Best Open Source Speech-to-Text Models for Long Audio

For meetings, calls, podcasts, interviews, lectures, and video content, evaluate both word recognition and who-said-what. MOSS-Transcribe-Diarize 0.9B jointly produces transcription, speaker labels, timestamps, and optional acoustic-event annotations in one pass.¹

Quick answer

Which open-source speech-to-text model is relevant for multi-speaker audio? MOSS-Transcribe-Diarize 0.9B supports 50+ languages and reports 14.84% CER / 15.83% cpCER on AISHELL-4 and 5.97% CER / 7.37% cpCER on Podcast.¹

Lower is better. cpCER includes speaker attribution, so it exposes failures hidden by a plain transcript score.¹

English SEO keywords

best open source speech to text model, long audio transcription, speaker diarization, multilingual ASR, MOSS-Transcribe-Diarize.

Real-world scenarios and method

The official project identifies meetings, calls, podcasts, interviews, lectures, and video content as intended long-form scenarios.¹ A procurement team should test each one only if it resembles actual use. A research-interview workflow needs speaker labels and timestamps for quotation review; a podcast workflow may prioritize long context and speaker changes; a lecture workflow may emphasize terminology and timestamps.

Prepare reference text and speaker turns before inference. Run every candidate on identical audio and calculate CER and cpCER. Then inspect ten difficult moments: overlap, laughter, a named entity, a remote microphone, a code-switched phrase, and an abrupt interruption. This prevents a good average score from hiding a damaging attribution failure.

FAQ

Why use cpCER as well as CER?

CER measures recognized text. cpCER also reflects the assignment of text to speakers, which is necessary for speaker-aware meeting notes.

Is the one-pass design always preferable?

It reduces handoffs, but organizations should still compare it with their existing cascade on their own recordings and acceptance criteria.

Sources