AitopkitNvidia on Wednesday, October 28, 2026, released Nemotron 3.5 Transcribe, a free, open-weight speech model that transcribes audio in up to 40 languages. Alongside it, the company released version 2 of Granary, its open-source speech dataset.
The headline specs are aimed squarely at developers building meeting and medical transcription apps — two of AI's most established workplace footholds. According to Nvidia, the model can translate 31 languages to and from English, handle recordings up to an hour long, and tell up to eight speakers apart in a single recording. It also claims to beat rivals on benchmarks, including Hugging Face's Open ASR Leaderboard — though those are Nvidia's own reported results, and independent tests on messy real-world audio will be the real verdict.
Coverage is organised into tiers. Nineteen locales, including English (US and GB), Spanish, French, German, Arabic, Hindi, Japanese and Korean, are rated transcription-ready out of the box; a further 13 locales get production-grade coverage, and 8 more are adaptation-ready, meaning they need a fine-tuning pass on in-domain data to unlock full accuracy. That breadth lets one model replace a fleet of single-language deployments — a big deal for global meeting tools and call-centre software.
At 30 billion parameters, Nvidia says Transcribe can run on a single chip, from the Jetson AGX Thor robotics controller and the DGX Spark chip powering new laptops, all the way to Blackwell datacenter chips. It joins Nvidia's growing open-weight voice lineup, including the much smaller Parakeet model (1.1 billion parameters), and lands in a busy month for AI voice: Mistral shipped Voxtral 3 earlier in October, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni.
One limitation to note: Transcribe works with recorded audio, not livestreams, so real-time captioning use cases will need a different tool for now. Developers can combine the model with the updated Granary dataset and tools like Nvidia's NeMo Speech model to customise apps for an industry's jargon in each language — the kind of tuning that decides whether a transcription tool is actually usable in a hospital or a boardroom.