NVIDIA has announced Nemotron 3.5 Transcribe, a new open-weights speech language model featuring approximately four billion active parameters within a thirty billion mixture-of-experts architecture. According to reporting by MarkTechPost, the model is built to handle transcription across forty languages, while also providing speaker labels and word-level timestamps. It achieved a four point seventy-one percent mean word error rate on the English Open ASR Leaderboard.
Open Models in Speech Processing
The release marks another step by hardware and software firms into open-weights audio models. Mixture-of-experts architectures allow models to maintain large total parameter counts while keeping active compute costs lower during inference, a design choice increasingly common in both text and speech systems.
Unknowns and Future Performance
While the initial benchmark score on the English Open ASR Leaderboard provides a baseline, the announcement did not specify hardware requirements for running the thirty billion parameter model locally, nor did it detail licensing restrictions or API availability for enterprise deployment. Further testing will determine how the model performs across its full span of forty supported languages outside of controlled benchmarks.
Developers and researchers can expect more implementation details as community evaluations of the open-weights release begin.
Lead image: panumas nikhomkhai / Pexels
