Transcribe long recordings with every speaker labeled

You hand it an audio or video file, up to 90 minutes long, and it returns a transcript where every line carries a timestamp and a speaker tag: the first voice it hears becomes [S01], the second becomes [S02], and each line is marked with the tag of whoever is talking. It does not know anyone's name, but you can follow the whole conversation voice by voice. Most tools do this with two separate systems, one writing down the words and another guessing who spoke, then gluing the results together. Here a single model decides what was said, when, and by whom in one listen. It handles more than 50 languages, and you can feed it hotwords so domain terms like product names or people's names come out spelled right.

Behind this is a 0.9B model from OpenMOSS, the Fudan University lab. Small as it is, it beats the transcription services from ElevenLabs, Doubao, and Gemini on speaker-attributed error rates across meeting, podcast, and movie test sets. It can also mark sounds that are not speech, like laughter or applause, so a summary tool reading the transcript knows what was happening in the room.

The demo Space is the fastest test: upload a recording where two people talk over each other, and watch whether the [S01] and [S02] labels stay consistent through the crosstalk. That is the exact spot where stitched-together pipelines fall apart.

For local use the weights load through Transformers with trust_remote_code, and there are OpenAI-compatible endpoints via vLLM or SGLang. The GitHub package even ships a small subtitle web app: upload a video, review the segments, export SRT or burn the subtitles into an MP4.

Speaker labels are anonymous and only consistent within one file: [S01] here is not [S01] in your next recording, and mapping labels to real names stays your job. On long multi-speaker audio you also need to raise the generation token limit, or the transcript stops before the recording ends.

• Turn meeting recordings into minutes where every point has a name-ready speaker tag.
• Draft podcast show notes with accurate speaker turns and timestamps.
• Transcribe interviews without manually marking who is asking and who is answering.
• Generate speaker-labeled subtitles for videos with the bundled subtitle app.
• Feed structured, speaker-aware transcripts into your own summarization workflow.

Try the demo

View model page

Read this issue on neuralexpedition.com →