The Recognition of Bimodal Produced Speech based on Multi-style Training

Abstract

In this paper an analysis on the recognition of words from speech database Whi-Spe in normal and whispered phonation, based on the conventional HMM/GMM framework, is presented. The analysis based on multi-style training is performed in the speaker dependent (SD) and speaker independent mode (SI). The analysis showed that a small portion of whisper data in training (10%) is required for the recognition of whisper higher than 90%, for both the SD and SI recognition.