Voice AI in vogue: Nvidia releases new open transcription model

In Brief

Nvidia connected Wednesday released Nemotron 3.5 Transcribe, a free, open-weight exemplary that transcribes audio successful up to 40 languages. It besides released mentation 2 of Granary, its open-source code dataset.

Nvidia boasts that Nemotron 3.5 tin construe 31 languages to and from English, grip recordings up to an hr long, and archer up to 8 speakers apart. It claims to bushed rivals connected benchmarks, including Hugging Face’s Open ASR Leaderboard.

These tools people developers gathering apps for concern meetings and aesculapian transcription, 2 of AI’s astir established footholds successful the workplace. Nvidia says developers tin usage the model, the updated Granary dataset, and tools similar its NeMo Speech exemplary to customize apps to admit an industry’s jargon successful each language. One drawback: Transcribe works with recorded audio, not unrecorded streams.

Still, the institution promises Transcribe, astatine 30 cardinal parameters, tin tally connected a azygous chip, from the Jetson AGX Thor robotics controller and DGX Spark spot powering caller laptops, to Blackwell datacenter chips. It joins Nvidia’s open-weight dependable models, including the ace tiny yet powerful Parakeet (1.1 cardinal parameters).

Nvidia’s latest dependable introduction comes during engaged play of AI dependable releases: Mistral shipped Voxtral 3 earlier this month; ElevenLabs launched Realtime for instant code successful dependable agents; and Salesforce released Launa-Omni, its dependable model.

Subscribe for the industry’s biggest tech news