COMPARAISON OF SEVERAL ACOUSTIC MODELING TECHNIQUES AND DECODING ALGORITHMS FOR EMBEDDED SPEECH RECOGNITION SYSTEMS

In this paper, we present several methods for mapping recognition engine requirements to mobile phone resource. The proposed techniques are evaluated on a digit recognition task using both French and English corpora. We investigate particularly three aspects of speech processing: acoustic parameterization, recognition algorithms and acoustic modeling. Several parameterization algorithms (LPCC, MFCC and PLP) are compared to the Linear Predictive Coding (LPC) included in the GSM norm. The two parameterization algorithms, MFCC and PLP, perform significantly better than the other ones. Moreover, feature vector size can be reduced from 13 to 6 PLP coefficients without a significant lost of performance. We compare Dynamic Time Warping (DTW) and Hidden Markov Model (HMM) based systems for clean conditions. For HMM systems, we propose an automatic building of phonetic lexicon which increases the system robustness to phoneme substitutions. The experiments show that HMM systems outperform DTW in speaker independent task. Nevertheless, the complexity of this approach is significantly higher than the DTW one. Finally, in order to achieve good performance with reasonable resource needs, we try to reduce the HMM model complexity. First, we reduce the number of (Gaussian) components per state for each model. In a second step, we reduce the global amount of Gaussian by tying together all the Gaussian components into a generic speech Gaussian Mixture Model (GMM). The phoneme models are derived from this GMM using a well-known MAP (Maximum A Posteriori) adaptation algorithm