First International Conference on Spoken Language Processing (ICSLP 90)
This paper describes methods for improving the accuracy of our speaker independent speech recognition system in a noisy environment. Phoneme templates were selected from several varieties of phoneme templates, which were generated in typical noise environments, to simulate actual background noise. If none of these templates suits the noise environment, a template adaptation is performed by adding the estimated spectrum of the background noise to the mean vector of each phoneme template. A speech period detector and segmentation algorithm were also improved with respect to changes in environment noise. These approaches were applied to a hardware system and evaluated. Average phoneme recognition score is 71.6% by use of the template adaptation method. Input speech data was achieved by adding a noise to clean speech. Signal-to-noise ratio is 20dB. In a actual noisy environment(70-85dBA), an average word recognition score is 97%, using a twenty four word vocabulary.
Bibliographic reference. Morii, Shuji / Morii, Toshiyuki / Hoshimi, Masakatsu / Hiraoka, Shoji / Watanabe, Taisuke / Niyada, Katsuyuki (1990): "Noise robustness in speaker independent speech recognition", In ICSLP-1990, 1145-1148.