A Measure of Speech and Pitch Reliability from Voicing

We propose a CASA labelling method of the TF representation, which is based on the periodicity of the speech, related to the voicing. A local voicing index is estimated in four subbands after demodulation of the signal. This index is used as a reliability measure for both pitch identification and speech recognition. First, this model allows robust f0 identification thanks to the voicing index, which is a consistent reliability measure associated to the f0 measure. Since the task is to recognise speech corrupted with additive noise, the periodicity is specific to the signal. In our model, the evaluation of "speech reliability" is not direct. It also depends on a priori knowledge about the relationship between SNR and the voicing index. The goal is to assign to each TF region a probability "clean" enough to feed a multistream recogniser only adapted to clean data. The model is able to localise regions where the target speech dominates over the background noise. This probability is evaluated according to a function established a priori, the SNR-feature mapping, and the choice of a SNR decision threshold. This is adapted to a new multistream recognition approach [7], since the previous probabilities serve to weight the streams' posteriors.