Nvidia launches Nemotron 3.5 Transcribe multilingual transcription model and Granary v2 speech dataset
nvidia open-source speech
| Source: Techmeme | Original article
Nvidia has launched Nemotron 3.5 Transcribe, an open‑weight model for multilingual audio transcription and translation, together with Granary v2, its new open‑source speech dataset.
Nvidia has unveiled Nemotron 3.5 Transcribe, an open‑weight speech‑augmented language model that can transcribe, translate, label speakers and answer audio‑based questions from a single checkpoint. The model, now hosted on Hugging Face and ModelScope, joins Nvidia’s growing portfolio of open AI tools and is accompanied by Granary v2, an open‑source speech dataset exceeding 800,000 hours of multilingual audio.
Nemotron 3.5 Transcribe is built as a speech‑augmented language model (SALM) that pairs a speech encoder with a large language model. At inference it activates roughly four billion of its 30 billion parameters, allowing the system to run on a single GPU—from Nvidia’s entry‑level DGX Spark up to its latest Blackwell data‑center chips. The model can be steered with natural‑language prompts, letting users request specific output formats such as translation, speaker attribution or medically accurate terminology without retraining.
The release matters because it narrows the gap between proprietary, high‑cost speech services and accessible, community‑driven alternatives. By delivering multilingual transcription and translation in a single, promptable model, Nvidia addresses the fragmented tooling landscape that developers currently face when building applications for call‑center analytics, real‑time meetings or clinical dictation. Granary v2 further lowers the barrier to entry, providing a massive, curated corpus for fine‑tuning and research.
Going forward, observers will watch how the model performs on established benchmarks and whether it can sustain its early leaderboard lead as competitors release newer open‑weight alternatives. Nvidia’s rollout also raises questions about the ecosystem of tools for prompt engineering, fine‑tuning pipelines and integration with existing enterprise workflows. The next update from Nvidia is likely to detail performance metrics across languages and domains, and to announce any partnerships that bring Nemotron 3.5 Transcribe into production‑grade services.
Sources
Back to AIPULSEN