JP2621376B2 - Multi-pulse encoder - Google Patents

Abstract

PURPOSE:To obtain the multi-pulse encoding device which is tolerant to a transmission line error by providing a linear predicting means which inputs the output signal of a 2nd fixed filtering means. CONSTITUTION:A linear prediction LPC coefficient extracting means 3 equipped with the 2nd fixed filtering means 31 and a linear prediction LPC means 33 extracts the linear prediction coefficient of a pre-emphasized voice signal. Further, spectrum structure conversion including partial pre-emphasis is carried out for the voice signal by a 1st fixed filtering means 81 and a time variation filtering means 83 included in a spectrum structure converting means 8 and then a multi-pulse analysis is taken to obtain multi-pulses. The pre-emphasis is therefore taken without spoiling the effect of equivalent S/N improvement by noise shaping. Consequently, the multi-pule encoding device is made tolerant to transmission line errors.

Description

Description: TECHNICAL FIELD The present invention relates to a multi-pulse encoder.

[Conventional technology]

Multi-pulse encoding is one of the methods for encoding an audio signal efficiently.

In multi-pulse coding, on the analysis side, spectral envelope information is represented by LPC coefficients such as α parameters and K parameters obtained by linear predictive coding (LPC) analysis of an input speech signal, and an impulse sequence (multi-pulse ) Represents sound source information. The spectrum envelope information represents spectrum distribution information of a vocal tract system that generates a voice. The sound source information represents a fine structure of a spectrum envelope, and is information on a residual signal obtained by removing spectrum distribution information from an input audio signal. Each pulse of the above-described multi-pulse is determined so as to have an occurrence time position and an amplitude corresponding to the characteristic of the waveform of the residual signal. On the synthesis side, an output voice signal can be synthesized by exciting an LPC synthesis filter, which is an all-pole digital filter using the α parameter as a coefficient, with multiple pulses. Since the excitation information includes the above-described waveform information, multi-pulse encoding is efficient.

In multi-pulse coding, noise shaping is performed. This noise shaping is performed as described below.

On the analysis side, the input audio signal is subjected to spectral structure conversion so as to suppress a change in its spectral envelope, and the spectral structure converted audio signal is analyzed to obtain a multipulse. On the analysis side, an audio signal (corresponding to the audio signal after the spectral structure conversion) is synthesized using the multi-pulse. The spectral envelope of the quantization noise included in the synthesized speech signal is flat. This audio signal is subjected to spectrum structure conversion by inverse conversion of the spectrum structure conversion on the synthesis side to obtain an output audio signal corresponding to the input audio signal. The spectral envelope of the quantization noise included in the output audio signal has the same shape as the difference between the spectral envelope of the input audio signal and the spectral envelope of the audio signal after the spectral structure conversion. In other words, the noise envelope is shaped so that the spectral envelope of the quantization noise in the output audio signal is somewhat similar to the spectral envelope of the output audio signal. The noise in speech coding is dominated by quantization noise.Hearing has the property that noise components at frequencies close to the signal components are masked by the signal components when the S / N exceeds a certain value and are not felt. S / N can be equivalently improved by performing noise shaping. This S / N
The degree of improvement does not increase monotonically with the degree of noise shaping, in other words, the degree of suppression of changes in the spectral envelope on the analysis side.There is an optimum value for the degree of noise shaping, and this optimum value is Selected. As will be described later in detail, an output audio signal corresponding to an input audio signal can be directly synthesized from a multi-pulse without obtaining an intermediate audio signal corresponding to an audio signal subjected to spectral structure conversion.

[Problems to be solved by the invention]

The conventional multi-pulse encoding apparatus that performs the above-described multi-pulse encoding has a disadvantage that it is vulnerable to transmission path errors. The cause is that the coefficient sensitivity of the LPC coefficient is high.

Pre-emphasis is known as a general and rational method for reducing the coefficient sensitivity. Since the speech spectrum has a primary slope of -6 dB / octave on average, if pre-emphasis is performed to compensate for this primary slope before encoding, less information needs to be transmitted from the analysis side to the synthesis side. Alternatively, if the bit transmission rates are the same, the information to be transmitted can be given more redundancy, so that the coefficient sensitivity can be reduced. However, the multi-pulse coding apparatus performs noise shaping as described above, and pre-emphasizes the input speech signal to eliminate the (average) first-order gradient and then inputs the signal to the conventional multi-pulse coding apparatus. Then, the (average) 1 of the synthesized output audio signal and the quantization noise included in the audio signal is obtained.
The next slopes match. In other words, too much noise shaping, equivalent S / N due to noise shaping
The improvement effect is impaired. Therefore, pre-emphasis cannot be directly applied to a conventional multi-pulse encoder.

An object of the present invention is to provide a multi-pulse encoding device that can perform pre-emphasis without impairing the effect of noise shaping and is therefore resistant to transmission path errors.

[Means for Solving the Problems] A multi-pulse encoding apparatus according to the present invention is characterized in that an input speech signal is analyzed by a linear prediction coefficient extraction unit for each analysis frame, and a linear prediction coefficient obtained from the spectral envelope information of the input speech signal is obtained. And the input voice signal is analyzed by the multi-pulse analysis means using the linear prediction coefficient and a predetermined attenuation coefficient for each analysis frame, and the occurrence time position and amplitude corresponding to the feature of the sound source information of the input voice signal In a multi-pulse encoding apparatus that obtains an impulse sequence composed of a plurality of impulses having the following and expresses the excitation information by the impulse sequence, a first fixed mode that reduces a first-order gradient of an average spectrum of the input speech signal Using the filtering means and the linear prediction coefficient and the attenuation coefficient before the input audio signal. A time-varying filtering means for relaxing a spectrum envelope for each analysis frame, and a spectrum structure converting means for converting a spectrum structure of the input voice signal and outputting a conversion result to the multi-pulse analyzing means; The extraction means includes second fixed filtering means for compensating for the first-order gradient of the average spectrum of the input voice signal, and linear prediction means for inputting an output signal of the second fixed filtering means.

The first fixed filtering means inputs a unit delay means, first and second multiplication means having the output signal of the unit delay means as one input, and a multiplication result of the first multiplication means. A subtraction means for subtracting the subtraction result from the signal and outputting the subtraction result to the unit delay means; and an addition means for adding the subtraction result and the multiplication result of the second multiplication means to produce an output signal. You may.

Further, the time-varying filtering means comprises a cascade-connected n
(N is an integer of 2 or more) unit delay means, n first multiplying means having respective output signals of the unit delay means as one input, and multiplication of these first multiplying means First adding means for adding a result, subtracting means for subtracting the result of the addition by the first adding means from an input signal and outputting a subtraction result to the first unit delay means, And n second multiplying means each of which has one output as the input signal, and second adding means which adds the multiplication result of the second multiplication means and the subtraction result to produce an output signal. It may be configured as:

Further, the second fixed filtering means includes a unit delay means, a multiplication means having an output signal of the unit delay means as one input, and a multiplication result of the multiplication means and an input signal of the unit delay means. And an adding means for converting the output signal into an output signal.

〔Example〕

 Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an analysis side of one embodiment of the present invention, and FIG. 2 is a block diagram showing a synthesis side of the embodiment.

The analysis side shown in FIG. 1 includes a low-pass filter 1 having a cut-off frequency of 3.4 kHz for inputting an audio signal Si, and an output signal of the low-pass filter 1 sampled at a sampling frequency of 8 kHz, quantized to a required number of bits, and quantized. A / D output as
A converter 2, a fixed filter 31 for realizing a second fixed filtering means for compensating the first-order gradient of the average spectrum of the input audio signal, and a waveform extracting circuit 32 cascade-connected to the fixed filter 31; A / A with LPC analyzer 33
LPC coefficient extraction unit 3 for extracting and outputting K parameters k _{1 to} k _p from the output signal of D converter 2, and K quantization for quantizing K parameters k _{1 to} k _p output by LPC coefficient extraction unit 3 Vessel 4,
A K decoder 5 for decoding the output signal of the K quantizer 4 and outputting a K parameter; and a K / for converting the K parameter output from the K decoder 5 into α parameters α _{1 to} α _p and outputting the same. α
A converter 6, alpha parameter alpha ₁ to? _P and the damping coefficient application circuit 7 for generating and outputting a parameter γα ₁ ~γ _^p α ^p from attenuation coefficient gamma, 1 order slope of the average spectrum of the input speech signal Filter 81 that realizes a first fixed filtering means for reducing noise, and a time-varying filter that is cascade-connected to the fixed filter 81 and realizes a time-varying filtering means for reducing the spectral envelope of each analysis frame of the input audio signal. alpha parameter to the coefficients of the time-varying filter 82 has a filter 82 alpha ₁ to? _p and parameters γα ₁ ~
A spectrum structure conversion unit 8 for inputting an output signal of the A / D converter 2 using γ ^p α _p , a waveform cutout circuit 9 for inputting an output signal of the spectrum structure conversion unit 8, and parameters γ α ₁ to
A multi-pulse analysis circuit 10 that performs multi-pulse analysis on an output signal of the waveform extraction circuit 9 and outputs a multi-pulse mp using γ ^p α _p , a maximum amplitude pulse of each pulse of the multi-pulse mp, A multi-pulse quantizer 11 for outputting data representing the normalized amplitude of the other pulses and the intervals between the pulses, and an output signal of the K quantizer 4 and output data of the multi-pulse quantizer 11 And a multiplexing circuit 12.

2 includes a separating circuit 13 for separating the signal output from the multiplexing circuit 12 on the analyzing side and transmitted to a state before the multiplexing by the multiplexing circuit 12, and one of the separating circuits 13 K decoder 14 for decoding the output and outputting the K parameter
And a K / α converter 15 for converting the K parameter into α parameters α _{1 to} α _p and outputting the same, and a multi-pulse decoder 16 for generating a multi-pulse mp from the other separated output of the separating circuit 13
And a multi-pulse using α parameters α _{1 to} α _p as coefficients
It comprises an LPC synthesis filter 17 excited by mp, a fixed filter 18, a D / A converter 19, and a reduction filter 20 sequentially connected to the output terminal of the LPC synthesis filter.

 FIG. 3 is a block diagram of the fixed filter 31.

The fixed filter 31 includes a unit delay circuit 311 for delaying an input signal, and a multiplier 312 for multiplying the output signal of the unit delay circuit 311 by −inverting the sign of an average first-order autocorrelation coefficient (= 0.9) of voice. And an adder 313 that adds the output signal of the multiplier 312 to the input signal and generates an output signal. The adder 313 is configured to output the audio signal represented by the quantized sample sequence output from the A / D converter 2. It outputs a quantized sample sequence of a pre-emphasized audio signal in which the average primary gradient of the spectrum has been compensated, in other words, a pre-emphasized audio signal. The waveform extraction circuit 32 performs window extraction for multiplying the quantized sample for 30 ms (240 points) output from the fixed filter 31 by the Hamming window function every 20 ms which is the analysis frame period (at an analysis frame frequency of 50 Hz). is there). The LPC analyzer 33 performs LPC analysis on the signal input from the waveform extraction circuit 32 for each analysis frame, and extracts K parameters k _{1 to} k _p , which are P-order partial autocorrelation coefficients. Therefore, the obtained K parameters k _{1 to} k _p are the spectral envelope information of the pre-emphasized audio signal. The order of the fixed filter 31 and the waveform extracting circuit 32 may be reversed from that in FIG. Further, the fixed filter 31 is configured as a transversal filter having one filter coefficient, and the waveform extraction circuit 32 can be realized by a transversal filter using a 240-point hamming coefficient as a filter coefficient. The fixed filter 31 and the waveform extracting circuit 32 can be replaced by a transversal filter having filter coefficients at points.

 FIG. 4 is a block diagram of the fixed filter 81.

The fixed filter 81 includes a unit delay circuit 811, a multiplier 812 that multiplies the output signal of the unit delay circuit 811 by −γ, and subtracts the output signal of the multiplier 812 from the input signal and outputs a subtraction result to the unit delay circuit 811. Subtractor 813 and unit delay circuit 811
Of the multiplier 814, and an adder 815 that adds the output signals of the subtractor 813 and the multiplier 814 to produce an output signal. γ is an attenuation coefficient, and has the same value as an attenuation coefficient used by a time-varying filter 82 described later. Assuming that the meaningful maximum response order of the loop including the unit delay circuit 811, the multiplier 812, and the adder 813 is Nth order, N unit delay circuits 816, N multipliers 817, and an adder 818.
The fixed filter 81 can be replaced by an N + 1 tap transversal filter shown in FIG. The tap coefficient of the first tap of this transversal filter is 1, so that there is no multiplier at the first tap, and the tap coefficients of the second and subsequent taps are (γ-1), ² (γ-1) γ , ³ (γ
-1) γ ² ,... ^N (γ-1) γ ^N−1 is set.

 FIG. 6 is a block diagram of the time-varying filter 82.

The time-varying filter 82 is composed of P unit delay circuits connected in cascade.
821, P multipliers 822 having the output signal of each unit delay circuit 821 as one input, an adder 823 for adding the output signal of each multiplier 822, and the addition result of the adder 823 from the input signal. A subtractor 824 for subtracting and outputting the subtraction result to the first unit delay circuit 821, a P number of multipliers 825 having the output signal of each unit delay circuit 821 as one input, a subtraction result of the subtractor 824 and each And an adder 826 for adding the multiplication results of the multiplier 825 to produce an output signal. As the other input of each multiplier 822, the parameter γ from the attenuation coefficient applying circuit 7 is input.
α _{1 to} γ ^p α _p are sequentially assigned to each multiplier 822 from the first multiplier 822.
To enter. Also, as the other input of each multiplier 825, K
The α parameters α _{1 to} α _p from the / α converter 6 are sequentially input to each multiplier 825 from the first multiplier 825. Note that the K / α converter 6 uses the K parameter output from the K
The reason why the K parameters k _{1 to} k _p from the C coefficient extraction unit 3 are not used is that the K side including the quantization distortion by the k quantizer 4 is used on the analysis side, and the influence of this quantization distortion on the synthesis side is obtained. In order to cancel out.

The quantized sample sequence output from the A / D converter 2 is filtered by a fixed filter 81 and a time-varying filter 82 and is subjected to spectral structure conversion. The attenuation coefficient γ is a constant that determines the degree of suppression of the change in the spectral envelope due to the spectral structure conversion. If γ = 1, the output signal of the time-varying filter 82 becomes the same as the input signal, and if γ = 0, The output signal becomes a quantized sample sequence of the residual signal. An optimal value of γ is selected in the range of 0 <γ <1. In this embodiment, γ = 0.8.

The fixed filter 81 filters the quantized sample sequence output from the A / D converter 2 with a transfer function different from that of the fixed filter 31. As can be seen from a comparison of the circuit configurations of FIGS. 4 and 6, the fixed filter 81 compensates for the primary slope to the same extent as the time-varying filter 82 compensates for the primary slope of the voice spectrum to some extent on average. Is not completely flattened, so that noise shaping is not excessively applied by the spectral structure conversion by the fixed filter 81 and the time-varying filter 82, and the fixed filter 81 itself has a certain (partial) pre-emphasis effect. . Note that the order of the fixed filter 81 and the time-varying filter 82 may be reversed from that in FIG.

The quantized sample sequence that has undergone spectral structure conversion including partial pre-emphasis by the spectral structure conversion unit 8 is cut out for each analysis frame by the waveform cutout circuit 9 and is input to the multi-pulse analysis circuit 10. The window extraction by the waveform extraction circuit 9 is the same as the window extraction by the waveform extraction circuit 32 except that a rectangular waveform having a width of 25 ms is used as a window function. The multi-pulse analysis circuit 10 uses a sequence of impulses of an appropriate position, amplitude, and polarity as a multi-pulse mp so as to represent the waveform of the residual signal of the audio signal represented by the signal from the waveform extraction circuit 9 as accurately as possible. Occur. This generation is performed by a known method using an impulse response of a filter expressing a spectral envelope of a waveform represented by a signal supplied from the waveform extracting circuit 9. This spectral envelope can be expressed by a combination of an all-pole filter defined by parameters γα _{1 to} γ ^p α _p and a first-order all-pole filter which is a (incomplete) spectral envelope filter in a macro sense. . Both parameters instead of the filter combinations of impulse responses γα ₁ ~γ ^p be utilized impulse response of the all-pole filter only defined by alpha _p practically sufficient. Multi-pulse analysis circuit 10 generates a multi pulse mp using parameters γα ₁ ~γ _^p α ^p supplied from the damping coefficient application circuit 7. The information of the multipulse mp obtained by the multipulse quantizer 11 is multiplexed by the multiplexing circuit 12 with the K parameter quantized by the K quantizer 4 and transmitted to the combining side.

On the combining side, first, a quantized sample sequence corresponding to the output signal of the fixed filter 31 is obtained using the multipulse mp and α _{1 to} α _p obtained from the transmitted signal. An operation of filtering the multi-pulse mp output from the multi-pulse decoder 16 with an all-pole filter defined by parameters γα _{1 to} γ ^p α _p to obtain a quantized sample sequence corresponding to the output signal of the spectrum structure conversion unit 8 Offsets a part of the operation of obtaining a quantized sample sequence corresponding to the output signal of the fixed filter 31 from the quantized sample sequence, and filters the multipulse mp (the remaining operation) by the LPC synthesis filter 17. Only with this, a quantized sample sequence corresponding to the output signal of the fixed filter 31 is obtained. The obtained quantized sample sequence is input to the fixed filter 18.

 FIG. 7 is a block diagram of the fixed filter 18.

The fixed filter 18 includes a unit delay circuit 181 that delays the output signal, a multiplier 182 that multiplies the output signal of the unit delay circuit 181 by −, and a subtraction result of the output signal of the multiplier 182 from the input signal and an output signal. And a de-emphasis characteristic corresponding to the pre-emphasis by the fixed filter 31 to de-emphasize the input quantized sample sequence, and output the D / A converter 2 on the analysis side. A quantized sample sequence corresponding to the quantized sample sequence thus output is output. The quantized samples sequence is an analog of the D / A converter 19, unnecessary high-frequency components in the low pass filter 20 is removed the audio signal S ₀ with. The quantization noise included in the audio signal S ₀ has a spectrum envelope having the same shape as the change in the spectrum envelope generated by the spectrum structure conversion by the spectrum structure conversion unit 8, in other words, required noise shaping is performed.

〔The invention's effect〕

As described above, the present invention extracts a linear prediction coefficient of a speech signal pre-emphasized by a linear prediction coefficient extraction means provided with a second fixed filtering means and a linear prediction means, (1) Speech signal is subjected to spectral structure conversion including partial pre-emphasis by a fixed filtering means and a time-varying filtering means, and then multi-pulse analysis is performed to obtain multi-pulses, thereby enabling equivalent S / N improvement by noise shaping. Since pre-emphasis can be performed without impairing the effect of the above, there is an effect that the multi-pulse encoder can be made robust against transmission path errors.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an analysis side of one embodiment of the present invention,
FIG. 2 is a block diagram showing the synthesis side of the embodiment, FIG. 3 is a block diagram of the fixed filter 31 on the analysis side shown in FIG. 1, FIG.
FIG. 6 is a block diagram of a transversal filter that can replace the fixed filter 81, FIG. 6 is a block diagram of a time-varying filter 82 on the analysis side shown in FIG. 1, and FIG. 7 is a fixed filter 18 on the synthesis side shown in FIG. It is a block diagram of. 1 low-pass filter, 2 A / D converter, 3 LPC coefficient extraction unit, 4 k quantizer, 5 k decoder, 6
k / α converter, 7: Attenuation coefficient application circuit, 8: Spectral structure converter, 9: Waveform extraction circuit, 10: Multipulse analysis circuit, 11: Multipulse quantizer, 12 ... Multiplexing circuit, 13 ... separating circuit, 14 ... k decoder, 15 ... k /
α converter, 16 multi-pulse decoder, 17 LPC synthesis filter, 18 fixed filter, 19 D / A converter, 2
0: Low-pass filter, 31: Fixed filter, 32: Waveform extraction circuit, 33: LPC analyzer, 81: Fixed filter, 82
...... Time-varying filter.

Claims (4)

(57) [Claims] An input speech signal which is analyzed by a linear prediction coefficient extracting means for each analysis frame to express spectral envelope information of the input speech signal, and the input speech signal is expressed for each analysis frame; Using a linear prediction coefficient and a predetermined attenuation coefficient, the multi-pulse analysis means performs analysis to obtain an impulse sequence including a plurality of impulses having an occurrence time position and an amplitude corresponding to the characteristics of the sound source information of the input audio signal. In a multi-pulse encoding device that expresses the sound source information by an impulse sequence, a first fixed filtering unit that reduces a first-order gradient of an average spectrum of the input speech signal and the linear prediction coefficient and the attenuation coefficient are used. A time-varying filter for alleviating the spectral envelope of the input audio signal for each analysis frame. Filtering means for converting the spectral structure of the input voice signal and outputting a conversion result to the multi-pulse analyzing means, wherein the linear prediction coefficient extracting means includes an average spectrum of the input voice signal. A second fixed filtering means for compensating for the first-order slope and a linear prediction means for inputting an output signal of the second fixed filtering means. 2. The first fixed filtering means includes a unit delay means, first and second multiplication means each having one input of an output signal of the unit delay means, and a first multiplication means. Subtraction means for subtracting the multiplication result from the input signal and outputting the subtraction result to the unit delay means; and addition means for adding the subtraction result and the multiplication result of the second multiplication means to obtain an output signal. The multi-pulse encoding device according to claim 1, wherein the multi-pulse encoding device is configured as follows. 3. The time-varying filtering means includes cascaded n (n is an integer of 2 or more) unit delay means, and n output signals of each of the unit delay means having one input thereof. A first multiplying means, a first adding means for adding the multiplication results of the first multiplying means, and a result of the addition by the first adding means being subtracted from the input signal, and the result of the subtraction being the first one. Subtracting means for outputting to the unit delay means, n second multiplying means having each output signal of the unit delay means as one input, a multiplication result of the second multiplication means and the subtraction result 2. The multi-pulse encoding apparatus according to claim 1, further comprising: a second adding means for adding the second signal to an output signal. 4. The second fixed filtering means includes a unit delay means, a multiplication means having an output signal of the unit delay means as one input, a multiplication result of the multiplication means, and an input signal of the unit delay means. 2. The multi-pulse encoding apparatus according to claim 1, further comprising an adding unit that adds the signals to output signals.