Noise-tolerant speech recognition: the SNN-TA approach