BigSSL: Exploring the Frontier of Large-Scale Semi-Supervised Learning for Automatic Speech Recognition

TL;DR

This paper demonstrates that large-scale pre-training and self-training significantly improve speech recognition performance and data efficiency, achieving state-of-the-art results across various tasks with massive unlabeled datasets.

Contribution

It introduces a large-scale semi-supervised learning approach using billion-parameter models, showing substantial improvements in ASR and downstream tasks over prior methods.

Findings

Pre-training with large datasets boosts data efficiency in ASR.

Fine-tuning an 8B parameter model matches state-of-the-art with only 3% of labeled data.

Large pre-trained models improve performance across diverse speech tasks.

Abstract

We summarize the results of a host of efforts using giant automatic speech recognition (ASR) models pre-trained using large, diverse unlabeled datasets containing approximately a million hours of audio. We find that the combination of pre-training, self-training and scaling up model size greatly increases data efficiency, even for extremely large tasks with tens of thousands of hours of labeled data. In particular, on an ASR task with 34k hours of labeled data, by fine-tuning an 8 billion parameter pre-trained Conformer model we can match state-of-the-art (SoTA) performance with only 3% of the training data and significantly improve SoTA with the full training set. We also report on the universal benefits gained from using big pre-trained and self-trained models for a large set of downstream tasks that cover a wide range of speech domains and span multiple orders of magnitudes of dataset sizes, including obtaining SoTA performance on many public benchmarks. In addition, we utilize the learned representation of pre-trained networks to achieve SoTA results on non-ASR tasks.