This project studies
the nonlinear techniques used for prediction of speech in case of low delay
coders. Speech application conventionally requires the computation of linear
prediction (LP) models for modeling the vocal tract. LPC requires many
calculations to be done in floating point; if your machine does not have a math
coprocessor, it will almost certainly be unable to do LPC compression and
decompression in real time. LPC compression is extremely sensitive to high
frequency noise and clipping caused by setting the audio input level too high.
Users with high-pitched voices may not be able to use LPC compression as it
loses too much high-frequency information
The
frequency response of LPC (or AR) coefficients ak are is very sensitive to small changes in ak (such as quantizing errors in
coding). There is no easy way to verify that the filter is stable.
Interpolating between the parameters that correspond to two different filters
will not vary the frequency response smoothly from one to the other.
The LPC coder is
optimized by using a Codebook (look up table) to find the best match for the
signal. This method reduces the processing complexity and the required data
transmission rate. Most digital cellular systems use Codebook excited linear
prediction (CELP) coders. But the CELP coders have the following disadvantages:
Processing delay
• Analysis of each block of digitized samples at the
encoder
• Reconstruction at the decoder
• Time to buffer the block of samples
• May include lookahead - buffering samples from the next successive block
LP techniques can
predict the analysis order but it is unable to model the nonlinearities
involved in the speech production mechanism. Using nonlinear prediction the
speech signal is better fit and can be adapted more easily to the application.
Nonlinear prediction is carried out using local linear prediction (LLP) and gives improved performance when compared to linear prediction methods in terms of gain and residual signal. The high bit rate used to transmit the prediction parameters is reduced by using an adaptive CELP coding algorithm. Unlike linear prediction, which optimizes parameters over a continuous time frame of speech data, the method of local prediction optimizes the parameters over local volumes of state space.The LLP method gives better predictor gain and ‘whiter’ residual as compared to the conventional linear predictors.
References:
[1] Arun Kumar and
Allen Gersho,”LD-CELP Speech Coding with
Nonlinear Prediction”, IEEE Signal
Processing Letters, Vol. 4, No.4,
April 1997
[2] A.C. Singer,
G.W. Wornell and A.V. Openheim, “Codebook Prediction:
A nonlinear signal modeling paradigm” in Proc.
Int. Conf. Acoustics,
Speech, Signal Processing,
1992,Vol 5,pp. 325-328