NVIDIA open-weights Nemotron 3.5 Transcribe: 30B-param speech model in 40 languages
dl_weekly · x · 2026-11-08
NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech transcription model, alongside v2 of its open-source Granary speech dataset.
- 30 billion parameters, 40 transcription languages, 31 translation pairs to/from English
- Handles hour-long recordings and distinguishes up to 8 speakers
- Claims benchmark wins including Hugging Face's Open ASR Leaderboard
- Runs on a single chip, from Jetson AGX Thor robotics controllers to DGX
- Targets meeting and medical transcription; developers can customize with NeMo Speech tools to recognize industry jargon per language
Caveat: works on recorded audio only, not livestreams.
More from Infra
- Chamath: Compute Spot Prices Doubling Would Force SpaceX to Reprice Anthropic Deal via 90-Day Exit — rohanpaul_ai · 2026-11-08
- Study: data centers' impact on residential electricity prices is only marginal — AccBalanced · 2026-11-08
- RustingStrata hits 160 tok/s on RTX 3060 running 35B MoE, ~2x faster than llama.cpp — Serious_Scarcity7818 · 2026-11-08
- Geek's Throne wishlist gets a $5M exabox (256x MI300X) fully funded — gnukeith · 2026-11-08
- Dynatrace Closes $915M Arize Deal as 1,100+ Customers Run Its Agents — shashib · 2026-11-08
- Running Multiple LLM Slots on One 7900 XTX Boosts Total Tokens/Sec — c2btw · 2026-11-08