Voice AI in vogue: Nvidia releases new open transcription model - TechCrunch

Nvidia has released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes audio in up to 40 languages, TechCrunch reported. The company also published version 2 of Granary, described in the article as its open-source speech dataset. The article characterises the model as both free and open-weight but does not name a licence, state a release date, or say where the weights can be downloaded.

According to the article, the model can translate 31 languages to and from English, handle recordings up to an hour long, and tell up to eight speakers apart. The 40-language transcription figure and the 31-language translation figure are given separately, and the speaker-separation capability is stated without detail on accuracy. TechCrunch does not report benchmark results, parameter counts, comparisons with other speech models, or what hardware the model runs on.

The article's headline places the release under the heading of voice AI being 'in vogue'. Beyond transcription and translation, the excerpt does not describe intended applications for the model or for the Granary dataset. Nor does it state how either will be distributed, whether commercial use is permitted, or how the contents of the dataset's second version differ from the first, which the article references only by version number.