| Computer technologies have improved significantly in both capacity and speed. The human-computer interface still lags behind, in that human-computer interaction lacks the naturalness essential for efficient communication. In order to make human-computer interaction more natural, novel sensory modalities have been used. Speech, gestures, and emotional states can be detected and understood to some extent. However, integrating these information sources to achieve performances superior to any single modal alone remains a challenging problem. Humans naturally and effortlessly perform sensory information fusion. The most important human-to-human communication tool is speech. Automatic speech recognition by machines has been able to achieve very high recognition accuracy for large vocabulary sets and speaker-independent tasks. However, in some environments, unexpected sources of noise will degrade system performance. We propose a novel integration technique that can efficiently incorporate both visual and acoustic speech signals to achieve better speech recognition accuracy than that achieved by either of the single modalities alone. The proposed fusion schemes have been tested in different situations. The experiment results show consistently improved performance by using multimodal fusion schemes. |
