01
Before you run the checkpoint
The final checkpoint is nvidia/Nemotron-3-Diarization on Hugging Face. NVIDIA documents 16 kHz audio for the reference workflow and the final OpenMDW 1.1 license for commercial or non-commercial use.
The output is anonymous speaker activity and timestamps, not words or verified identity. Pair it with ASR when you need speaker-attributed transcription, and keep diarization errors separate from word-recognition errors during evaluation.
02
Run with NVIDIA NeMo
NVIDIA's model card documents NeMo Speech as the reference runtime. Load the checkpoint through SortformerEncLabelModel.from_pretrained and use the diarization workflow with audio paths, arrays or supported manifests.
Returned segments identify start time, end time and an anonymous speaker index. Preserve overlapping speaker activity when the application cares about interruptions or cross-talk instead of forcing every instant to one speaker.
03
Choose and preserve a streaming profile
The documented recommended input-buffer profiles are 30.4 seconds, 1.04 seconds, 0.64 seconds and 0.32 seconds. Smaller buffers trade context for latency, so the lowest value is not automatically the best production choice.
Streaming state carries speaker information across chunks through the model's speaker cache and context. Resetting that state effectively starts a new diarization session and can break label continuity.
04
Run with Hugging Face Transformers
Hugging Face added native Transformers support using AutoProcessor and AutoModelForAudioFrameClassification, with separate offline and streaming examples. At verification time, the documentation is on the main branch and may require a sufficiently recent or source installation.
The streaming integration exposes low_latency, very_low_latency and ultra_low_latency modes corresponding to the documented 1.04, 0.64 and 0.32 second profiles. Pass the returned speaker cache into subsequent chunks and flush the final chunk correctly.
05
Production checklist and hardware caveats
Measure latency end to end, not just the audio buffer. Validate sample rate, streaming-state continuity, overlap handling, the eight-speaker limit, ASR alignment and behavior on the real meetings or calls your system will process.
The final model card does not publish a universal minimum RAM or VRAM requirement. NVIDIA's disclosed throughput tests use specific hardware and batching, so do not convert those benchmark environments into a claimed minimum device specification.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.