Nvidia launches a promptable transcription model
Nvidia on Wednesday launched Nemotron 3.5 Transcribe, its latest multilingual transcription model. The new open-weight model, the company says, ranks The post Nvidia launches a promptable transcription model appeared first on The New Stack .
Nvidia announced the launch of Nemotron 3.5 Transcribe, a multilingual transcription model on Wednesday. The open-weight model boasts the best performance on Hugging Face’s Open ASR open model leaderboard for English, with a word error rate of just 5.33%. It also leads two speaker-attribution benchmarks. What sets Nemotron apart is its ability to be promptable, allowing users to provide specific instructions before the model processes the audio.
This means users can set the context for the transcription, label speakers, or ask the model questions about the audio. They can also request a verbatim transcript or a cleaned-up version and dictate how numbers should be formatted. This is all bundled into one checkpoint, making it easier to build applications around these features.
The trade-off is that the model is larger than traditional transcription models, with 30 billion mixture-of-experts and 4 billion active parameters. Despite this, Nemotron’s word error rate of 2.81% on Artificial Analysis’s benchmark is competitive with Voxtral Small’s 2.77%, but with significantly fewer active parameters. The model supports 40 languages, with translation capabilities for 31 of them.
Speaker attribution works for 13 languages and can distinguish up to eight speakers. Audio Q&A is trained in English and can handle recordings up to an hour long. Developers can serve the model using vLLM or run it locally with NeMo-Speech.cpp. Nvidia claims the model runs on a single GPU and supports hardware from Jetson AGX Thor to Blackwell data-center GPUs.
The model is open-weight (under the OpenMDW-1.1 license), enabling users to fine-tune it for better accuracy. A medical transcription case study showed a 45.8% relative reduction in word error rate when fine-tuning. Nvidia has also released Granary v2, a reusable data curation pipeline built with NeMo Curator, and public manifests covering approximately 463,062 hours across 48 languages and 16 source corpora. This release provides the dataset, not the full training mixture or its sampling weights.
Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
