AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Voice AI in vogue: Nvidia releases new open transcription model

Collected Oct 28, 2026

Nvidia released Nemotron 3.5 Transcribe on Wednesday, a free, open-weight speech-to-text model that handles audio in up to 40 languages. Alongside it, the company published version 2 of Granary, its open source speech dataset. The release targets developers building transcription features into applications.

The model's stated capabilities cover three distinct jobs. It can translate 31 languages to and from English, accept recordings up to an hour long, and separate up to eight speakers in a single piece of audio. Speaker separation of that order is a practical requirement for meeting tools, where the value of a transcript often depends on knowing who said what rather than just what was said.

Nvidia claims the model beats competitors on benchmarks, including Hugging Face's Open ASR Leaderboard. That claim is the company's own and has yet to be independently verified across the languages and audio conditions real applications encounter.

A significant constraint is that Transcribe works with recorded audio, not livestreams. That narrows the addressable use case considerably: batch or post-hoc transcription of meetings, interviews, clinical visits and similar sessions, rather than live captioning or voice agents that respond in the moment. Nvidia positions the model for business meeting and medical transcription, two of the more established workplace applications of AI. The gap leaves real-time voice work to competitors.

On deployment, Nvidia says the 30-billion-parameter model can run on a single chip. That spans its Jetson AGX Thor robotics controller and the DGX Spark chip powering new laptops, up to Blackwell datacenter chips. A model of this size running on one accelerator rather than a multi-GPU cluster is notable, because self-hosting speech recognition on modest hardware changes the economics for teams that cannot send sensitive audio to a third-party API. This suggests the medical transcription use case is a natural fit, since patient audio is exactly the kind of data organizations prefer to keep in-house. That is my reading, not something Nvidia stated.

Transcribe joins Nvidia's existing open-weight voice lineup, which includes Parakeet at 1.1 billion parameters. The size difference is stark: Parakeet is roughly twenty-seven times smaller than Transcribe. In practice, that spread implies a trade-off between footprint and capability, with the smaller model suited to latency- and resource-constrained settings and the larger one aimed at harder accuracy problems such as multilingual audio, hour-long recordings and multi-speaker separation. Nvidia also points developers to its NeMo Speech tooling, which together with the Granary dataset is intended to let teams fine-tune recognition of industry-specific vocabulary in each language. Domain jargon, whether clinical terminology or internal business shorthand, is a common weak point for general-purpose speech models, so a customization path matters as much as raw benchmark scores.

The release lands in a crowded stretch for voice AI. Mistral shipped Voxtral 3 earlier this month, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released its Launa-Omni voice model. The pattern is that voice has moved from a standalone product category to a component that model vendors expect developers to build on. Nvidia's differentiator here is the combination of open weights, broad language coverage and hardware it already sells.

Why it matters: Developers building transcription into meeting, clinical or back-office products get a free, self-hostable option with wide language coverage and multi-speaker support, which lowers the cost and data-handling friction of shipping such features. Teams working on live audio, voice agents or real-time captioning still need a different stack, since the model does not process streams. The open weights and dataset also give teams a base for domain-specific fine-tuning, an inference on my part, while the benchmark claims remain something to test against your own audio before committing.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Nvidia just released a free speech-to-text model that can transcribe up to 40 languages, but it only works on recordings, not live audio, so it’s a big step for meeting and medical transcription tools while still leaving real-time use to rivals like ElevenLabs.