Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognition
decoder ·
Nvidia has released Nemotron 3.5 Transcribe, an open speech model that integrates transcription, translation, and speaker recognition into a single system. The model supports up to 40 languages for transcription and handles audio files of up to one hour with multiple speakers. Although it contains 30 billion parameters, only 4 billion are active per step, allowing it to run efficiently on a single GPU. It achieves top rankings on several benchmarks, including Hugging Face's English Open ASR Leaderboard and VoiceArena's speaker attribution tests. Alongside the model, Nvidia introduced Granary v2, a dataset comprising 844,000 hours of audio across 48 languages. The company also provided tools for fine-tuning the model in specialized domains, demonstrating significant accuracy improvements in medical contexts. While the base model offers robust performance, users are advised to spot-check outputs and note that real-time processing is not supported.
decoder