Root homomorphic deconvolution schemes for speech processing in car noise environments

It is well known that cepstrum parameters are particularly adapted in speech recognition. Nevertheless, the severe whitening effect of the logarithmic spectrum compression used in cepstral analysis is very sensitive to noise contamination. We propose the use a Root homomorphic approach instead of the classical Log, in order to improve cepstral representation in noise. Recognition tests demonstrate the robustness of the Root-based analysis for noisy speech processing. We show empirically that Root cepstral coefficients permit a heuristic extension of noise subtraction techniques into the cepstral domain. Finally, we extend our work to LP-based cepstral analysis and show that the Root approach gives equivalent results when comparing direct (LFCC) and LP-based (LPCC) analysis schemes. A unified model is then proposed and we conclude that both classical LPCC and LFCC analyses are sub-optimal solutions of our more general scheme.