Nvidia on Wednesday released Nemotron 3.5 Transcribe, a multilingual transcription model that achieved a 5.33% average word error rate for English, placing it first on Hugging Face's Open ASR open model leaderboard. The open-weight model also leads two speaker-attribution benchmarks, according to the company. What sets the release apart is its promptable design, allowing users to provide specific instructions before the model processes audio.
The model can transcribe 40 languages and translate 31 of those to or from English. Speaker attribution functions in 13 languages and can identify up to eight distinct voices. Audio question-and-answer capability was trained exclusively in English, and the system processes recordings lasting up to one hour, though it doesn't support live streaming. On Artificial Analysis's benchmark, Nemotron achieved a 2.81% word error rate while using roughly 4 billion active parameters, compared with Voxtral Small's 2.77% error rate with 24 billion active parameters. The model is a 30 billion mixture-of-experts architecture with 4 billion active parameters, considerably larger than traditional transcription models that typically run under a billion parameters in size.
Being promptable means users can set context before transcription begins by telling the model whether an audio file contains a medical discussion in English, for instance, and supplying a list of relevant medical terms. Users can also request speaker labels, ask questions about the audio content, choose between verbatim or cleaned-up transcripts, and specify number formatting preferences. Nvidia bundles these features, along with translation, into a single checkpoint, which could streamline application development. In a medical transcription case study, the company says fine-tuning dropped word error rate on held-out speakers from 22.44% with its best prompt to 12.17%, representing a 45.8% relative reduction, with the complete adaptation workflow requiring about 14 GPU-hours on a single GB300.
The promptable design addresses a core challenge in speech-to-text systems: context matters when interpreting specialized vocabulary, speaker dynamics, and formatting conventions. By allowing users to specify these parameters upfront, the model can theoretically reduce errors without requiring separate fine-tuned versions for each use case. The trade-off is computational: these models demand substantially more resources than conventional transcription systems. Developers can serve the model with vLLM or run it locally with NeMo-Speech.cpp, and Nvidia says it operates on a single GPU across hardware ranging from Jetson AGX Thor and DGX Spark to Blackwell data-center GPUs. Since this is an open-weight model under the OpenMDW-1.1 license, users can fine-tune it using NeMo Speech's tools or other options to further reduce error rates for specific applications.
Nvidia is also releasing Granary v2, a reusable data curation pipeline built with NeMo Curator, alongside public manifests covering approximately 463,062 hours across 48 languages and 16 source corpora. Developers can run the pipeline on their own recordings to prepare data for customization. The company positions the release as part of a broader shift toward promptable speech-to-text models, a concept also explored by Mistral's Voxtral models and AssemblyAI's Universal 3 Pro, though Nvidia achieves comparable accuracy with fewer active parameters. The question for enterprises will be whether the flexibility of context-setting justifies the infrastructure demands, particularly in sectors where specialized terminology and speaker identification drive accuracy requirements. Organizations investing in voice infrastructure may find themselves choosing between lightweight simplicity and adaptable power, with neither path offering a clear universal advantage.