NVIDIA Corp. on Wednesday launched Nemotron 3.5 Transcribe, a free, open-weight artificial intelligence (AI) model capable of transcribing speech across 40 languages and translating 31 of them to and from English.
Alongside the release, the chipmaker made Granary v2 publicly available — a massive training corpus spanning 463,062 hours of speech in 48 languages.
The new model processes pre-recorded audio files up to an hour long and can identify and separate as many as eight individual speakers. Beyond basic transcription, the system can answer contextual questions about recordings. It responds to natural-language prompting, allowing developers to request verbatim transcripts, clean readouts, or heightened sensitivity to domain-specific jargon such as medical or technical terminology.
Designed for operational efficiency, the architecture comprises roughly 30 to 34 billion total parameters yet activates only about 4 billion per computational step. This mixture allows the model to run on a single chip, spanning compact edge devices like NVIDIA’s Jetson boards and laptop processors up to high-end Blackwell enterprise GPUs. The model is licensed for commercial deployment.
NVIDIA claimed top standing for the system on Hugging Face’s open speech-recognition benchmark, citing an average English word error rate of 5.33%. To demonstrate adaptability, the company fine-tuned the model on speech from Indian physicians, reporting a 45.8% error reduction on evaluation sets after roughly 14 hours of training on a single processor.
Unlike streaming engines, Nemotron 3.5 Transcribe is engineered exclusively for pre-recorded media rather than live broadcasts. The tooling targets developers building enterprise meeting summaries, documentation pipelines, and clinical transcription platforms — sectors that represent early, lucrative commercial deployments for generative AI.
Granary v2 significantly expands its predecessor, nearly doubling language coverage from 25 to 48 tongues. Developers can access the underlying audio from source repositories under respective licenses, presenting an alternative to fee-based commercial alternatives such as xAI’s Grok transcription, which runs at 10 cents per hour.
The release bolsters NVIDIA’s expanding footprint in open-weight software, supplementing its 1.1-billion-parameter Parakeet audio model and building upon foundational work in the Nemotron family, including April’s Nano Omni and August’s Lightning framework.
The launch lands amid fierce competition across conversational and voice technologies. Recent weeks have seen Mistral AI ship its Voxtral 3 system, ElevenLabs roll out real-time conversational agents, and Salesforce Inc. introduce Launa-Omni. By providing open-weight models that pair directly with its hardware stack, NVIDIA seeks to ensure its compute infrastructure remains the foundational layer for both enterprise fine-tuning and end-user deployment.
