Acquiring variable length speech bases for factorisation-based noise robust speech recognition

Studies from multiple disciplines show that spectro-temporal units of natural languages and human speech perception are longer than short-time frames commonly employed in automatic speech recognition. Extended temporal context is also beneficial for separation of concurrent sound sources such as speech and noise. However, the length of patterns in speech varies greatly, making it difficult to model with fixed-length units. We propose methods for acquiring variable length speech atom bases for accurate yet compact representation of speech with a large temporal context. Bases are generated from spectral features, from assigned state labels, and as a combination of both. Results for factorisation-based speech recognition in noisy conditions show equal or better separation and recognition quality in comparison to fixed length units, while model sizes are reduced by up to 40%.