NVIDIA Launches Open-Weight Nemotron 3.5 Transcribe for 40 Languages

29 Oct 2026, 4:07 pm · 2h ago · 1 min read · Analytics India MagazineNew

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech recognition model supporting 40 languages, translation, and speaker diarisation. Built with approximately 4.3 billion active parameters across a 30B-A4B architecture, the model combines a Transformer speech encoder with NVIDIA’s Nemotron 3.5 Lightning language model. It enables developers to format transcripts using natural language instructions without fine-tuning. On the English and multilingual Open ASR Leaderboards, NVIDIA reported record mean word error rates of 4.71% and 4.62% respectively. It also delivers high-throughput inference of up to 2,238 seconds of audio per second on RTX PRO 6000 GPUs.