New time-frequency derived cepstral coefficients for automatic speech recognition

The goal is to improve recognition rate by optimisation of Mel Frequency Cepstral Coefficients (MFCCs): modifications concern the time-frequency representation used to estimate these coefficients. There are many ways to obtain a spectrum out of a signal which differ in the method itself (Fourier, Wavelets,…), and in the normalisation. We show here that we can obtain noise resistant cepstral coefficients, for speaker independent connected word recognition. The recognition system is based on a continuous whole word hidden Markov model. An error reduction rate of approximately 50% is achieved. Moreover evaluation tests demonstrate that these results can be obtained with smaller databases: halving the training database have small effects on recognition rates (which is not the case with traditional MFCCs).