Reverb

Reverb is Rev's release of the speech models behind its transcription business: an ASR model built on the WeNet framework plus speaker diarization built on Pyannote, trained on 200,000 hours of human-transcribed English audio, which Rev describes as the largest such corpus ever used for an open model. A distinctive verbatimicity control slides output between strict verbatim, keeping every false start and filler, and a cleaned-up reading style suited to captions. On the long-form Earnings21 benchmark the ASR scores 9.68 percent word error rate against 14.26 percent for Whisper large-v3, and the diarization component reaches a 0.046 error rate on the same data. The pipeline installs with pip on Python 3.10 or newer or runs through Docker, and pulling the weights requires Hugging Face authentication plus Git LFS; a GPU is the practical choice for long-form workloads. The inference code is Apache-2.0, but the weights ship under the Rev Model Non-Production License, which limits them to testing, research, and evaluation, with commercial terms available from Rev, whose paper is at arXiv 2410.03930.