (2) |,-,-,-, | R (M) |, R (L)｝ (64) , Rs (0) is estimated, then Rn (0) can be estimated by the aforementioned equation (32). Rn (0) = R (0) -Rs (0) (32) The noise change rate is expressed by the above-described equation (33). Statistical properties of noise included in the K = Rn (0) / Rn m (0) (33) the audio signal is to be considered the same as when the section having no audio signal. Then, the aforementioned equation (34) is established. Rn (k) = K * Rn m (k); k = 1,2 -, -, -, the following methods can be considered as a precise estimation than M (34). <Example B> Rs (0) = function {R (1), R (2), -, -, R (M), Rn m (0), R (1) n m, -, -, Rn m ( M), R (L)｝ (65) <Example Ba> As a specific example, the value of K is expressed by the following equation (66). K = Min (K ₁ , K ₂ , K _3a , K _3b , K _4a , K _4b , K _L ) (66) Min means taking the minimum value, and K ₁ , K ₂ , K _3a , K _3b ,
K _4a, K _4b, K _L is determined by the following method.

<How to determine K ₁ , K ₂ , K _3a , K _3b , K _4a , K _4b , and K _L >
_{K 1, K 2, K 3a} , K 3b, K 4a, how to determine the K _4b will be described only method of determining the K _L for detailed above in the first embodiment. The value determined by the pitch period autocorrelation function _{R (L), K L =} [R (0) - R (L)] / Rn m (0) (67) employs a noise either small value Is assumed to be the time variation rate K.

<Example Bb> The following example will be described as another specific example. K is obtained by the method described in <Example 2-b>, and its value and the autocorrelation function R of the pitch period are calculated.
There is a method of setting the smaller one of the following values determined by (L) as the time variation rate K of noise. K _L = [R (0) −R (L)] / Rn _m (0) (67) The autocorrelation function estimating circuit 106 uses the value of K obtained by the above method to calculate the current speech signal. The autocorrelation function of the mixed noise is estimated as follows. Rn (j) = K * Rn m (j); j = 1,2, -, -, M (59) Next, the autocorrelation function of the speech signal only, Rs (0), Rs ( 1), Rs
(2) Calculate-,-, Rs (M) by the following formula. Rs (j) = R (j ) -K * Rn m (j); j = 1,2, -, -, Rs obtained by M (60) or (0), the value of Rs (j), The signal is passed to the prediction coefficient calculation circuit 1304 via the voice / noise switch SW102B.

[Speech / Noise Switching SW 102B], [Prediction Coefficient Calculation Circuit 1304] The subsequent operations are exactly the same as in the first embodiment, and will not be described.

As described above, according to the third embodiment of the present invention, in the first or second embodiment, the short-term prediction coefficient calculation of the speech signal existing in the speech encoding device is performed. Circuit, coefficient conversion circuit, coefficient quantization circuit, the autocorrelation function Rs (0), Rs (1), Rs
(2) By operating using-,-, Rs (M), a speech coding apparatus capable of reducing ambient noise is made possible.

As described above, it has become clear that an adaptive speech coding apparatus or method can be realized which adapts to the amount of noise currently contained in the speech signal and removes it.

(Fourth Embodiment) FIGS. 21, 22, 23 and 24 illustrate the invention according to claim 13, and FIGS. 21 and 22 show a fourth embodiment. FIG. 23 is a functional block diagram illustrating an operation of a speech signal section including noise of the adaptive noise suppression speech coding apparatus corresponding to the first embodiment. FIGS. 23 and 24 illustrate the fourth embodiment. FIG. 9 is a functional block diagram illustrating an operation of a speech signal section including noise of the adaptive noise suppression speech coder according to the second embodiment.

In FIG. 21, FIG. 22, FIG. 23, and FIG. 24, all the reference numerals are the same as those in FIG. 1 and FIG. 2 used in the first embodiment, and will not be described because they have already been described.

The operation of the adaptive noise suppression speech coder according to the first embodiment configured as described above will be described with reference to FIGS. 21 and 22. Most of the operations are performed using the first operation described with reference to FIGS. 1, 2, 3, 4, and 5.
This embodiment is the same as that of the first embodiment, and only the differences will be described below.

The operation in the noise-only section is the same as that already described with reference to FIG. Next, an operation in a section in which a voice signal is present will be described with reference to FIGS.

[Autocorrelation function storage circuit 103] The autocorrelation function in the section where the voice signal exists calculated by the autocorrelation function storage circuit 103 is input to the autocorrelation function estimation circuit 106.

[Speech / Noise Switching SW102A], [Autocorrelation Function Storage Circuit 103] The autocorrelation function Rn _m (0), R (1) n of the noise stored in the noise section
_{m, R (2) n m} , -, -, Rn m (M) is input read and the autocorrelation function estimation circuit 106.

[Noise change rate estimating circuit 105] The final value of the gain of the first fixed code gain adjusting means 117A controlled by the data of the immediately preceding fixed section length is the noise change rate estimating circuit 105.
Is input to A value determined as a function of this value is determined as a time variation rate K of noise.

[Autocorrelation function estimating circuit 106] The autocorrelation function calculating circuit 1303 outputs the autocorrelation functions R (0), R (1), R (2),-,-, R ( M) is inputted, and the autocorrelation function Rn _m (0),
_{R (1) n m, R} (2) n m, -, -, Rn m (M) are input.

Using these values and the value of K determined by the noise change rate estimation circuit 105, the autocorrelation function of only the speech signal, Rs
(0), Rn (0) , Rs (j), Rn (j); j = 1,2, -, -, -, and M, Rs (j) = R (j) -K * Rn m (j ); Estimate as j = 1,2-,-,-, M (60).

The values of Rs (0) and Rs (j) obtained as described above are input to the prediction coefficient calculation circuit 1 via the voice / noise switch SW102B.
Pass to 304.

[Speech / Noise Switching SW102B], [Prediction Coefficient Calculation Circuit 1304] The prediction coefficient calculation circuit 1304 employs the Levinson-Durbin algorithm using the autocorrelation function of only the noise-free speech signal to obtain the optimum prediction coefficient. a ₁ , a ₂ ,-,-, a ₁₀ are obtained. The following operation is the same as that described in the first embodiment, and a description thereof will be omitted.

Next, the operation of the adaptive noise suppression speech coder of the fourth embodiment corresponding to the second embodiment in the speech signal section containing noise will be described with reference to FIGS. 23 and 24. However, since it is the same as the description of the operation of the adaptive noise suppression speech coding apparatus corresponding to the first embodiment in the fourth embodiment described with reference to FIGS.

Next, the operation of the noise change rate estimating circuit 105 is described as follows: "The final value of the gain of the first fixed code gain adjusting means 117A controlled by the data of the immediately preceding fixed section length is the noise change rate estimating circuit The value which is input to the circuit 105 and is determined as a function of this value is determined as the time variation rate K of noise. "

The prediction error e (n) is represented by the above-mentioned equation (3). e (n) = s '( n) -s "(n) = Σa k s'(nk); k =, 0,1,2,3 ,, - ,, 10; a 0 = 1 (3) Here, let e (n) = Δu (n) (68), where u (n) is an impulse and its amplitude is 1.
Δ takes a positive or negative value. e (n) * e (n ) = Δ 2 = Σ k = 0 k = M a k s '(nk) * Σ j = 0 j = M a j s' (n-j) = Σ k = 0 k ^{= M} Σ _{j = 0} ^{j = M} a _k a _j s '(nk) * s' (n-j) = Σ _{k = 0} ^{k = M} Σ _{j = 0} ^{j = M} a _k a _j R (k-j ) (69) e (n) * e (n) = Δ ² = Σ _{k = 0} ^{k = M} _ak Σ _{j = 0} ^{j = M} a _j R (k-j) (70) From equation (8), the following equation (71) holds. Σ _{j = 0} ^{j = M} a _j R (k−j) = − R (k) (71) By substituting equation (71) into equation (70), the following equation (72) is obtained. e (n) * e (n ) = Δ 2 = Σ k = 0 k = M a k R (k) = {R (0) + Σ k = 1 k = M a k R (k)} (72) Therefore, the following equation (73) is obtained. e (n) = Δu '( n) = {R (0) + Σ k = 1 k = M a k R (k)} 1/2 u' (n) (73) i.e., the impulse u '(n) Δ = {R (0) + Σ _{k = 1} ^{k =} _Mak R (k)} ^1/2 (74)

Here, considering that the fixed codebook 116 issues 10 pulses for 40 sample points, u ′
Since it has 1/4 of the power of (n), it is necessary to correct it. The following equation (61) has already been described for the gain of the fixed gain adjustment circuit 117A. u (n) = g _a C (n); n = 0,1,2,-,-,-, 39 (61) In consideration of the above, the following equation (75) is obtained. g _a = 2 * {R (0) + Σ _{k = 1} ^{k = M} a _k R (k)} ^1/2 (75)

When the section shifts from the section containing only noise to the section containing an audio signal containing noise, the value set in the fixed gain adjustment circuit 117A first is the value shown in the aforementioned equation (63). g _a = 2 * in _{{Rn m (0) + Σ} k = 1 k = M a k Rn m (k)} 1/2 (76) , and optimized operation is completed, the audio signal containing noise It becomes another value adapted to the noise in the section and the following equation (77)
Holds. g _a '= 2 * {KR nm _m (0) + Σ _{k = 1} ^{k = M} _ak KR nm _m (k)} ^1/2 (77) That is, the following equations (78) and (79) hold. K = {g _a '/ g _a } ² (78) K = {g _a '} ² / [4 * {Rn _m (0) + Σ _{k = 1} ^{k = M} _ak R _nm (k)}] ( 79) If the ratio of the actual number of pulses to the total number of samples in the fixed codebook 116 is Y, the following equation (80) is obtained. K = {g _a '} ² Y / {R _nm (0) + Σ _{k = 1} ^{k = M} _ak R _nm (k)} (80) In this way, the noise change rate K can be determined.

As described above, it has become clear that an adaptive speech coding apparatus or method can be realized which adapts to the amount of noise currently contained in the speech signal and removes it.

[0212]

As described above, according to the present invention, the autocorrelation function of the speech signal and the autocorrelation function of the noise have different characteristics and the autocorrelation function of the signal in which the speech signal and the noise are superimposed. The signal processing is performed using the features, the noise suppression operation is performed in accordance with the level of the noise included in the currently-talking frame, and the autocorrelation function of the voice signal including the noise is calculated by the voice coding apparatus. In addition, since it is also used in the process of speech encoding processing, by integrating the noise suppression device and the speech encoding device, and by adding noise encoding means having the same function as the speech encoding device, There is an effect that it is possible to realize an adaptive noise suppression speech coding apparatus which does not malfunction even when excessive noise is added to the speech coding apparatus and is adapted to noise during a call.

[Brief description of the drawings]

FIG. 1 is a left half of a functional block diagram of an adaptive noise suppression speech coding apparatus according to a first embodiment of the present invention;

FIG. 2 is a right half part of a functional block diagram of the adaptive noise suppression speech coding apparatus according to the first embodiment of the present invention;

FIG. 3 is a functional block diagram (an operation explanatory diagram of a noise section) of the adaptive noise suppression speech coding apparatus according to the first embodiment of the present invention;

FIG. 4 is a left half part of a functional block diagram (an operation explanatory diagram of a speech section including noise) of the adaptive noise suppression speech coder according to the first embodiment of the present invention;

FIG. 5 is a functional block diagram of the adaptive noise-suppressed speech coding apparatus according to the first embodiment of the present invention (an operation explanatory diagram of a speech section including noise);

FIG. 6 is a simplified functional block diagram of the adaptive noise suppression speech coding apparatus according to the first embodiment of the present invention;

FIG. 7 is a simplified functional block diagram (another method) of the adaptive noise suppression speech coding apparatus according to the first embodiment of the present invention;

FIG. 8 is a simplified functional block diagram of an adaptive noise suppression speech coder according to a second embodiment of the present invention;

FIG. 9 is a graph of an autocorrelation function of an audio signal,

FIG. 10 is a diagram illustrating a calculation basis of a noise change rate K according to the first embodiment of the present invention;

FIG. 11 is a left half of a functional block diagram of an adaptive noise suppression speech coding apparatus according to a second embodiment of the present invention;

FIG. 12 is a right half of a functional block diagram of an adaptive noise suppression speech coder according to a second embodiment of the present invention;

FIG. 13 is a functional block diagram of an adaptive noise-suppressed speech coding apparatus according to a second embodiment of the present invention (operational explanatory diagram of a noise section);

FIG. 14 is a left half part of a functional block diagram (an operation explanatory diagram of a speech section including noise) of the adaptive noise suppression speech coding apparatus according to the second embodiment of the present invention;

FIG. 15 is a functional block diagram (an operation explanatory diagram of a speech section including noise) of the adaptive noise suppression speech coder according to the second embodiment of the present invention;

FIG. 16 is a left half of a functional block diagram of an adaptive noise suppression speech coder according to a third embodiment of the present invention;

FIG. 17 is a right half of a functional block diagram of an adaptive noise suppression speech coder according to a third embodiment of the present invention;

FIG. 18 is a left half part of a functional block diagram (an operation explanatory diagram of a speech section including noise) of the adaptive noise suppression speech coder according to the third embodiment of the present invention;

FIG. 19 is a right half part of a functional block diagram (an operation explanatory diagram of a speech section including noise) of the adaptive noise suppression speech coding apparatus according to the third embodiment of the present invention;

FIG. 20 is a graph of an autocorrelation function of an audio signal in consideration of a pitch period;

FIG. 21 is a functional block diagram of an adaptive noise-suppressed speech coding apparatus corresponding to the first embodiment of the fourth embodiment of the present invention (an operation explanatory diagram of a speech section including noise);

FIG. 22 is a right half part of a functional block diagram (operation explanatory diagram of a speech section including noise) of the adaptive noise suppression speech coding apparatus corresponding to the first embodiment of the fourth embodiment of the present invention;

FIG. 23 is a left half part of a functional block diagram (operation explanatory diagram of a speech section including noise) of an adaptive noise suppression speech coder corresponding to the second embodiment of the fourth embodiment of the present invention;

FIG. 24 is a right half part of a functional block diagram (an operation explanatory diagram of a speech section including noise) of the adaptive noise suppression speech coder corresponding to the second embodiment of the fourth embodiment of the present invention;

FIG. 25 is a block diagram showing a configuration of a conventional first noise suppression device.

FIG. 26 is an operation flowchart of a conventional first noise suppression device;

FIG. 27 is a block diagram showing a configuration of a second conventional noise suppression device.

FIG. 28 is an operation flowchart of a second conventional noise suppression device;

FIG. 29 is a simplified block diagram of a conventional ACELP speech coding apparatus;

FIG. 30 is a detailed functional block diagram of a conventional ACELP speech coding apparatus;

FIG. 31 is a diagram of a conventional audio signal windowing capture,

FIG. 32 is a codebook configuration table of a conventional ACELP speech encoding device.

[Explanation of symbols]

 101A voice / noise section determination circuit 101B voice / noise switch SW control circuit 102A voice / noise switch SW 102B voice / noise switch SW 102C voice / noise switch SW 102D voice / noise switch SW 102E voice / noise switch SW 102F voice / noise switch SW 103 Autocorrelation function storage circuit 105 Noise change rate estimation circuit 106 Autocorrelation function estimation circuit 107 Pitch autocorrelation function storage circuit 116 Fixed codebook 117A Fixed gain adjustment circuit 117B Impulse response circuit 117C Error calculation circuit 117D Error minimization circuit 117E Fixed Codebook optimal value discovery circuit 118 Fixed codebook gain quantization circuit 119 Excitation calculation circuit 120 Optimal filter storage circuit circuit 901 Extraction unit 902 Fourier analysis unit 903 Noise spectrum average calculation unit 904 Noise spectrum memory 905 Noise signal determination unit 906 Amplitude subtraction Section 907 Half-wave rectification section 908 Inverse Fourier transform section 909 Waveform reproduction section 1001 Step windowing 1002 step FFT analysis 1003 steps Step Voice section detection 1004 steps Noise storage 1005 steps Average spectrum calculation 1006 steps Noise subtraction unit 1007 steps Negative coefficient zeroization 1008 steps IFFT 1009 steps Waveform synthesis 1101 S / N measurement unit for each channel 1102 Gain control unit for each channel 1103 Each channel Variable gain amplifier 1104 step S / N measurement for each channel 1105 step Gain control for each channel 1106 step Variable gain amplification for each channel 1201 Fixed codebook 1202 Amplifier 1203 Adaptive codebook 1204 Amplifier 1205 Adder 1206 Linear prediction synthesis circuit 1207 Hearing correction circuit 1301 Preprocessing circuit 1302 Windowing circuit 1303 Autocorrelation function calculation circuit 1304 Prediction coefficient calculation circuit 1305 Coefficient conversion circuit 1306 Coefficient quantization circuit 1307 Subframe coefficient interpolation circuit 1308A Subframe coefficient interpolation circuit (before quantization) 1308B Impulse response Calculation circuit 1309 4 subframes without weight Voice calculation circuit 1310A Open loop pitch delay detection circuit 1310B Open loop pitch gain detection circuit 1311A Adaptive codebook 1311B Pitch gain adjustment circuit 1311C Impulse response circuit 1311D Error calculation circuit 1311E Error minimization circuit 1312 Optimal delay and gain detection circuit 1313 Pitch and gain Quantization circuit 1314 Applied codebook contribution calculation circuit 1316 Fixed codebook 1317A Fixed gain adjustment circuit 1317B Impulse response circuit 1317C Error calculation circuit 1317D Error minimization circuit 1317E Fixed codebook optimal value finding circuit 1318 Fixed codebook gain quantization circuit 1319 Excitation calculation circuit 1320 Optimal filter storage circuit

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Shogo Iizuka 3-1-4 Tsunashimahigashi, Kohoku-ku, Yokohama-shi, Kanagawa Prefecture Inside Matsushita Communication Industrial Co., Ltd. (72) Kazuki Hoshino 4-1-1 Mita, Meguro-ku, Tokyo No.4 Ebisu View Tower Room 3102 F term (reference) 5D045 CA01 9A001 GG01 HH15

Claims (14)

[Claims] 1. A means for converting an audio signal on which ambient noise is superimposed into a digital signal by an analog-to-digital converter, a means for extracting it as data of a fixed length section, an autocorrelation function R (0), R (1), means for calculating R (2),-,-, R (M), determination means for determining whether the extracted signal is a section of only noise or a voice signal section in which noise is superimposed, If the determination means determines that the section is only noise, the value of the autocorrelation function R (0), R (1), R (2),-,-, R (M) in that section is Rn
_m (0), Rn _m (1), Rn _m (2),-,-, noise autocorrelation function storage means to store as Rn _m (M), autocorrelation function R (0), R (1), R
(2) means for determining constants of the first predicting means based on the values of-,-, R (M), first impulse response calculating means for calculating an impulse response of a reciprocal of a transfer function of the first predicting means , A first fixed codebook, a first fixed code gain adjusting means,
A first convolution integrator for obtaining a convolution integral output of the output pulse train of the first fixed codebook via the first fixed code gain adjuster and the impulse response calculated by the first impulse response calculator; The first
A first error calculating means for calculating an error between the output of the convolution integrator and noise, calculating a mean square value of the error, and an output code of the first fixed book so as to minimize the error. A first control means for controlling the gain of the first fixed code gain adjustment means, a first noise state storage means for storing a constant of the first impulse response and an initial state after the minimization is achieved, and If it is determined that the means is a voice signal section in which noise is superimposed, Rn _m (0), Rn _m (1), Rn _m (2),-, stored in the noise autocorrelation function storage means -, The values of Rn _m (M) are read out and there are R (0), R (1), R (2),-,-, R (M) calculated by the autocorrelation function calculation means. Means for calculating a value K determined as a function as a time variation rate of noise;
R (0) -KRn as the autocorrelation function Rs (0), Rs (1), Rs (2),-,-, Rs (M) of only the speech signal without noise using the value of
_m (0), R (1) −KRn _m (1), R (2) −KRn _m (2), −, −, R (M) −KRn _m (M)
, Rs (0), Rs (1), Rs (2),-,-, Rs (M) to determine the constant of the second predicting means, and transmission of the second predicting means. Second to calculate the impulse response of the reciprocal of the function
Impulse response calculation means, a second fixed codebook,
Second fixed code gain adjusting means, second convolution integral for obtaining a convolution integrated output of the output pulse train of the second fixed codebook via the second fixed code gain adjusting means and the second impulse response Means, said second
Second error calculating means for calculating an error between the output of the convolution integrator and the signal from which noise has been removed from the speech signal on which noise is superimposed, and calculating the mean square value of the error and minimizing the error As described above, the second control means for controlling the output code of the second fixed book and the gain of the second fixed code gain adjustment means, after achieving the minimization, sets the constant and the initial state of the second impulse response. Adaptive noise suppression speech coding configured to include a second speech signal state storage means for storing and adapted to adapt to ambient noise during a call and reduce the influence of ambient noise to encode a speech signal. apparatus. 2. The method according to claim 1, wherein when the determining unit determines that the voice section has noise superimposed thereon, the output of the second error calculating unit is used as an input of the control unit that minimizes the first error. 2. The adaptive noise-suppressed speech coding apparatus according to claim 1, wherein said adaptive noise suppression speech coding apparatus is connected to stabilize the operation of said first error calculating means. 3. Rn stored in said first storage means.
_m (0), Rn _m (1), Rn _m (2),-,-, Rn _m (M) are read out, and those values and the voice in which the autocorrelation function calculating means is superimposed with noise R (0), R (1), R (2),-,-, R calculated as signal values
As a time variation rate value K of the noise determined as a function of (M), R (k) , Rn m (k) (k = 1,2, -, -, M) takes the respective positive or negative value, further R (k) and its sign, in view of the magnitude of _{Rn m (k), R (} k), Rn m (k) (k = 1,2, -, -, M) into a plurality of states 3. The method according to claim 1, wherein the classification is performed, and a minimum value of K is further selected from the minimum value of K calculated for each state using a different calculation formula of K for each state, and is set as the value of K. Adaptive noise suppression speech coding apparatus. 4. The noise time variation value K is represented by R (k), Rn
_m (k) (k = 1,2,-,-, M) takes a positive or negative value, respectively, and considering the magnitude relationship between the positive and negative and R (k), Rn _m (k), R
(k), Rn _m (k) (k = 1,2,-,-, M)
(k), Rn _m (k) (k = 1, 2,-,-, M) are classified into four states according to the positive / negative relationship, and further R (k), Rn _m (k) (k = 1, When (2,-,-, M) has the same sign, R (k) and Rn _m (k) are classified into two categories according to the magnitude relation, and classified into a total of six states. 4. The adaptive noise-suppressed speech coding apparatus according to claim 3, wherein a minimum value of K is further selected from the minimum value of K calculated for each state by the equation, and the selected value is used as the value of K. 5. The adaptive noise-suppressed speech coding apparatus according to claim 4, wherein R (k), Rn _m (k) (k = 1, 2,-,-, M) are classified into six states. 5. The adaptive noise-suppressed speech coding apparatus according to claim 4, wherein a calculation formula of K different in each state is K as shown in Table 1 below. [Table 1] 6. The method of claim 1 or the adaptive noise suppression speech encoding apparatus according to claim 2, wherein the first Rn _m stored in the storage means _{(0), Rn m (1} ), Rn m (2 ),-,-, Rn _m (M) are read out, and these values and R (0), R calculated by the autocorrelation function calculation means as a value of a voice signal on which noise is superimposed.
(1), R (2),-,-, R (M) As a time variation value K of noise determined as a certain function, for k = 1, 2,-,-, M, 3. The adaptive noise-suppressed speech coding apparatus according to claim 1, wherein K is calculated and the minimum value is selected as K. K = {b / a} [1− {1− (ca / b ² )} ^1/2 ] Here, a, b, and c have the following values. a = {R ² _nm (0) −R ² _nm (k)} b = {R (0) * R _nm (0) −R (k) * R _nm (k)} c = {R ² ( 0) −R ² (k)} 7. The apparatus further comprises means for calculating an autocorrelation function R (L) in a pitch period L of a voice signal, and means for storing the value, wherein Rn _m (0) read from the noise autocorrelation function storage means. ), Rn _m (1), Rn _m (2),-,-, the value of Rn _m (M), the value of the autocorrelation function R (L) of the pitch period read from the storage means, R (0), R (1), R calculated by the autocorrelation function calculation means as the autocorrelation function of the speech signal with noise superimposed
3. The adaptive noise-suppressed speech coding apparatus according to claim 1, wherein a value K determined as a certain function with a value of (2),-,-, R (M) is determined as a time variation rate of noise. 8. The adaptive noise-suppressed speech coding apparatus according to claim 7, wherein R (k), Rn
_m (k) (k = 1,2,-,-, M) takes a positive or negative value, respectively, and considering the magnitude relationship between the positive and negative and R (k), Rn _m (k), R
(k), Rn _m (k) (k = 1,2,-,-, M) are classified into multiple states,
A value determined as a function of a value obtained by further selecting the minimum K value from the minimum value of K calculated for each state using a different K calculation formula in each state, and a pitch period autocorrelation function R (L). 8. The adaptive noise-suppressed speech coding apparatus according to claim 7, wherein a smaller one of the values is a time variation rate K of noise. 9. The adaptive noise-suppressed speech coding apparatus according to claim 8, wherein R (k), Rn
_m (k) (k = 1,2,-,-, M) takes a positive or negative value, respectively, and considering the magnitude relationship between the positive and negative and R (k), Rn _m (k), R
_{(k), Rn m (k} ) (k = 1,2, -, -, M) when classifying the plurality of _{states, R (k), Rn m} (k) (k = 1,2, - , -, classified into four states by positive and negative relationship M), further _{R (k), Rn m (} k) (k = 1,2, -,
-, M) have the same sign, R (k), Rn _m (k) are classified into two states according to the magnitude relation, and classified into a total of six states. Either a value obtained by further selecting the minimum value of K from the minimum value of K calculated for each time or a value determined by the autocorrelation function R (L) of the pitch period is a smaller value as the time variation rate K of noise. An adaptive noise-suppressed speech coding apparatus according to claim 8. 10. The adaptive noise-suppressed speech coding apparatus according to claim 9, wherein R (k),
Rn _m (k) (k = 1,2,-,-, M) takes a positive or negative value, respectively.
Considering the magnitude relationship between the positive and negative and R (k), Rn _m (k), R
When (k), Rn _m (k) (k = 1,2,-,-, M) are classified into a plurality of states, they are classified into six states. It used those described in Table 1 shown in claim 5, and as the value determined by the pitch period autocorrelation function R (L), the minimum value calculated in terms of calculating the K _L by a formula shown below, adaptive noise suppression speech encoding apparatus according to claim 9, wherein either the small value was set as the time variation rate K of the noise values of K and K _L. K _L = [R (0) −R (L)] / Rn _m (0) 11. The value of the time variation rate K of the noise is a value of K determined by the expression described in claim 6, and a function of the autocorrelation function R (L) of the pitch period described in claim 7. adaptive noise reduced speech coding apparatus of any claim 6 or claim 7, wherein selecting smaller becomes the value as the value of K and the value of K _L defined as. 12. The adaptive noise suppression speech coding apparatus according to claim 11, wherein the expression defined in claim 10 is used as a value determined by an autocorrelation function R (L) of a pitch period. 12. The adaptive noise-suppressed speech coding apparatus according to claim 11. 13. When the determining means determines that the signal is a voice signal section in which noise is superimposed, the first fixed code gain adjusting means controlled by the data of the immediately preceding fixed section length. Means for calculating a value determined by using the final value of the gain as a time variation rate K of noise, and using the K, an autocorrelation function of only a noise-free speech signal;
(0), Rs (1), Rs (2),-,-, Rs (M) is R (0) −KRn _m (0), R (1) −KRn
3. The adaptive noise-suppressed speech code according to claim 1, wherein _m (1), R (2) −KRn _m (2), −, −, R (M) −KRn _m (M). Device. 14. The adaptive noise suppression speech coding apparatus according to claim 13, wherein the time variation rate K of the noise is the first fixed code gain adjustment controlled by data of a preceding fixed section length. The value of K is obtained by using the following equation, where g _a ′ is the final value of the gain of the means.
An adaptive noise-suppressed speech coding apparatus according to claim 1. K = {g _a ′} ² Y / {R _nm (0) + Σ _{k = 1} ^{k = M} _ak R _nm (k)} where Y is the actual number of pulses relative to the total number of samples in the fixed codebook. And a _k is a prediction coefficient of the linear prediction circuit.