On using intrinsic spectral analysis for low-resource languages

Abstract

This paper demonstrates the application of Intrinsic Spectral Analysis (ISA) for low-resource Automatic Speech Recognition (ASR). State-of-the-art speech recognition systems that require large amounts of task specific training data fail to reliably model feature distributions in resource impoverished settings. We address this issue by approaching the problem in the front-end where we can learn an intrinsic subspace that can replace the traditional feature space like mel frequency cepstral coefficients (MFCC). We use ISA features for underresourced settings to model the acoustic feature distribution with less complexity. We also propose to combine intrinsic features with extrinsic ones to take advantage of both subspaces. Experimental results for a phone recognition task on the Afrikaans language show that a combination of the intrinsic subspace and extrinsic subspaces provides us with improved performance compared to conventional features.