A new approach for speech recognition

Year 2003
Project team Thomas Knight

Improved machine communication

Speech is our most natural communication medium, but difficulties with present-day speech-recognition algorithms prevent its widespread use in our interaction with machines. Current algorithms have extreme difficulty in noisy environments and with new speakers. Speech recognizers have high error rates – even the best in ideal conditions misunderstand 1 in 20 words, with recognition dropping to 50% in noisy environments. They also have difficulty with previously unheard speakers and dialects. These failures, which suggest that we don’t really yet understand what differentiates particular sounds in the language, have limited the usefulness of current speech recognition technology. This project will develop a novel, non-linear representation for speech that addresses the shortcomings of existing representations. It will greatly improve the ability of machines to understand human speech and thus make communicating with machines as easy as talking on the phone.