September 23, 2026
Conversation transcription will distinguish up to 8 voices: Nemotron 3 Diarization
On September 23, 2026, NVIDIA released Nemotron 3 Diarization: 99.2 million parameters for 8 speakers. The open-weight model labels audio in streaming and after recording, returning segments with a start time, end time, and speaker number.
In Voice Arena, the model ranked first among 12 systems: 14.72% speaker-diarization errors versus 19.3% for the next system across 139 English-language conversations totaling about 22 hours. On DIHARD III, Nemotron 3 Diarization with a 1.04-second buffer delivered 13.18% errors versus 19.60% for the previous Sortformer.
From file to labels. Previously, this kind of labeling required assembling the stack yourself. Now NeMo installs with `uv pip install 'nemo-toolkit[asr]'` after Python 3.12, Cython, a recent PyTorch version, libsndfile1, and ffmpeg, while `diarize()` returns utterance boundaries and a speaker index.
For streaming. NVIDIA provides a separate stack with speech recognition: the input file must be mono at 16 kHz. Before processing, profiles wait for 0.32, 0.64, or 1.04 seconds of audio; computation time is not included in those figures.
For streaming transcription, NVIDIA has already paired the model with `multitalker-parakeet-streaming-0.6b-v1` or `nemotron-3.5-asr-streaming-0.6b`.
