Emotion-detecting Based Model Selection for Emotional Speech Recognition

As known to all, the performance of speech recognition degrades dramatically in the presence of emotion. How to deal with emotion issue properly is crucial. Most widely used approaches include robust feature extraction, speaker normalization and model tuning/retraining. In the study, a novel method is proposed, that is, adaptation technique is adopted to transform a general model into emotion-specific one with a small amount of emotion speech. Moreover, a model-selection strategy based on emotion-detection was proposed and proven to be effective, and the overall mean recognition rate increased to 80.79% with an Error Rate Reduction (ERR) of 16.55% compared to the neutral speech Acoustic Model (AM).