Vowel, Digit and Continuous Speech Recognition Based on Statistical, Neural and Hybrid Modelling by Using ASRS_RL

In the first part of this paper a recognizer based on hidden Markov models (HMMs) is compared in the simple task of vowel recognition with a recognizer based on the multilayer perceptron (MLP). In this situation, we have obtained better results for the last recognizer, fact which highlights the advantage of the discriminative training of the perceptron versus the maximum likelihood training of the HMM. Because MLPs have problems with accommodating time sequences like speech, a combination of a HMM with a MLP could be a good idea. In the second part of the paper, the hybrid structure HMMMLP is compared with the simple HMM in a digit recognition task. The hybrid structure has recognition rates improved with around 2%. In the last part of the paper are describes the continuous speech recognition experiments for Romanian language, by using HMM modelling. The progresses concern enhancement of modelling by taking into account the context in form of triphones, improvement of speaker independence by applying a gender specific training and enlargement of the feature categories used to describe speech sequences. In order to easier handling the recognition experiments an Automatic Speech Recognition System for Romanian Language (ASRS_RL) was designed.