# Nemotron 3.5 Transcribe — NVIDIA's open speech model you steer with prompts > One open checkpoint for transcripts, translation, speaker labels and audio Q&A, controlled by plain-language prompts. - Org: NVIDIA - Date: 2026-10-28 - Category: model, dataset - Source: https://huggingface.co/blog/nvidia/nemotron-transcribe-granary - Canonical page: https://ai-tldr.dev/releases/nvidia-nemotron-3-5-transcribe/ Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model. It transcribes 40 languages, translates 31, and labels up to eight speakers in hour-long recordings. NVIDIA also released Granary v2, a 463,062-hour speech dataset. ## What it is Nemotron 3.5 Transcribe lets you ask for the transcript you want in plain words: label each speaker, translate into English, drop the filler words, or watch for a list of domain terms. It works on recorded audio, not live streams. NVIDIA released it alongside Granary v2, an open speech dataset built with NeMo Curator. ## How it works A Transformer speech encoder and a speaker encoder feed a hybrid Mamba-2–Transformer Mixture-of-Experts language model built on Nemotron 3.5 Lightning, so only about 4.3B of its 34B parameters run per token. Granary v2 gives each recording three transcript styles, from verbatim to clean, and those are the styles the prompt can select. The model is served through vLLM with an OpenAI-compatible chat endpoint. ## Why it matters Teams that send meetings, calls or clinical notes to paid speech APIs get an open alternative that runs on a single GPU and can be tuned to their own words. In NVIDIA's Indian medical speech case study, a better prompt cut errors from 30.8% to 23.8% and a 14 GPU-hour fine-tune brought them to 11.8%. ## Who it's for developers building transcription, meeting and medical speech apps ## Try it hf download nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B ## Quick facts - Maker: NVIDIA - Parameters: 34B total, about 4.3B active - License: OpenMDW-1.1 - Languages: 40 to transcribe, 31 to translate - Input limit: Recordings up to 1 hour, up to 8 speakers - Availability: Open weights on Hugging Face and ModelScope - Companion dataset: Granary v2 — 463,062 hours, 48 languages ## FAQ ### Can Nemotron 3.5 Transcribe handle live audio? No. NVIDIA built Nemotron 3.5 Transcribe for recorded audio only. It accepts recordings up to one hour long with up to eight labeled speakers, as 16 kHz single-channel audio in WAV, FLAC, MP3 or OGG, and returns text with optional word timestamps. Live captioning needs a streaming model instead. ### How does Nemotron 3.5 Transcribe rank against other open speech models? NVIDIA reports that Nemotron 3.5 Transcribe is first among open-weight models on the Hugging Face Open ASR English leaderboard with a 5.33% average word error rate. On Artificial Analysis it scores 2.8% AA-WER, tied for first among open-weight models, and it posts the lowest diarization error of 14 systems on David AI's 21-language test. ### What GPU do I need to run Nemotron 3.5 Transcribe? The Nemotron 3.5 Transcribe model card lists a single H100 80 GB as the reference setup, and NVIDIA says it runs on one GPU from Jetson AGX Thor and DGX Spark up to Blackwell data-center cards. FP8 and NVFP4 checkpoints are published alongside the BF16 weights for smaller memory budgets. ### What is in Granary v2? Granary v2 is NVIDIA's open speech dataset released with Nemotron 3.5 Transcribe. It covers 463,062 hours across 48 languages from 16 source corpora, up from 25 languages in Granary v1. Each recording gets three transcript styles; the source audio must be downloaded separately under each corpus's own license. ## Links - [NVIDIA Launch Post](https://huggingface.co/blog/nvidia/nemotron-transcribe-granary) - [Model Card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B) - [Live Demo](https://huggingface.co/spaces/nvidia/nemotron-transcribe) - [NeMo Speech Repo](https://github.com/NVIDIA-NeMo/Speech) - [TechCrunch Coverage](https://techcrunch.com/2026/10/28/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model/)