Fully adaptive SVD-based noise removal for robust speech recognition
Kris Hermus, Ioannis Dologlou, Patrick Wambacq, Dirk Van Compernolle · 1999
This paper deals with the problem of the recognition of speech corrupted by additive noise at moderate SNR ratios. The proposed technique - based on Singular Value Decomposition (SVD), and fully adaptive - outperforms well-known approaches as Nonlinear Spectral Estimation and SNR-Normalisation for the recognition of large vocabulary continuous speech. Current techniques for robust speech recognition take advantage of slowly varying and/or accurately modeled environments. Deviations from these prior assumptions greatly compromise the performance. Our new approach is based on SVD and tries to overcome these limitations. This technique automatically removes additive noise by suppressing low energetic, noise related, singular components of the Hankel matrix constructed from the original signal. Providing the SVD-algorithm with prior knowledge about the noise highly improves the efficiency. The algorithm is fully adaptive, and works in real-time. Recognition experiments on a database with large vocabulary, continuous speech (Resource Management) show that the WER is more than halved.