LeBenchmark: A Reproducible Framework for Assessing Self-Supervised Representation Learning (SSL) from Speech

LeBenchmark is a reproducible and multifaceted benchmark for evaluating speech SSL models. We propose an initial set of four tasks, measuring specific speech challenges in French language: Speech Recognition (ASR), Spoken Language Understanding (SLU), Speech Translation (AST) and Emotion Recognition (AER).

For reproducibility, we also provide pre-trained SSL models learnt on different subsets of a large and heterogeneous collection of French speech utterances (read, prepared, and spontaneous). The gathered collection is diverse, informed with relevant metadata and each sub-corpus’s license is clearly documented in our published paper.

A leaderboard will be updated frequently here.