Nvidia has released a free Nvidia speech to text model called Nemotron 3.5 Transcribe that can turn recorded speech into text in up to 40 languages. The company says the open-weight system is meant for developers building apps for business meetings and medical transcription, and that it does not process live audio streams.
What Nvidia released
On Wednesday, Nvidia published Nemotron 3.5 Transcribe and version 2 of Granary, its open-source speech dataset. The firm says the Nvidia speech to text model can translate 31 languages to and from English, work with recordings as long as an hour, and separate as many as eight speakers. Nvidia also says the system performs strongly on public speech-recognition benchmarks, including Hugging Face’s Open ASR Leaderboard.
At roughly 30 billion parameters, Transcribe is designed to run on a single chip, Nvidia says—from the Jetson AGX Thor robotics controller and the DGX Spark chip used in new laptops through to Blackwell datacenter processors. It sits alongside Nvidia’s other open-weight voice models, including the much smaller Parakeet system at 1.1 billion parameters.
What it can and cannot do
The main limit is timing: Transcribe accepts recorded audio, not live streams. That matters for anyone who needs instant captions or real-time voice agents. For batch work—uploaded meeting files, clinic dictation, archived interviews—the feature set is broader: long recordings, multi-speaker labels, and translation paths into and out of English for 31 languages, according to Nvidia.
Developers can pair the model with the updated Granary dataset and Nvidia AI tools such as NeMo Speech to tune apps so they recognise industry jargon in each language, the company says. That path is aimed at specialised workplace software rather than a consumer chat app.
Why Pakistani teams may care
For software houses, freelancers, and health-tech builders in Pakistan, a free Nvidia speech to text model makes it easier to prototype meeting notes and clinical transcription without starting from a paid closed service. Teams will still need hardware that can host a 30-billion-parameter model, and they should check whether the languages they need are among the up to 40 supported. Because the system is built for recordings, it fits overnight or after-call processing better than live call centres or broadcast captioning.
Aasan alfaaz mein
Nvidia ne ek naya speech-to-text model diya hai, Nemotron 3.5 Transcribe, jo free aur open-weight hai. Yeh recording sun kar text bana sakta hai, taqreeban 40 zabaanon tak. Live audio pe yeh abhi kaam nahi karta. Meeting notes aur medical transcription apps banane walon ke liye yeh useful ho sakta hai, lekin turant live captioning ke liye nahi.
