Décodage avec ancrage pour la reconnaissance automatique de la parole
Daniel Moraru, Guillaume Gravier.
Automatic speech recognition mainly rely on hidden Markov models (HMM)
which makes little use of phonetic knowledge. As an alternative,
landmark based recognizers rely mainly on precise phonetic knowledge
and exploit distinctive features. We propose a theoretical framework
to combine both approaches by introducing prior (phonetic) knowledge
in a non stationary HMM decoder. As a case study, we investigate how
broad phonetic landmarks can be used to improve a HMM decoder by
focusing the best path search. We show that every broad phonetic class
bring a small improvement, the best improvement being obtained with
glides. Using all broad phonetic classes brings a significant
improvement by reducing the error rate from 22% to 14% on a broadcast
news transcription task. We also experimentally demonstrate that
landmarks does not need to be detected with precise boundaries and can
be used to fasten the beam search algorithm.