NVIDIA released an open-weight voice model called Nemotron 3.5 Transcribe—and it’s aiming to kill the “tool chain” problem in speech AI. Instead of stitching together ASR + translation + speaker diarization + post-processing, this single model can take one audio clip and follow natural-language instructions to generate different output styles. Want a verbatim transcript that keeps filler words and speaking vibe? Done. Need a cleaner, readable version with “ums” removed? Also done. Want speaker-labeled segments (up to 8 people) with word-level timestamps so you can jump to exact moments? Nemotron 3.5 Transcribe has you covered. It supports 40 languages for transcription (31 for English ↔ translation) and processes up to ~1 hour per run. Heads up: speaker separation is based on voice differences—not guaranteed real identity. Developers get “instructionized outputs,” including domain terminology handling and custom formatting. NVIDIA reports strong benchmarks (WER ~2.8% on English). It’s MoE (340B total, ~43B active per token). vLLM deployment + datasets/tools are included too. #NVIDIA #ASR #SpeechAI #OpenSource #MachineLearning #NLP
Want to learn more? Visit Explore the world, stay updated on travel insights and international affairs, and discover authentic stories from real life
评论
发表评论