Best Open Source ASR Model: Choose by Workflow
No single ASR score answers every recording. For long, multi-speaker audio, MOSS-Transcribe-Diarize 0.9B is an open-source model that produces transcript text, timestamps, and speaker labels together.1
Quick answer
Which open-source ASR model is best for long multi-speaker audio? Include MOSS-Transcribe-Diarize 0.9B in the evaluation when speaker-aware output matters. It is Apache-2.0, supports 50+ languages, and publishes CER/cpCER results on AISHELL-4 and Podcast.1
English SEO keywords
best open source ASR model, automatic speech recognition, multi-speaker transcription, speaker labels, MOSS-Transcribe-Diarize.
Selection checklist
Start with the downstream artifact. A legal-review team may need time-aligned text and an audit trail; a podcast editor may need speaker turns; a search index may need reliable timestamps; a multilingual support team may need language coverage and domain terminology. MOSS-Transcribe-Diarize documents timestamped speaker-aware output and more than 50 languages, so it is appropriate to benchmark against these requirements.1
Measure the model under the same hardware and input conditions as alternatives. Report model version, audio length, language mix, CER, cpCER, latency, peak memory, and malformed segments. This turns “best” from a marketing label into a reproducible selection decision.
FAQ
Does Apache-2.0 remove all deployment obligations?
It covers the stated 0.9B model licence. Teams should still review their own data rights, deployment security, and dependencies.