The performance of speech recognition systems is significantly degraded in the presence of noise. To solve the noise problem, there is a need to reconsider standard approaches by taking into account this new constraint. We first envisage two well-known cepstral representations (parametric and non-parametric) of speech signals and propose a unifying view of both schemes. We introduce a pseudo-autocorrelation domain, which can be interpreted as a ''Root-cepstral domain'', and we show how non-parametric cepstral and linear predictive analyses converge to the same optimal solution. Experiments are carried out using an HMM-based isolated word recogniser for speaker-dependent and speaker-independent tasks in car noise environments.
