Connectionist Probability Estimators in HMM Arabic Speech Recognition Using Fuzzy Logic

Hidden Markov Models (HMM) are nowadays the most successful modeling approach for speech recognition. However, standard HMM require the assumption that adjacent feature vectors are statistically independent and identically distributed. These assumptions can be relaxed by introducing neural networks in the HMM frame work. These neural networks particularly the Multi-Layer Perceptrons (MLP) estimate the posterior probabilities used by the HMM. We started in the frame work of this work, to investigate smoothing techniques combining MLP probabilities with those from others estimators with better properties for small values (e.g., a single Gaussian) in the framework of the learning of our MLP. The main goal of this paper is to compare the performance of speech recognition of an isolated speech Arabic databases obtained with (1) discrete HMM, (2) hybrid HMM/MLP approaches using a MLP to estimate the HMM emission probabilities and (3) hybrid FCM/HMM/MLP approaches using the Fuzzy C-Means (FCM) algorithm to segment the acoustic vectors.