India, Oct. 30 -- NVIDIA is addressing those requirements with Nemotron 3.5 Transcribe, a speech model that lets users select transcription styles through natural-language instructions, without retraining it for each request.

Announced on October 28, the model supports 40 languages and recordings of up to one hour. It combines transcription, speech translation, spoken-language identification, and speaker attribution, with word-level timestamps available on request. NVIDIA has released model weights and a fine-tuning toolkit.

Conventional automatic speech recognition (ASR) systems typically learn a particular transcription convention during training. That convention may determine whether numbers appear as spoken words or numerals, whethe...