Nemotron 3 Diarization
Nemotron 3 Diarization is NVIDIA's model for marking who spoke when. The card body says 100 million parameters, and the sidebar says 99.2 million.
One checkpoint, four printed latencies
The model card, dated 23 September 2026, describes one checkpoint that marks who spoke when, on a live stream or a finished recording. It says that checkpoint can take an input buffer as short as 80 ms, and that the lowest configuration it recommends is 0.32 s. The offline setting on the same card uses a 30.4 s buffer. The NeMo streaming table lists four setups: 30.4 s, 1.04 s, 0.64 s, and 0.32 s. The Transformers streaming table lists the last three, and calls 1.04 s the default. The card's latency is the audio the model waits for, and the note says that wait does not include compute time. Output frames default to every 10 ms, and can be set to another multiple of 10 ms. The tensor has 8 speaker channels, ordered by who started talking first. The card says the model handles up to eight speakers. The DIHARD III eval set on the same page includes recordings with one to nine speakers, including 40 recordings with five to nine.
100 million in the text, 99.2 million on the side
The architecture section says 100 million parameters, written as 1.0 × 10⁸. The card sidebar says 99.2 million parameters, in 32-bit floats. The encoder is a 31-layer transformer. Input is 16 kHz, one channel, as wav, flac, opus, or mp3. With chunked inference, the card says duration is not limited. The preferred operating system line says Linux. The speed table was measured on an NVIDIA RTX PRO 5000, in BF16. At 0.32 s, with a batch of one, it prints a real-time factor of 12.5 without compile and 54 with compile. Training used 8 nodes of eight A100 80GB GPUs: first offline training on simulated mixtures only, then a streaming fine-tune that adds real conversations. The card counts about 10,000 hours of real conversation and 82,611 hours of simulated multi-talker audio.
Error rate and headcount do not move together
The comparison model is diar_streaming_sortformer_4spk-v2.1, named as a 4-speaker model. The card says to reproduce the scores with the reference labels it links. Collar is 0 seconds for AMI and 0.25 seconds for CALLHOME, so those tables are separate scales. On AMI Test with mixed headsets, at 30.4 s, diarization error goes from 15.81 to 9.25, and speaker-count accuracy goes from 93.75 to 87.50. On CALLHOME with two speakers, at 30.4 s, the 4-speaker model's error is 5.68 and this model's is 5.98. On the full CALLHOME set at 30.4 s, this model's error is 9.10 and the baseline's is 10.32. On full DIHARD III at 30.4 s, error goes from 19.09 to 12.73, and speaker-count accuracy goes from 75.29 to 81.47.
The quick-start script sets a chunk of 340 frames, 40 frames of right context, a FIFO of 40, and a speaker-cache update every 300 frames. That row is the 30.4 s setup. Mel features start at 10 ms, are stacked by eight into an 80 ms encoder frame, and a convolution brings the prediction back to 10 ms. On AliMeeting near-field, at 30.4 s, error goes from 11.57 to 6.40 and speaker-count accuracy goes from 80 to 90. The version line says General Access. Languages named for the training audio include English, Mandarin, Hindi, Kannada, Telugu, and Bengali.
Three ways to run it
The NeMo path asks for Python 3.12 or newer, Cython, and a recent PyTorch. The install line is nemo-toolkit[asr]. A load example on the card says a Hugging Face token is required to load the weights from the card. NeMo-Speech.cpp exposes nemo-speech diarize for a wav file, and the same command can add speaker tags to a transcript. Transformers support is written as an install from the Transformers git repository. The card links a demo at nvidia/nemotron-diarization, and says the model is not deployed by any inference provider. It prints no dollar price. Use is governed by the OpenMDW License Agreement, version 1.1, and the card says the model is ready for commercial or non-commercial use. Several David AI training sets on the same page are marked licensed under agreement. The NeMo configs and the evaluation script are in NVIDIA-NeMo/Speech.