Feature sets in continuous speech recognition for the Portuguese language

We evaluate the performance of different feature sets in continuous speech recognition systems for the Portuguese language. Results were obtained for the task of recognizing sequences of digits spoken in a fluent manner. We have investigated five parametric descriptions of speech, selected among the most-used ones in present continuous speech recognition systems. We show that the feature set providing the best results for the Portuguese language comprises 18 parameters, 15 derived from the PLP-cepstrum and 3 from the energy. In the speaker-independent mode, a word accuracy of 99.5% was obtained. The performance of a Mel-cepstrum-based set with 39 parameters was 99.3% word-accurate.