ESC Benchmark

The End-to-end Speech Challenge (ESC) is a benchmark for assessing ASR systems on a collection of eight speech recognition datasets. ESC consists of:

                🤗 Datasets
        

    📜 Official Checkpoints
        

    🏆 ESC Leaderboard
        

The ESC datasets are sourced from 11 different domains and cover a range of audio and text distributions (speaking styles, background noise, transcription requirements). There is no restriction on architecture or training data: any system capable of processing audio inputs and generating the corresponding transcriptions is eligible to participate. The only constraint is that the same training and evaluation algorithms must be used across datasets and systems may not use any dataset-specific pre- or post-processing. The objective of ESC is to encourage the research of more generalisable, multi-domain ASR systems.

ESC was proposed in ESC: A Benchmark For Multi-Domain End-to-End Speech Recognition by ... For more information, see the official submission on OpenReview.net or the blog post at ESC Benchmark (TODO).