A multi-task neural network approach to speech recognition

Improving the speaker-independent generalization exhibited by neural network approaches to phoneme identification is an area of continuing interest to speech recognition researchers. The author reports research exploring the combined impact of multiple task constraints and differing speech input representations on network generalization. The multiple tasks required of the networks are based on a psychological model of speech perception. Using 12 American vowels to train and test the networks, the differing input representations are motivated by current theories of vowel perception and human audition. Network results compare favorably to baseline performance results established by a K-nearest neighbor classification and the classification performance of human listeners on the same task. These results are also extremely good when compared to performance reported by other researchers