Abstract: A two-stage detection method was proposed for the identification of vocal effort modes in robust speech recognition. Firstly, whisper identification of speech signal is performed by using Gaussian mixture model(GMMs) which are trained by global spectrum features. Secondly, vowels are acquired based on landmark detection for the speech signal which does not belong to the whisper mode, and the vocal effort mode of the speech signal is determined by vowel template matching. Experiments conducted on 863-test show that, accompanied by a slight decline for whisper mode, the significant improvement of recognition accuracy for the remaining four vocal effort modes can be achieved.

Key words: speech recognition, vocal effort, vowel, adaptive modulation, Gaussian mixture model