Influence de la corrélation entre le pitch et les paramètres acoustiques en reconnaissance de la parole
Gwenael CLOAREC, Denis Jouvet, Jean MONNE.
In this paper we compare the role played by the pitch frequency on speaker independent speech recognition performances for two tasks: an isolated word recognition task and a continuous speech recognition task. While introducing pitch and/or voicing directly into the acoustic vector leads to significant improvements on the isolated word recognition task, this method does not bring any improvement on the continuous speech recognition task. On the contrary, modelling the pitch frequency independently of the acoustic parameters leads to small but similar improvements on the two tasks. Those results could be explained by the fact that the improvement brought when introducing pitch and voicing directly into the acoustic vector is related to the correlation between the pitch frequency and the acoustic parameters. This correlation is much less important in the case of the continuous speech recognition task in which prosody can lead to very different pitch values depending on the prosodic context.