Significance of Indic Self-supervised Speech Representations for Indic Under-Resourced ASR

Abstract

Automatic speech recognition is a mature speech technology, almost able to attend human label recognition performance conditioned on the availability of sufficient labeled training data. However, the performance of the system struggles to achieve deployable performance in the under-resourced scenario. In such a scenario, most of the work suggests traditional frameworks are preferable over state-of-the-art deep learning frameworks. This work creates a dataset for the Lambani language of 6 hours duration, and attempts to develop an ASR system. The system provides a character error rate (CER) of $39.1\%$ and $24.1\%$ using the GMM-HMM framework and TDNN framework, respectively for Lambani dataset. The language doesn’t have enough publicly available speech and corresponding text transcription resources of its own. Motivating by the same, this work uses the publicly available wav2vec2.0 (W2V) pre-trained model (trained on 23 Indian languages’ unlabeled speech data) and fine-tuned it with the labeled data of the Lambani language. After that using the fine-tuned framework as a non-linear feature extractor, the ASR task is performed with GMM-HMM and TDNN framework. The proposed approach provides a relative improvement of $53.4\%$ and $32.1\%$ for the GMM-HMM and TDNN frameworks, respectively.