Fast channel and noise compensation in the spectral domain

We compare in this work several methods for fast adaptation of speech models to convolutional and additive noise. The tested algorithms are Parallel Model Combination (PMC), Cepstral Mean Subtraction (CMS), and an algorithm that combines PMC and CMS in the spectral domain. Experiments are realized on a natural numbers recognition task in French. We have trained the acoustic models on the SPEECHDAT database (recorded through telephone lines), and we have tested the system on the VODIS database (recorded in three different cars).