Acoustic model adaptation in noisy and reverberated scenarios using multi-task learned embeddings

Distortions, such as revebererations and background noise in an acoustic environment pose a significant challenge for large-scale Automatic Speech Recognition (ASR) systems. In this paper we explore acoustic model adaption, to be used in conjuction with multi-conditioning, to address this challenge. Recently, it has been standard to use i-vectors for Acoustic Model (AM) adaptation. With a varying degree of success, Single Task Learned (STL) neural embeddings such as x-vector and r-vector, have also been used for AM adaptation. This paper proposes the use of Multi Task Learned (MTL) embeddings for large vocabulary hybrid acoustic model adaptation in noisy and reverberant environments. Our experiments on noise added Librispeech database as well as real reverberated HRRE database show that this approach provides significant performance gain when used along with the traditional i-vectors.