# NVIDIA releases 34B speech model with manifests covering 463,062 hours > Source: > Published: 2026-10-28 15:30:40+00:00 # NVIDIA releases 34B speech model with manifests covering 463,062 hours **Nemotron 3.5 Transcribe combines transcription, translation, speaker labels and audio Q&A; Granary v2 adds public manifests and a reusable curation pipeline.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) · Published Primary source: [Hugging Face Newsroom](https://huggingface.co/blog/nvidia/nemotron-transcribe-granary) ## Why it matters NVIDIA is pairing an open-weight speech model with reusable data and fine-tuning tools, giving developers a route to adapt transcription locally while making the hardware and benchmark caveats visible. NVIDIA released an open-weight speech model that can transcribe, translate, label speakers and answer questions about recordings, alongside public manifests for 463,062 hours of multilingual speech data. The [October 28th announcement](https://huggingface.co/blog/nvidia/nemotron-transcribe-granary?ref=runtimewire) puts the model and its data-curation pipeline in developers' hands, with the promise that teams can tune speech recognition to their own terminology, accents and output formats.… When you purchase through links in our articles, [we may earn a small commission](https://runtimewire.com/affiliate-monetization-standards). This doesn’t affect our editorial independence.