Articulatory Feature Extraction from Voice and Their Impact on Hybrid Acoustic Models

Abstract

There is a great amount of information in the speech signal, although current speech recognizers do not exploit it completely. In this paper articulatory information is extracted from speech and fused to standard acoustic models to obtain a better hybrid acoustic model which provides improvements on speech recognition. The paper also studies the best input signal for the system in terms of type of speech features and time resolution to obtain a better articulatory information extractor. Then this information is fused to a standard acoustic model obtained with neural networks to perform the speech recognition achieving better results.