GB2327021A

GB2327021A - Speech coding

Info

Publication number: GB2327021A
Application number: GB9713814A
Authority: GB
Inventors: Daniel Brighenti
Original assignee: Telefonaktiebolaget LM Ericsson AB
Current assignee: Telefonaktiebolaget LM Ericsson AB
Priority date: 1997-06-30
Filing date: 1997-06-30
Publication date: 1999-01-06
Also published as: WO1999001866A1; AU8339598A; GB9713814D0

Abstract

A CELP speech coder estimates the LPC coefficient values on a continuously adaptive basis, for example using the LMS algorithm. An LPC analysis circuit 4 which may be used in a conventional speech coder (fig.1 not shown) has samples of input speech signal applied to a processor 6 and a predictor 8. The input speech samples are subtractively combined at 10 to obtain a residual value which is also applied to the processor 6 and the processor outputs the LPC coefficients.

Description

SPEECH CODING TECHNICAL FIELD OF THE INVENTION This invention relates to a speech coder, for example for use in a radio telecommunications system, such as a mobile telephone system. In particular, the invention relates to the coder itself and to a method of coding a speech signal.

DESCRIPTION OF RELATED ART Radio communications devices, in particular those used in digital communications systems, typically include a codec, comprising a coder for coding speech signals into a digital signal for transmission and a decoder for retrieving speech signals from received digital signals.

A conventional speech codec is disclosed in the TIA/EIA standard "TDMA Cellular/PCS - Radio Interface Enhanced Full Rate Speech Codec", published under number TIA/EIA/IS-641, May 1996. That document discloses a CELP encoder, which includes a linear predictive coding (LPC) analysis module, in which one part of the coding takes place. In this module, a frame of speech samples, for example lasting 20ms, is processed. A window is applied, covering the frame itself, some samples from the last frame, and some samples from a future frame, and the autocorrelation of the signal is estimated as a measure of the variation of the signal with time. From this estimated autocorrelation, the LPC coefficients, which are used to represent the speech signal, are estimated from an algorithm known as the Levinson-Durbin algorithm, which obtains filter coefficients from the autocorrelations by an iterative process. A new set of LPC coefficients is estimated in this way for every frame.

The LPC coefficients, which are estimated in this way, are then used to encode these speech signals, and are themselves quantized and transmitted to the receiver, for use in decoding the received signal. The estimation of the LPC coefficients is optimised by attempting to minimise the power of a residual signal, namely the difference between the signal received at the coder and the signal as coded, that is the error made by the prediction.

A potential problem which arises from this method of estimating the LPC coefficients is that, because the coefficients are calculated only once per frame, there can be significant discontinuities between consecutive values of the coefficients, associated with consecutive frames of the speech signal.

SUMMARY OF THE INVENTION The present invention seeks to overcome this potential disadvantage of the conventional coder, by estimating the LPC coefficients for each sample of the signal. Because the coefficients are estimated more often, this can result in coefficients which are more continuous, or "smoother" with time, and which therefore allow the signal to be encoded more accurately.

In particular, there is proposed the use of the LMS (Least Mean Square) algorithm for estimating and coding the speech spectrum. This algorithm is relatively simple to implement, and its properties have been thoroughly investigated due to its use in Adaptive Differential Pulse Code Modulation (ADPCM) waveform coding, where it is used for a different purpose, namely to minimise the signal that is applied for quantization.

Aspects of the invention relate to the speech coder, and to the method of encoding the speech signal.

BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a block schematic diagram of an encoder in a CELP transmitter.

Figure 2 is a block schematic diagram of an LPCanalysis module forming part of a speech coder in accordance with the invention.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS Figure 1 is a block schematic diagram of a generally conventional code excited linear prediction (CELP) encoder modified in accordance with the invention.

The purpose of Figure 1 is to give a general overview of a CELP encoder, the nature of which is familiar to a person skilled in the art. In Figure 1, the input X is the true speech signal, x(n) is the predicted speech signal, i iopt is the transmitted code book index, (b,N) ops are the transmitted pitch gain and lag(period), and a is the set of LPC coefficients. In view of the well-known form of the CELP encoder, Figure 1 is not considered to need further explanation.

Figure 2 is a block schematic diagram of a part of the speech coder relevant to the present invention.

Coder 2 is the CELP codec, in which the speech spectrum is estimated and coded by a procedure known as Linear Predictive Coding. The coder 2 is largely conventional, but differs from a conventional device in the form of the LPC-analysis module 4, which is illustrated in more detail. The module 4 receives the same input signals as a LPC-analysis module in a conventional CELP coder, and produces output signals in the same format, which means that the other modules in the CELP encoder, and their interfaces to the LPCanalysis module 4, can be unchanged.

The LPC-analysis module 4 receives the sampled speech signal x(n), which it supplies to a processor 6 and a predictor 8. The processor 6 produces coefficients ai(n), using an algorithm described in more detail below, and these coefficients are used by the predictor 8 to obtain a predicted value x(n) of the signal x(n), according to the equation below:

where ai are the LPC coefficients, n is the sample index and N is the predictor order, that is the number of coefficients which are used to characterise the signal.

The input signal x(n), and the predicted value x(n) thereof, are supplied to a subtractor 10, to obtain a residual value e(n), which represents the error in the prediction, according to the equation: e(n) = x(n) - (n) The residual e(n) is also supplied to the processor 6, which derives the next set of LPC coefficients using the LMS algorithm: ai(n + 1) = a1(n) + A e(n) x(n - i) in which d is the step size parameter determining the amount by which the coefficients are changed in each iteration.

The properties of this algorithm are well-known in themselves.

The LPC coefficients ai(n) are then output from the analysis module 4. Whenever LPC coefficients are to be used in the analysis-by-synthesis loop in the speech coder, or transmitted to the receiver, they are extracted from the LMS algorithm in the LPC module.

Thus, since coefficients are transmitted to the receiver at the same rate as in a conventional device, there is no additional overhead. However, by estimating the LPC coefficients in such a way that they are continuously adapted from the speech signal, or adapted on a sample basis, for example using the LMS algorithm in this illustrated embodiment, consecutive values of the coefficients will have smaller discontinuities, which will mean that the speech spectrum can be followed more accurately over time.

This should result in less degradation in the synthesized speech spectrum.

Claims

1. A speech coder, comprising a linear predictive coder using linear predictive coding coefficients, characterized in that the linear predictive coding coefficients are derived using an algorithm which is adapted on a sample basis.

2. A speech coder as claimed in claim 1, characterized in that the linear predictive coding coefficients are derived using the LMS algorithm.

3. A speech coder as claimed in claim 1 or 2, comprising means for coding a speech signal using the derived coefficients, and means for quantizing the coefficients for transmission.

4. A CELP speech coder, in which the LPC coefficients are estimated on a sample basis.

5. A CELP speech coder as claimed in claim 4, in which the LPC coefficients are estimated using the LMS algorithm.

6. A mobile communications device, including a speech coder as claimed in any preceding claim.

7. A method of coding a speech signal, the method comprising using a linear predictive coder using linear predictive coding coefficients, characterized in that the linear predictive coding coefficients are derived using an algorithm which is adapted on a sample basis.

8. A method of coding a speech signal as claimed in claim 7, characterized in that the linear predictive coding coefficients are derived using the LMS algorithm.

9. A method of coding a speech signal as claimed in claim 7 or 8, comprising coding the speech signal using the derived coefficients, and quantizing the coefficients for transmission.

10. A method of coding a speech signal using a CELP speech coder, in which the LPC coefficients are estimated on a sample basis.

11. A method of coding a speech signal using a CELP speech coder as claimed in claim 10, in which the LPC coefficients are estimated using the LMS algorithm.