Nvidia's free 40-language transcription model: what it means for India
Nvidia has released Nemotron 3.5 Transcribe, a free 30-billion-parameter speech model for recorded audio. Can it serve India's many languages and strict data rules?
Key points
- Nvidia has released Nemotron 3.5 Transcribe, a free open-weight speech-to-text model hosted on Hugging Face.
- The 30-billion-parameter model transcribes up to 40 languages and handles recordings of up to one hour, but not live audio.
- Indian firms could self-host and fine-tune it for regional vocabulary, though Indic language coverage is not yet confirmed.
The News
Nvidia has put out a free, open-weight speech-recognition model called Nemotron 3.5 Transcribe, which the company says can turn recorded audio in up to 40 languages into text. The release was reported on 28 October 2026, with the weights hosted on Hugging Face.
The model has 30 billion parameters and, by Nvidia's account, can also translate 31 languages to and from English, process recordings of up to one hour and tell apart as many as eight speakers. It is built to run on a single chip, ranging from the Jetson AGX Thor robotics controller and the DGX Spark through to Blackwell data-centre processors.
There is an important catch. The system handles only pre-recorded files, not live streams, so it is aimed at jobs such as minuting business meetings and producing medical notes rather than powering real-time voice agents. Alongside it, Nvidia published version 2 of Granary, its open speech dataset, which companies can combine with NeMo Speech tools to tune the model for sector-specific vocabulary.
Why It Matters
Nvidia claims the model beats rivals on benchmarks including Hugging Face's Open ASR Leaderboard, though no precise accuracy figures were shared. Even so, a 30-billion-parameter open model from the dominant chip supplier raises the floor for anyone selling transcription as a paid service. The closest precedent is OpenAI's open-sourcing of Whisper in September 2022, which became the default engine inside many note-taking products and squeezed standalone speech vendors.
The voice segment is getting crowded. Mistral shipped Voxtral 3 earlier this month, ElevenLabs is pushing its Realtime product for instant speech in agents, and Salesforce has its own voice model. By leaving live audio to competitors while giving away a strong batch model, Nvidia is effectively commoditising one layer of the stack - and nudging developers towards running workloads on its hardware, from edge devices to the data centre.
Indian Angle
India is one of the hardest and most valuable markets for speech technology, with 22 scheduled languages, heavy code-mixing and a large share of users who prefer voice to typing. Nvidia's coverage has not yet specified which 40 languages are supported, so Indian developers should test carefully before assuming strong Hindi, Tamil or Bengali performance. Domestic players such as Sarvam AI, the AI4Bharat research group and MeitY's Bhashini language mission have spent years building Indic speech data, a moat if global models treat Indian languages as an afterthought.
The bigger opportunity may be the open Granary dataset and NeMo customisation route. An Indian health-tech firm or a bank's contact-centre team could fine-tune the model on regional terminology and run it on its own servers. That matters under the Digital Personal Data Protection Act, 2023, because recordings of doctor consultations or customer calls contain sensitive personal data that many organisations would rather not send to a foreign cloud API.
Cost is the third lever. Single-chip deployment and free weights mean Indian start-ups can avoid per-minute API charges billed in dollars, a real saving when the rupee is under pressure. The trade-off is the batch-only design: live vernacular voice bots for lending or insurance still need a real-time engine.
FAQ
What exactly has Nvidia released?
Nemotron 3.5 Transcribe, a 30-billion-parameter open-weight speech-to-text model, plus version 2 of the Granary speech dataset. The model transcribes up to 40 languages, translates 31 languages to and from English, handles recordings of up to an hour and can separate up to eight speakers. It is available on Hugging Face at no charge.
Can it be used for live calls or voice assistants?
No. The model works only on recorded audio, not live streams. It suits meeting minutes, medical documentation and archive processing, but real-time voice agents will need a different tool, such as ElevenLabs' live-speech products.
Does it support Indian languages?
The initial reporting does not list the 40 supported languages, so Indian-language coverage and accuracy are unconfirmed. Developers should benchmark it against Indic-focused systems from Sarvam AI or AI4Bharat on their own audio, and consider fine-tuning with the Granary dataset and NeMo Speech tools for regional vocabulary.
Why might Indian hospitals and banks find it useful?
Because the weights are open and the model runs on a single chip, organisations can host it on their own infrastructure. That keeps sensitive patient or customer recordings in-house, which eases compliance with the Digital Personal Data Protection Act, 2023, and avoids recurring dollar-denominated API fees.
This story was reported by TechCrunch. Read the full original coverage at TechCrunch.