Nvidia unveiled Nemotron 3.5 Transcribe on October 28, 2026, an open-weight artificial intelligence model designed for processing recorded audio. The model can transcribe speech in approximately 40 languages, translate 31 of those languages to or from English, and identify up to eight distinct speakers within a single audio file. Nemotron 3.5 Transcribe is engineered to process recordings up to one hour in length.
The model's architecture utilizes a mixture-of-experts approach, with 34 billion total parameters but only around 4 billion active per processing step. This design allows the model to run on a single graphics processing unit (GPU), ranging from Nvidia's edge devices like Jetson boards to its data center GPUs. Nvidia states that Nemotron 3.5 Transcribe leads in performance among open models on Hugging Face's English speech-recognition leaderboard, achieving an average word error rate of 5.33%. The model accepts instructions through plain-language prompts, enabling users to request specific outputs such as a clean transcript, a verbatim record, or to incorporate domain-specific terminology for improved accuracy.
Nvidia also released Granary v2, a complementary open dataset containing 463,062 hours of speech across 48 languages. This dataset, built using Nvidia's NeMo Curator tools, provides audio in various transcription styles, including verbatim with numbers spelled out, verbatim with symbols, and a cleaned version that omits filler words. The company has also provided tools for developers to fine-tune the model, demonstrating its adaptability with a case study on Indian medical speech that reportedly reduced word error rate by 45.8%.
Nemotron 3.5 Transcribe is specifically built for recorded audio and does not support live transcription. This distinction places it in a different category than real-time transcription services. Nvidia offers a separate model, Nemotron 3.5 ASR, which is designed for streaming audio. The licensing for Nemotron 3.5 Transcribe permits commercial use.