Nvidia’s new transcription release works at a very different scale from the Liquid models. Nemotron Transcribe supports transcription in 40 languages, translation to and from English for 31, and speaker attribution in 13. The supported task and language combinations vary. It can take recordings up to an hour, with a maximum of eight speakers for attribution.
And it’s an offline system, so the natural thing to assess is a completed recording. The reference BF16 server uses one H100 with 80 gigabytes. Including the encoder, the model has 34 billion total parameters and 4.3 billion active. The smaller active count does not make that reference deployment a lightweight local transcription setup.
The audio question-and-answer training is English. That adds another language boundary to the transcription, translation and speaker-attribution counts. For a multilingual recording, the supported combination of those tasks matters more than the broadest language total in the announcement.
Quantized versions and NeMo-Speech.cpp are described in the announcement as forthcoming. That leaves the published reference setup as the concrete starting point, rather than a promised smaller version. And the vendor’s short-clip throughput figures don’t establish the time it would take to process a full hour-long recording. That longer job is what I’d want to measure.
The release is under OpenMDW 1.1, with third-party notices still applicable.