XTREME-S: Evaluating Cross-lingual Speech Representations

TL;DR

XTREME-S is a comprehensive benchmark designed to evaluate universal cross-lingual speech representations across 102 languages, covering multiple tasks to advance research in multilingual speech understanding.

Contribution

The paper introduces XTREME-S, a new benchmark for evaluating speech models across many languages and tasks, with established baselines using XLS-R and mSLAM.

Findings

First speech-only and speech-text baselines on all tasks

Benchmark covers 102 languages from diverse families

Facilitates evaluation and development of universal speech models

Abstract

We introduce XTREME-S, a new benchmark to evaluate universal cross-lingual speech representations in many languages. XTREME-S covers four task families: speech recognition, classification, speech-to-text translation and retrieval. Covering 102 languages from 10+ language families, 3 different domains and 4 task families, XTREME-S aims to simplify multilingual speech representation evaluation, as well as catalyze research in "universal" speech representation learning. This paper describes the new benchmark and establishes the first speech-only and speech-text baselines using XLS-R and mSLAM on all downstream tasks. We motivate the design choices and detail how to use the benchmark. Datasets and fine-tuning scripts are made easily accessible at https://hf.co/datasets/google/xtreme_s.