ROBUST SPEECH RECOGNITION USIN INTRA-SPEAKER ADA

Inter-speaker variation can be coped rather well in speech recognition by speaker adaptation techniques such as MLLR and MAP. However, when dealing with speech other than reading style, such as conversational speech, emo-tional speech and so on, current recognition systems can-not achieve a satisfactory performance even after speaker adaptation. In view of this situation, two-level adaptation method was newly proposed, where adaptation technique was applied in two levels to handle inter-speaker and in-tra-speaker variations. A speaker independent model is first adapted to a specific speaker to generate a speaker dependent model. Then, after classifying the training data into several categories, the speaker dependent model is fur-ther adapted to each category using data classified to it (category dependent model). The recognition is done in parallel using the speaker dependent model and each cat-egory dependent model, and the result with highest likeli-hood is selected as the final recognition result. Recognition experiments were conducted for speech with various emo-tions (emotion of input speech is unknown), and the results showed that the proposed method outperformed the con-ventional MLLR-based speaker adaptation. 1