Nemotron 3 Diarization demo with NeMo and Parakeet

This walkthrough pairs NVIDIA Nemotron 3 Diarization with speech recognition to produce transcripts that label who spoke when. The speaker describes an open-weights model with a license allowing commercial use, roughly 100 million parameters excluding ASR, and a GPU memory requirement of about 4 GB for the diarization model alone, not the complete ASR application. He says it supports up to eight speakers, overlapping speech, and language-agnostic diarization.

The demo uses a Next.js interface on a Mac, connected over Tailscale to a DGX Spark. A Docker container runs NVIDIA NeMo, with FastAPI exposing the diarization and ASR models. Parakeet handles the demonstrated transcription; a separate multilingual ASR option is also available. The interface lets users replace numbered speaker labels with names and inspect talk time, word counts, and speaking turns. The model does not automatically identify people by name.

The speaker reports processing an All-In podcast episode longer than an hour in 148 seconds on this setup, then demonstrates text and SRT export. That timing describes his hardware and workflow, rather than a general performance guarantee. He also explains diarization error rate and compares the model with Sortformer. For recorded audio, offline processing can use the whole recording; streaming makes decisions as audio arrives. The model supports streaming, but the demonstrated app does not implement it.

The speaker discloses NVIDIA sponsorship for this video. The numbered labels are renamed by the user; the model does not recognize personal identities.