Reproduce “Reverb: Open-Source ASR and Diarization from Rev” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nishchal Bhandari, Danny Chen, Miguel Ángel del Río Fernández, Natalie Delworth, Jennifer Drexler Fox, Migüel Jetté, Quinten McNamara, Corey Miller and 5 more
Rev open-sourced Reverb, a suite of Automatic Speech Recognition (ASR) and diarization models trained on the world's largest human-transcribed English dataset. These models demonstrate superior performance on long-form audio benchmarks compared to existing open-source alternatives and introduce a unique verbatimicity control feature for ASR output.
Key insights
- Training on an unparalleled volume of high-quality, human-transcribed real-world data is critical for achieving state-of-the-art performance in challenging long-form speech recognition and diarization.
- Architectural modifications within existing frameworks, such as adapting language-specific layers for verbatimicity control, can provide nuanced and practically useful output customization.
- A comprehensive production pipeline, including advanced decoding and post-processing, is essential to maximize the performance of core ASR models on complex, long-form audio.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.