Massively Multilingual ASR on 70 Languages: Tokenization, Architecture, and Generalization Capabilities

TL;DR

This paper investigates large-scale multilingual end-to-end ASR models across 70 languages, focusing on tokenization, architecture choices, and their impact on generalization and performance transfer.

Contribution

It introduces an optimal tokenization strategy and compares two architectures, demonstrating significant WER improvements and strong generalization to unseen datasets.

Findings

Multilingual ASR improves WER by 13.9%-15.6% over monolingual models.

Optimal tokenization is crucial for multilingual model performance.

Models generalize well to unseen datasets, achieving 9.5% and 7.5% WER.

Abstract

End-to-end multilingual ASR has become more appealing because of several reasons such as simplifying the training and deployment process and positive performance transfer from high-resource to low-resource languages. However, scaling up the number of languages, total hours, and number of unique tokens is not a trivial task. This paper explores large-scale multilingual ASR models on 70 languages. We inspect two architectures: (1) Shared embedding and output and (2) Multiple embedding and output model. In the shared model experiments, we show the importance of tokenization strategy across different languages. Later, we use our optimal tokenization strategy to train multiple embedding and output model to further improve our result. Our multilingual ASR achieves 13.9%-15.6% average WER relative improvement compared to monolingual models. We show that our multilingual ASR generalizes well on an unseen dataset and domain, achieving 9.5% and 7.5% WER on Multilingual Librispeech (MLS) with zero-shot and finetuning, respectively.