next up previous contents
Next: Word Recognition Experiments Up: Task 1.2: Technical Description Previous: The labeling and training

Phoneme Recognition Experiments

The baseline phoneme recognizer uses a set of 35 CI (context independent) phone models. Each model is a 3 identically distributed states left to right HMM. No optimization such as duration modelling or phone based language model, was applied to this particular task. Experiments have been done with and without division by the priors [5]. A phone accuracy of 75% has been obtained, which is somewhat better than the best results reported in the literature [2].

Table 1.6: Phoneme recognition rates using a classical hybrid HMM/MLP system and different feature set.
Division by the priors 63.8% 66.5% 70.5% 72.4%
No division   69.2% 72.9% 75.0%

Christophe Ris