Recorded speech is the target
Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight speech model, TechCrunch reports. The report says the model transcribes audio in up to 40 languages and accompanies a second version of Nvidia’s open-source Granary speech dataset.
Nvidia says the model can translate 31 languages to and from English, handle recordings up to an hour long and distinguish up to eight speakers. These are different capabilities with different limits. A language supported for transcription should not automatically be assumed to have the same translation coverage.
A useful boundary for application builders
Transcribe works with recorded audio rather than live streams. That makes an uploaded meeting or archived conversation a more relevant starting point than an interactive voice agent that needs immediate responses.
The report identifies meetings and medical transcription as intended application areas. Those are target uses, not evidence of validated accuracy for a particular organization’s speakers, recording conditions or specialist vocabulary. Developers would still need to assess errors on representative audio before relying on a transcript downstream.
Check the complete workload
TechCrunch reports Nvidia’s claim that the 30-billion-parameter model can run on a single chip across several hardware classes. That does not establish one throughput figure or memory configuration for all of those systems. The same caution applies to the company’s benchmark comparisons: they are reported vendor claims, not Franklin’s hands-on results.
A useful pilot would keep the original audio available, test speaker attribution separately from word accuracy, and examine terminology that matters to the workflow. The open weights and accompanying speech resources make customization a concrete possibility for developers. The recorded-audio requirement remains the first filter when deciding whether the model fits a project.