Êtes-vous un étudiant de l'EPFL à la recherche d'un projet de semestre?
Travaillez avec nous sur des projets en science des données et en visualisation, et déployez votre projet sous forme d'application sur Graph Search.
Phonological studies suggest that the typical subword units such as phones or phonemes used in automatic speech recognition systems can be decomposed into a set of features based on the articulators used to produce the sound. Most of the current approaches to integrate articulatory feature (AF) representations into an automatic speech recognition (ASR) system are based on a deterministic knowledge-based phoneme-to-AF relationship. In this paper, we propose a novel two stage approach in the framework of probabilistic lexical modeling to integrate AF representations into an ASR system. In the first stage, the relationship between acoustic feature observations and various AFs is modeled. In the second stage, a probabilistic relationship between subword units and AFs is learned using transcribed speech data. Our studies on a continuous speech recognition task show that the proposed approach effectively integrates AFs into an ASR system. Furthermore, the studies show that either phonemes or graphemes can be used as subword units. Analysis of the probabilistic relationship captured by the parameters has shown that the approach is capable of adapting the knowledge-based phoneme-to-AF representations using speech data; and allows different AFs to evolve asynchronously.
Ronan Collobert, Dimitri Palaz