Efficient Utilization of Large Pre-Trained Models for Low Resource ASR

TL;DR

This paper explores efficient methods to leverage large pre-trained models for low-resource automatic speech recognition in medical telephony, demonstrating significant performance improvements through advanced unsupervised techniques and domain adaptation.

Contribution

It introduces novel strategies for adapting large pre-trained models to low-resource telephony ASR tasks, surpassing baselines with refined architectures and minimal in-domain data.

Findings

22% relative improvement over baselines with pretraining techniques

29% additional gains from architecture and training refinements

6% improvement using 0.8 hours of in-domain data

Abstract

Unsupervised representation learning has recently helped automatic speech recognition (ASR) to tackle tasks with limited labeled data. Following this, hardware limitations and applications give rise to the question how to take advantage of large pre-trained models efficiently and reduce their complexity. In this work, we study a challenging low resource conversational telephony speech corpus from the medical domain in Vietnamese and German. We show the benefits of using unsupervised techniques beyond simple fine-tuning of large pre-trained models, discuss how to adapt them to a practical telephony task including bandwidth transfer and investigate different data conditions for pre-training and fine-tuning. We outperform the project baselines by 22% relative using pretraining techniques. Further gains of 29% can be achieved by refinements of architecture and training and 6% by adding 0.8 h of in-domain adaptation data.