GigaSpeech 2: An Evolving, Large-Scale and Multi-domain ASR Corpus for Low-Resource Languages with Automated Crawling, Transcription and Refinement

2024-06-17

arXiv.org

AI Summary

To address the performance limitations of automatic speech recognition (ASR) for low-resource languages caused by scarce labeled data, this paper proposes an end-to-end automated multilingual corpus construction framework. Leveraging unlabeled YouTube videos, it integrates Whisper-based initial transcription, MMS-based forced alignment, and multi-dimensional quality filtering to generate high-fidelity pseudo-labels. We further introduce an enhanced Noisy Student self-training paradigm to enable iterative model refinement. To our knowledge, this is the first work to significantly outperform Whisper-large-v3 and leading commercial ASR systems using a lightweight model with only 10% of their parameter count. On real-world YouTube test sets for Thai, Indonesian, and Vietnamese, our approach achieves average WER reductions of 25–40%. The framework establishes a novel, efficient, and scalable paradigm for both data curation and modeling in low-resource ASR.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

The evolution of speech technology has been spurred by the rapid increase in dataset sizes. Traditional speech models generally depend on a large amount of labeled training data, which is scarce for low-resource languages. This paper presents GigaSpeech 2, a large-scale, multi-domain, multilingual speech recognition corpus. It is designed for low-resource languages and does not rely on paired speech and text data. GigaSpeech 2 comprises about 30,000 hours of automatically transcribed speech, including Thai, Indonesian, and Vietnamese, gathered from unlabeled YouTube videos. We also introduce an automated pipeline for data crawling, transcription, and label refinement. Specifically, this pipeline involves Whisper for initial transcription, MMS for forced alignment, and multi-dimensional filtering for data quality assurance. A modified Noisy Student Training is developed to further refine flawed pseudo labels iteratively, thereby enhancing model performance. Experimental results on our manually transcribed evaluation set and two public test sets from Common Voice and FLEURS confirm our corpus's high quality and broad applicability. Notably, ASR models trained on GigaSpeech 2 can reduce the word error rate for Thai, Indonesian, and Vietnamese on our challenging and realistic YouTube test set by 25% to 40% compared to Whisper large-v3, with merely 10% model parameters. Furthermore, our ASR models trained on GigaSpeech 2 yield superior performance compared to commercial services. We hope that our newly introduced corpus and pipeline will open a new avenue for low-resource speech recognition and significantly facilitate research in this area.

Problem

• Addresses low-resource language ASR with automated data pipeline
• Reduces word error rates for Thai, Indonesian, Vietnamese
• Enhances model performance with iterative pseudo-label refinement

Innovation

• Automated pipeline for data crawling and transcription
• Multi-dimensional filtering for quality assurance
• Modified Noisy Student Training for label refinement