Voice AI in vogue: Nvidia releases new open transcription model

Introduction

Nvidia has just unveiled Nemotron 3.5 Transcribe, a free, open‑weight speech‑to‑text model that can handle up to 40 languages, differentiate up to eight speakers, and process recordings as long as an hour. Coupled with the updated Granary v2 speech dataset and the NeMo Speech toolkit, the release is positioned as a developer‑friendly foundation for business‑meeting and medical‑transcription applications. While it does not support live streaming, the model’s modest hardware requirements—30 billion parameters that can run on a single Nvidia chip—make it attractive for a wide range of edge and datacenter deployments.

A versatile, multilingual engine

Nemotron 3.5 Transcribe is built to be multilingual out of the box. In addition to transcribing 40 languages, it can translate 31 of those to and from English, enabling cross‑language workflows such as:

• International conference call summaries that automatically produce bilingual minutes.
• Multilingual medical notes where a clinician records in a native language and the system outputs an English transcript for centralized records.

The model also includes speaker diarization, distinguishing up to eight distinct voices within a single recording. This feature is especially useful for board meetings, legal depositions, and multi‑doctor consultations where attribution of remarks matters.

Open data and customization tools

Nvidia’s release is more than a single model; it comes with Granary v2, an expanded open‑source speech dataset that covers diverse accents, domains, and acoustic conditions. Developers can fine‑tune Nemotron 3.5 using the NeMo Speech framework, tailoring the system to industry‑specific jargon—think pharmaceutical terminology, financial acronyms, or technical engineering lingo.

Because the entire stack is open‑weight, teams can:

1. Download the model and dataset without licensing barriers.
2. Fine‑tune on proprietary corpora to improve accuracy for niche vocabularies.
3. Deploy on a range of Nvidia hardware, from the Jetson AGX‑Thor edge controller to the DGX Spark and Blackwell datacenter GPUs.

Positioning in a crowded voice‑AI market

Nemotron 3.5 Transcribe arrives amid a flurry of new voice models:

• Mistral’s Voxtral 3, released earlier this month.
• ElevenLabs’ Realtime engine for instant speech synthesis.
• Salesforce’s Launa‑Omni, a multimodal voice model.

What sets Nvidia apart is the combination of open weight, multilingual breadth, and hardware flexibility. While competitors often focus on either streaming capabilities or ultra‑large parameter counts, Nvidia emphasizes a model that can be run on a single chip without sacrificing language coverage.

Practical considerations for adopters

• Recorded‑audio only: Transcribe does not yet support live streaming, so use cases requiring real‑time captions will need a separate solution.
• Hardware sizing: Although the model can run on a Jetson AGX‑Thor, performance scales with GPU power. For high‑throughput batch transcription, a Blackwell‑class datacenter GPU will deliver faster turnaround.
• Data privacy: Since the model and dataset are open, organizations can host everything on‑premises, keeping sensitive audio—such as patient consultations—within their own security perimeter.

Closing thoughts

Nvidia’s Nemotron 3.5 Transcribe represents a pragmatic step forward for voice AI: an openly available, multilingual transcription engine that balances accuracy, speaker awareness, and deployment flexibility. By pairing the model with the Granary v2 dataset and NeMo’s fine‑tuning tools, developers can quickly build domain‑specific transcription services without the licensing hurdles that often accompany commercial speech APIs. While live‑streaming remains out of scope, the model’s ability to run on a single Nvidia chip makes it a compelling option for both edge devices and large‑scale cloud workloads, potentially accelerating the adoption of voice AI across enterprise and healthcare settings.