Nvidia Corp. on October 28 released Nemotron 3.5 Transcribe, an open-weight speech model that consolidates multilingual transcription, translation, speaker attribution, and audio question answering into a single checkpoint1,4,5. The model transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. Alongside it, Nvidia published Granary v2, an open speech dataset and curation pipeline whose public manifest covers 463,062 hours across 48 languages.
Architecture and hardware envelope
Nemotron 3.5 Transcribe is a 30-billion-parameter mixture-of-experts model with roughly 4 billion active parameters per token, fitting inside a single GPU3. The speech side pairs a 600M-parameter Transformer ASR encoder with the approximately 99M-parameter Nemotron-3-Diarization speaker encoder and a connectionist temporal classification head for word timestamps; the language backbone is Nvidia's Nemotron-3.5 Lightning 30B-A3B, a hybrid Mamba-2–Transformer MoE model. The full checkpoint is 34 billion parameters with about 4.3 billion active per token and a 65,536-token context window.
Nvidia said a Blackwell Server Edition GPU running NVFP4 precision with one-quarter model compression processes 100 hours of audio in under three minutes. The model also runs on smaller hardware, from a Jetson AGX Thor edge computer to a DGX Spark desktop AI system. Weights ship under the OpenMDW-1.1 license for commercial or non-commercial use, with hosted endpoints offered by Baseten, DeepInfra, and Simplismart.
Promptability and benchmarks
Unlike conventional ASR pipelines that chain separate detection, transcription, diarization, and QA stages, Nemotron 3.5 Transcribe handles all tasks in a single model pass. The model is promptable: users can set context before inference, specifying domain terminology, requesting verbatim or cleaned-up output, choosing number formatting, or asking questions about the audio. It handles recordings up to one hour long and labels up to eight speakers.
On Hugging Face's Open ASR leaderboard for English, the model ranks first with an average word error rate of 5.33% and also leads two speaker-attribution benchmarks, according to Nvidia. The model does not support live streaming; it works only on recorded audio2. That constraint leaves real-time transcription to competitors such as ElevenLabs.
Fine-tuning pipeline and dataset
Granary v2, built with NVIDIA NeMo Curator, expands from 25 languages in Granary v1 to 48 languages across 16 source corpora. In a medical transcription case study using the Eka Care Indian medical speech benchmark, Nvidia reported that fine-tuning cut the word error rate on held-out speakers from 22.44% with its best prompt to 12.17%, a 45.8% relative reduction. The complete adaptation workflow took about 14 GPU-hours on a single GB300, according to Nvidia. Developers can serve the model with vLLM or run it locally with NeMo-Speech.cpp.
ANALYSIS Bundling transcription, translation, diarization, and audio QA into one open-weight checkpoint lowers the integration cost for teams that previously stitched together separate models for each task. The no-live-streaming limitation, however, confines the model's immediate utility to post-hoc workflows such as meeting recap, medical documentation, and media indexing.