Hi-index | 0.00 |
This work aims at improving the discrimination of confusable words like letters. We propose a new method which computes spectral parameters only in a discriminative part of the words and uses artificial neural networks to perform the recognition. Tests have been conducted on clean speech and Lombard speech with and without additive noise. They show a general improvement of the recognition accuracy compared with a continuous density hidden Markov models (HMM)recognition system.