Nvidia Open
The Decoder highlights a quietly strategic move: Nvidia is no longer just selling the picks and shovels — it is shipping open models that run on its own hardware. The efficiency claim (a sixth of the active parameters of the closest open rival) may matter more than the benchmark lead, since transcription is fundamentally a cost-per-minute business.… more
In our view, the 800,000-hour dataset could be the bigger gift: speech training data is far scarcer than text, and bundling it with fine-tuning tools may spark a wave of specialized voice models. The prompt-driven instruction interface also points where the field is heading — one multimodal model replacing a pipeline of separate ASR, diarization and translation systems. For anyone building voice agents, this release deserves a real test.
Nvidia's open Nemotron 3.5 Transcribe combines transcription, translation, and speaker recognition
Nvidia has released Nemotron 3.5 Transcribe, an open speech recognition model that handles transcription, speaker recognition, and translation and takes its instructions as text prompts. It leads several benchmarks and uses only a sixth of the active…
Source:
The Decoder
Discussion (0)