First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

A Speech Recognition Method for Noise Environments Using Dual Inputs

Yoshio Nakadai, Noboru Sugamura

NTT Human Interface Laboratories, Kanagawa, Japan

This paper describes an improved speaker-dependent isolated word recognition algorithm that overcomes the effect of variable noise environments. The noisy speech is received by dual inputs; primary and reference. Speech endpoints are detected by the power ratio of these two inputs. The noise power spectrum that overlaps the primary speech input is offset by the reference input noise spectrum where the spectrum is restricted to values lower than a mean noise spectrum. The Staggered Array DP algorithm is used for pattern matching and the inverse-variance weighted distance measure of LPC cepstrum and its' regression coefficients are applied as a spectral distance measure between the input pattern and the reference templates. These algorithms are applied to noisy speech uttered in an actual passenger car. Tests confirm the superiority of this approach over the conventional algorithms that use single speech input.

Full Paper

Bibliographic reference.  Nakadai, Yoshio / Sugamura, Noboru (1990): "A speech recognition method for noise environments using dual inputs", In ICSLP-1990, 1141-1144.