Nvidia's Nemotron 3 labels 8 speakers in real time, open weights
The free 100M model tops Voice Arena and streams down to 0.32s latency; pair it with ASR for speaker-tagged transcripts. Plus a Gemini CLI security patch.
Copy markdown
Open-weight ‘who spoke when’
Nvidia's Nemotron 3 Diarization is a 100M-param model that tags up to 8 speakers, overlapping speech included, and tops Voice Arena's Diarization-Bench at a 14.72% error rate versus 19.3% for the runner-up. The weights are on Hugging Face under the commercial-friendly OpenMDW-1.1 license.
Real time, down to 0.32s
Four buffer settings run from 30.4s offline to 0.32s ultra-low latency, so you can label speakers live in a call or on-device. Offline mode hits roughly 15,000x real-time, and at 1.04s latency it cuts error ~41% versus Nvidia's older Sortformer, with the biggest gains at 5+ speakers.
The weekend build: speaker-tagged transcripts
Load it in NeMo with SortformerEncLabelModel.from_pretrained and a .wav returns segments like `0.4 2.1 speaker_0`. Pair it with an ASR model (Nvidia demos Parakeet TDT) to turn recordings into ‘speaker_0: … speaker_1: …’ transcripts for meeting notes, podcast tooling, or call analytics.
Grab the right repo — the preview is eval-only
There are two artifacts. The OpenMDW-1.1 release is cleared for commercial use, but the gated preview ships under an evaluation-only license that bars production, redistribution, and even publishing your results. Check the license on the exact weights you pull before you build on them.
Elsewhere: Gemini CLI patches a prompt-injection hole
Gemini CLI 0.61.0 now blocks indirect prompt injection through build-file edits and untrusted flags, and hardens its sandbox's filesystem boundaries; it will ask before touching your build files. Worth the bump if you run it against untrusted repos.
Elsewhere: see how your model tokenizes
A browser-based compiler builds a TTF font where every token in a chosen tokenizer (Claude, OpenAI, DeepSeek) renders at equal width, making token boundaries visible in any app you install it in. It runs client-side via Pyodide and fontTools, with drop-in guides for Slack and Discord.