Automatic Speech Recognition and Identification of African Portuguese

This document deals with speech recognition of different Portuguese varieties, it resumes results from the author’s diploma thesis [9]. The performance of a hybrid large vocabulary continuous speech recognizer, which combines multi-layer perceptrons and Hidden Markov Models, degrades heavily in the presence of African Portuguese varieties in broadcast news. Variety-specific acoustic and language models are shown to improve recognition significantly by up to 21.1%, from 30.1% WER to 23.7% WER. Further, this document discusses a novel and efficient approach to automatically distinguish African from European Portuguese, first presented in [8] [10]. The phonotactic variety identification system, based on phone recognition and language modeling, focuses on a single tokenizer that combines distinctive knowledge about differences between the target varieties. This knowledge is introduced into a multi-layer perceptron phone recognizer by training variety-dependent phone models for two varieties as contrasting classes. Significant improvements were achieved, lowering the computational cost and reducing the equal error rate by more than 60%, from 11.4% EER to 4.1% EER, compared to conventional single and fused phonotactic and acoustic systems. The approach is extended to cover Brazilian Portuguese, where it also shows high variety identification performance.