NVIDIA Releases Nemotron 3.5 Transcribe, an Open Weight Speech Model
We track speech recognition releases. NVIDIA has released Nemotron 3.5 Transcribe, an open weight speech model. Open weight means anyone can download the trained model and run it on their own hardware. NVIDIA published the model on October 28, 2026 through Hugging Face, GitHub, its NGC catalog, and build.nvidia.com. A November 16 report from Tech and Space brought new attention to the model and its prompt features.
Several audio tasks in one model
NVIDIA says one checkpoint handles transcription, speaker attribution, translation, and audio question answering. According to NVIDIA, the model transcribes 40 languages. It translates 31 of those languages to or from English. It labels speakers in 13 languages and can tell apart up to eight speakers in one recording.
The model also adds word timestamps and identifies the spoken language. It accepts recordings up to one hour long. NVIDIA describes it as an offline model, which means it works on recorded files instead of live streams.
Guidance through prompts
Users can guide the output with plain language instructions. A user can ask the model to transcribe a meeting and label each speaker, to translate the audio into English, or to answer a question about the meeting.
Users can also give the model a list of likely terms. NVIDIA shows a medical example. The prompt lists terms such as HbA1c and metformin and asks the model to use them when the audio supports them. We see prompt vocabulary as a key feature for transcripts in medicine, law, and other fields with special terms.
NVIDIA also released Granary v2, an open data set and data pipeline, with tools for fine tuning. Fine tuning means further training a model on new examples. In a medical case study, NVIDIA says fine tuning lowered WER on new speakers by 45.8 percent compared with the best prompt alone. WER counts word mismatches against a reference transcript, so a lower number is better.
Size, hardware, and license
The model has 34 billion parameters in total, including its audio encoder. It uses a mixture of experts design, so about 4.3 billion parameters are active for each step. The full precision version runs on one NVIDIA H100 GPU with 80 GB of memory. NVIDIA also offers smaller FP8 and NVFP4 versions.
The model card lists the OpenMDW license and says the model is ready for commercial or noncommercial use. Teams can host the model themselves with vLLM. They can also use hosted versions from Baseten, DeepInfra, or Simplismart.
Reported results
NVIDIA reports that the model ranks first among open weight models for English on the Hugging Face Open ASR Leaderboard. Its mean WER across nine public test sets is 4.71, according to the model card. NVIDIA also reports the strongest result on a 21 language speaker labeling leaderboard from David AI. NVIDIA captured these results on October 12, 2026. The figures in this section come from NVIDIA.
NVIDIA advises users to check names, speaker changes, and answers against the recording. We will keep tracking open speech models as they reach transcription and translation workflows.