Adaptive noise immunity learning for word spotting

We have developed a word spotting method using adaptive noise immunity learning to achieve robust performance in noisy environments. Previously we proposed a noise immunity word spotting method based on the multiple similarity (MS) method. We also proposed a keyword-based spontaneous speech understanding method and constructed a real-time speech dialogue system, TOSBURG II, based on the noise immunity keyword spotting. By enhancing noise robustness against high-level nonstationary noise in various locations, we have extended noise immunity learning so that it can adapt to noise on-line. In the adaptation process, noisy speech data is synthesized by contaminating pure speech data in a speech database with on-line pick-up background noise data. The synthesized data is then used to modify word reference vectors to adapt to a time-variant noise environment. Massively parallel computers enable real-time adaptation for noise immunity learning. Experimental results indicate the effectiveness of the proposed method