A system for coding wide-band audio signals

Claim 3

A method as claimed in claim 1, characterized in that during coding the available bit rate is equally shared among the different sub-bands.

Claim 4

A method as claimed in claim 1, characterized in that during coding the available bit rate is unequally shared among the various sub-bands, and each sub-band is assigned the bit rate assuring in the average a uniform distortion in all sub-bands.

Claim 5

A method as claimed in claim 1 or 3, characterized in that each sub-band is coded by the technique which is better suited to represent the characteristics of its excitation and which is better suited to the bit rate chosen for that sub-band.

Claim 6

A method as claimed in claim 1, characterized in that the available bit rate is unequally and dynamically shared among the various sub-bands, each sub-band being allotted, at each frame and in a way varying from one frame to the subsequent one, the bit rate ensuring a uniform distortion during each frame.

Claim 8

A method as claimed in claim 6, characterized in that the coding technique is a linear prediction coding technique with multi-pulse excitation, in which a split-­codebook vector quantization of a representation of linear prediction coefficients resulting from the short-term analysis is effected, and in which long-term analysis and excitation pulse generation are performed in successive steps, by determining first the optimum values of the lag which separates a current sample from a preceding sample used for processing said current sample (long-term analysis lag) and of the gain by which said preceding sample is weighted to process the current sample (long-term analysis gain), and then the amplitudes and the positions of the excitation pulses, the long-term analysis lag and gain being also determined in successive steps.

Claim 9

A method as claimed in claim 6, characterized in that the coding technique is a codebook excited linear prediction coding technique, in which a split-codebook vector quantization of a representation of linear prediction coefficients resulting from the short-term analysis is effected, and in which the long-term analysis and excitation generation are performed in two successive steps, in the first of which an optimum value is determined for the lag which separates a current sample from a preceding sample used for processing the current sample (long-term analysis lag), while in the second step an optimum excitation model is determined, which consists of the optimum gain by which that preceding sample is weighted (long-­term analysis gain), of a sequence of vectors and of a respective scale factor, said long-term analysis gain and said vectors and scale factors being jointly determined.

Claim 10

A method as claimed in claim 8, characterized in that before said two steps the contribution [s0(p)] given to the analyses and syntheses relevant to a current frame by the analyses and syntheses previously performed is determined once for all by weighting a null signal by a weighting function [1/A(z,γ)], depending on the linear prediction coefficients reconstructed after split-codebook vector quantization, and said contribution is subtracted from a first weighted signal [s w (p)], obtained by spectrally shaping the signal to be coded, thus generating a first error signal [s1(p)].

Claim 11

A method as claimed in claim 9, characterized in that for the determination of the optimum value (M) of long-term analysis lag the energy of a second error signal [s3(p)] is minimized, which signal is the difference between said first error signal [s1(p)] and a second weighted signal [s2(p)], obtained by submitting a null signal to a first long-term synthesis filtering and by weighting with said weighting function the signal resulting from the synthesis filtering, said first long-term synthesis filtering being performed by using successively all possible values (m) of the long-term analysis lag and, for each of them, a corresponding optimum value [b(m)] of the long-term analysis gain; and in that, for determining the excitation, the energy of a third error signal [s5(p)] is minimized, which error signal is the difference between said first error signal and a third weighted signal [s4(p)], obtained by weighting with said weighting function [1/A(z,γ)] a signal which is the sum of an excitation vector [e(i)], weighted by the relevant scale factor (g), and of a further signal obtained by submitting a null signal to a second long-term synthesis filtering in which the lag used is said optimum value (M) of the long term analysis lag and the gain is an optimum long-term analysis gain for said vector.

Claim 12

A method as claimed in claim 9, characterized in that, for determining said optimum value (M) of long-term analysis lag, the energy of a fourth error signal [s7(p)] is minimized, which signal is the difference between said first error signal [s1(p)] and a fourth weighted signal [s6(p)], obtained by submitting a null signal to a third long-term synthesis filtering in which all possible values (m) of the long-term analysis lag are successively used in combination with a unitary gain, by weighting with said weighting function the signal resulting from the synthesis filtering and by multiplying the signal (t1) resulting from the weighting by a value [b(m)] of the long-term analysis gain which is the optimum value for that lag; and in that for said determination of the excitation the energy of a fifth error signal [s9(p)] is minimized, which signal is obtained by subtracting a fourth weighted signal [s8(p)], obtained by weighting by said weighting function [1/A(z,γ)] all possible excitation vectors e(i) and by multiplying each weighted vector by a value [g(i)] of ihe scale factor which is the optimum value for said vector, from a signal which consists of the fourth error signal [s7(p)] when the optimum value (M) of the long-term analysis lag and a value of the long-term analysis gain which is the optimum value for that vector are used in said third long-term synthesis filtering.

Claim 13

A method as claimed in claim 11, characterized in that the values of the scale factors are normalized with respect to the maximum or the mean value of said factors within a frame, and at each frame said maximum or mean value is optimized so as to minimize the energy of said fifth error signal [s9(p)].

Claim 17

A device as claimed in claim 14, characterized in that the coder (CD1, CD2) relevant to at least one of the sub-bands is a coder based on multi-pulse excitation linear prediction coding, in which the means (LTA) for the long-term analysis are apt to determine the lag (M) which separates a current sample from a preceding sample used to process said current sample and the gain (B) by which said preceding sample is weighted in two successive steps, which both precede a step in which the amplitudes and positions of the excitation pulses are determined by an excitation generator (EG), and the means (STA) for the short-term analysis are connected to a circuit (VQ) for split-codebook vector quantization of adjacent spectrum line pairs, or of differences between adjacent spectrum lines, forming a representation of linear prediction coefficients obtained as a result of the short-­term analysis.

Claim 18

A device as claimed in claim 14, characterized in that the coder (CD1, CD2) relevant to at least one of the sub-bands is a coder based on the codebook excited linear prediction technique, in which the means (LTA1) for the long-term analysis are apt to determine the lag (M) which separates a current sample from a preceding sample used to determine said current sample in a first operation step which precedes a step in which an excitation generator (EG1) determines an optimum excitation model, consisting of the optimum value of the gain (B) by which said preceding sample is weighted, of a sequence of vectors chosen out of the excitation codebook and of the respective scale factor, said gain, excitation vectors and scale factor being jointly determined, and in that the means (STA) for the short-term analysis are connected to a circuit (VQ) for split-codebook vector quantization of adjacent spectrum line pairs, or of differences between adjacent spectrum lines, obtained as a representation of linear prediction coefficients determined by short-term analysis.

Claim 19

A device as claimed in claim 16, characterized in that the means (LTA1) for long-­term analysis comprise a first weighting filter (F5) which determines the contribution [s0(p)] given to the analyses and syntheses relevant to a current frame by the analyses and syntheses previously performed, by weighting a null signal with a weighting function [1/A(z,γ)], depending on the linear prediction coefficients reconstructed after vector quantization, and a first adder (SM3) which subtracts said contribution from a first weighted signal [sw(p)], obtained by spectrally shaping the input signals of the coder (CD1, CD2), and supplies a first error signal [s1(p)], used at least for determining the long-term analysis lag (M).

Claim 20

A device as claimed in claim 17, characterized in that the means (LTA1) for long-­term analysis also comprises: - a first long-term synthesis filter (LTP4), which is fed with a null signal, presents a transfer function depending on long-term analysis lag and gain and successively uses all possible lag values and, for each of them, a value [b(m)] of the gain which is the optimum value for said lag; - a second weighting filter (F6) connected to the output of said long-term synthesis filter (LTP4) and having as transfer function said weighting function; - a second adder (SM4), which receives at a first input the first error signal [s1(p)] and at a second input the output signal of said second weighting filter (F6) and supplies a second error signal [s3(p)] which is the difference between the signals present at its first and second inputs; - a first processing unit (CMB1), which successively supplies said synthesis filter (LTP4) with all possible values (m) of the long-term analysis lag, receives said second error signal and determines the value of the lag which minimizes the energy thereof and forms the optimum lag (M), the processing unit resetting at each new value of the lag the memory of said second weighting filter (F6); and in that the means (EG1) for determining the optimum excitation model comprises: - a second long-term synthesis filter (LTP5), which is fed with a null signal, presents a transfer function analogous to that of the first synthesis filter (LTP4) and uses the optimum lag value (M) and a value [b(i)] of the gain which is the optimum value for a given excitation vector [e(i)]; - a third adder (SM5) which adds the output signal of said third synthesis filter (LTP5) and the output signal of a first multiplier (M1) which multiplies an excitation vector by a respective scale factor; - a third weighting filter (F7), connected to the output of said third adder (SM5) and having as transfer function said weighting function [1/A(z,γ)]; - a fourth adder (SM6), which receives at a first input the first error signal [s1(p)] and at a second input the output signal [s4(p)] of said third weighting filter (F7) and supplies a third error signal [s5(p)], which is the difference between the signals present at its first and second inputs; - a second processing unit (CE1) which supplies said second synthesis filter (LTP5) with the gain value [b(i)] and said first multiplier (M1) with the vectors and scale factors, receives said third error signal [s5(p)] and determines the combination of gain (B), excitation vectors and respective scale factors forming the optimum excitation model by minimizing the energy of said third error signal, the second processing unit (CE1) resetting at each new combination of gain, vector and scale factor the memories of the third weighting filter (F7).

Claim 21

A device as claimed in claim 17, characterized in that the means (LTA1) for long-­term analysis further comprises: - a third long-term synthesis filter (LTP6), which has a transfer function depending on the long-term analysis lag, is fed with a null signal and, while determining the optimum long-term analysis lag, successively uses all possible lag values, whereas, while determining the optimum excitation model, uses the optimum lag value; - a fourth weighting filter (F8), connected to the output of said third long-term synthesis filter (LTP2) and having as transfer function said weighting function; - a second multiplier (M2) which is connected to the output of the fourth weighting filter (F8) and which, during determination of the optimum long-term analysis lag, for each value of the lag used in the third synthesis filter (LTP6), multiplies the output signal (t1) of said fourth weighting filter (F8) by a value [b(m)] of the long-term analysis gain which is the optimum value for that lag, whereas during determination of the optimum excitation model, it multiplies the output signal of said fourth filter (F8) by a value [b(i)] of the long-term analysis gain which is the optimum value for a given excitation vector; - a fifth adder (SM7), which receives at a first input the first error signal [s1(p)] and at a second input the output signal of said second multiplier (M2) and supplies a fourth error signal [s7(p)], consisting of the difference between the signals present at the first and second inputs; - a third processing unit (CMB2), which supplies successively said third synthesis filter (LTP6) with all possible values (m) of the long-term analysis lag, resetting at each new value the memory of the fourth weighting filter (F8), and, while determining the optimum lag, supplies the second multiplier (M2) with the values of the gain [b(m)], receives said fourth error signal and determines the lag value which minimizes the energy of the fourth error signal and forms the optimum lag (M); and in that the means (EG1) for determining optimum long-term analysis gain (B), the vectors e(i) and their respective scale factors (g), forming the optimum excitation model, comprise: - a fifth weighting filter (F9), fed with the excitation vectors and having as transfer function said weighting function; - a third multiplier (M3), which is connected to the output of the fifth weighting filter (F9) and, for each vector it supplies to such fifth weighting filter, multiplies the output signal of the filter (F9) by an optimum scale factor; - a fifth adder (SM8), which receives at a first input the fourth error signal [s7(p)] and at a second input the output signal [s8(p)] of said fifth weighting filter (F9) and generates a fifth error signal [s9(p)], consisting of the difference between the signals present at the first and second inputs; - a fourth processing unit (CE2), which supplies said fifth weighting filter (F9) with the excitation vectors, said third multiplier (M3) with the optimum scale factor for each vector supplied to said fifth weighting filter (F9) and said second multiplier (M2) with the optimum long-term analysis gain for that vector, receives said fifth error signal [s9(p)] and determines the combination long-­term analysis gain (B), excitation vectors and respective scale factors which forms the optimum excitation model by minimizing the energy of said fifth error signal [s9(p)], the fourth processing unit resetting the memories of the fifth weighting filter (F9) at each new combination vector, scale factor, long-term analysis gain.