Method for doping (impurity introduction) unsupervised and for rejection of OOV words in speech recognition.

摘要

Procedure speech recognition comprising a step of Viterbi decoding type in a network Markov for a list of N candidates, characterized in that is determined for each phoneme source Ps an associated set EAPD of phonemes Pd very differentiable phoneme cited source Ps, and in that each of said candidates is subject to a treatment consisting in comparing each of these phonemes with the phonemes of said associated set EAPD and in rejecting said candidate if at least one of the evaluated phonemes is confused with one of said phonemes of said associated set.