Nvidia is pushing open voice AI into a new phase with a model built to handle long recordings, multiple speakers, and dozens of languages. On Wednesday, October 28, 2026, the company released Nemotron 3.5 Transcribe, a free, open-weight transcription model aimed at developers building practical speech applications.
The release arrives as voice AI fills a growing share of the technology spotlight. Mistral shipped Voxtral 3 earlier this month, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni, its voice model. Nvidia’s answer centers on giving developers a model they can run, adapt, and connect to a broader speech development toolkit.
A Transcription Model Built for Long, Multilingual Audio
Nemotron 3.5 Transcribe can transcribe audio in up to 40 languages, giving developers a single model for applications that serve users across multiple language groups. Nvidia also says the model can translate 31 languages to and from English, extending its role beyond simple speech-to-text conversion.
Long recordings are another key target. Nvidia claims Nemotron 3.5 can handle recordings up to an hour long, which fits use cases such as business meetings and medical transcription. The model can also tell up to eight speakers apart, helping applications organize conversations when several people contribute to the same recording.
That combination opens the door to richer transcripts. A meeting application could separate speakers while converting a multilingual recording into text, and a medical transcription application could process a longer session without splitting the audio into short pieces first.
There is one clear limit: Transcribe works with recorded audio, not livestreams. Developers building instant voice experiences will need a different approach, while teams working from saved meetings, consultations, or other recordings can focus on the model’s transcription and translation features.
Open Tools Give Developers More Control
Nvidia released version 2 of Granary alongside Nemotron 3.5. Granary is Nvidia’s open source speech dataset, and the updated dataset gives developers another part of the stack for building and refining speech applications.
Nvidia says developers can use the model, the updated Granary dataset, and tools such as its NeMo Speech model to customize apps for an industry’s jargon in each language. That matters because business and medical speech often depends on terms that general-purpose systems may not recognize well.
Instead of treating transcription as a fixed service, developers can shape an application around the language and vocabulary of a specific field. The same toolkit can support business meetings in one setting and medical transcription in another, with customization aimed at the words each industry uses.
Nvidia also claims Nemotron 3.5 can beat rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard. Those results place the model in direct competition with other open speech systems, while its free, open-weight release gives developers a path to inspect and adapt the technology.
Thirty Billion Parameters Across Nvidia Hardware
Nemotron 3.5 contains 30 billion parameters, yet Nvidia promises that it can run on a single chip. The company says Transcribe can run on the Jetson AGX Thor robotics controller and the DGX Spark chip powering new laptops, as well as Blackwell datacenter chips.
That range connects the model to several types of computing hardware, from a robotics controller to laptops and datacenter systems. Developers can consider the same transcription technology across different platforms instead of limiting it to one class of machine.
Nemotron 3.5 also joins Nvidia’s open-weight voice models, including Parakeet, a super small yet powerful model with 1.1 billion parameters. The contrast between Parakeet’s size and Nemotron 3.5’s 30 billion parameters shows how Nvidia is building a lineup for different speech workloads and hardware needs.
The larger model brings multilingual transcription, translation, long-recording support, and speaker separation into one release, while the broader collection gives developers more options for voice applications. With Granary version 2 and NeMo Speech included in the customization path, Nvidia is turning this launch into more than a single model announcement.
Voice AI has entered a busy period, and developers now have new choices from Nvidia, Mistral, ElevenLabs, and Salesforce. Nvidia’s open-weight strategy puts Nemotron 3.5 at the center of that movement, especially for teams that want to build specialized applications rather than rely on a closed transcription service.
Based on
- Voice AI in vogue: Nvidia releases new open transcription model — techcrunch.com
- Jensen Huang Says Open Models Can Beat The Frontier. Here’s How — forbes.com
- Block’s new voice model sounds eerily human | TechCrunch — techcrunch.com
- NVIDIA and Meta to Expand Collaboration on Muse and Muse Family of Models Amid Robust Growth | Markets Insider — markets.businessinsider.com
