Nvidia releases free Nemotron 3.5 Transcribe for voice transcription

Nvidia announced on Wednesday that it is releasing Nemotron 3.5 Transcribe, a free, open‑weight voice model that can transcribe audio in up to 40 languages. The company also updated its Granary speech dataset to version 2, providing more training material for developers. According to TechCrunch, the model can distinguish as many as eight speakers and translate 31 languages to and from English, positioning it for use in business meetings and medical transcription.

Nvidia's new Nemotron 3.5 Transcribe model

The key specifications shared by Nvidia include a 30‑billion‑parameter architecture that can run on a single chip, from the Jetson AGX Thor controller to the Blackwell datacenter processor. The company says the model “beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard.” In addition to the transcription engine, Nvidia released version 2 of its Granary open‑source speech dataset, which developers can use to fine‑tune the model for industry‑specific terminology. The model is designed for recorded audio only; it does not support live‑stream transcription.

Why the release matters for small businesses, nonprofits, and public‑sector teams

Many organizations already rely on voice AI for meeting notes, customer support logs, and medical record keeping. Until now, high‑quality multilingual transcription has often required paid services or large cloud contracts. By making Nemotron 3.5 Transcribe free and open‑weight, Nvidia lowers the barrier to entry for small teams that need to process audio in multiple languages. The ability to separate up to eight speakers helps organizations keep track of who said what, a feature useful for board meetings, community forums, or multidisciplinary medical consultations.

The model’s claim of running on a single chip also means it can be deployed on edge hardware, reducing dependence on internet bandwidth and external APIs. For nonprofits working in remote regions, this can translate into lower operating costs and improved data privacy, as audio never leaves the organization’s own device.

Getting started: practical steps

If you are interested in testing or deploying Nemotron 3.5 Transcribe, follow these steps:

• Review the licensing terms on Nvidia’s official repository to confirm the model fits your project’s open‑source or commercial use case.
• Download the model and Granary v2 dataset from Nvidia’s GitHub page.
• Set up an Nvidia‑compatible development environment, a Jetson device, a DGX Spark laptop, or a Blackwell‑based server will run the model efficiently.
• Fine‑tune with domain‑specific audio using the NeMo Speech toolkit, which Nvidia bundles with example scripts.
• Integrate the transcribed text into your workflow, for example, pipe the output into a note‑taking app, a document management system, or a secure electronic health‑record platform.
• Test speaker separation on a sample recording that includes multiple participants to ensure the model correctly tags each voice.
• Monitor performance for recordings longer than an hour, as the model’s official limit is one hour per file.

By following this checklist, teams can quickly prototype a transcription service without incurring additional licensing fees.

Limitations and considerations

While the release is promising, there are practical constraints to keep in mind. The model only processes recorded audio; live captioning or real‑time translation is not supported. Nvidia’s benchmark claims have not been independently verified for every language pair, so organizations should conduct their own accuracy tests, especially for critical medical or legal use cases. The 30‑billion‑parameter size, although runnable on a single chip, still requires a capable GPU; older or low‑power hardware may struggle with latency.

Security‑focused teams should also review the data handling practices of any downstream services that consume the transcribed text. Even though the model can run on‑premises, any cloud‑based storage or analytics component must meet your organization’s compliance requirements.

In the broader AI voice market, Nvidia’s release follows other recent launches such as Mistral’s Voxtral 3, ElevenLabs’ Realtime, and Salesforce’s Launa‑Omni. The timing suggests a competitive push toward more accessible voice tools, making it worthwhile for small organizations to evaluate several options before committing to a single vendor.

Overall, Nemotron 3.5 Transcribe offers a rare combination of multilingual coverage, speaker differentiation, and on‑device execution at no cost. For teams that need reliable transcription without ongoing subscription fees, it represents a practical option worth exploring.