Nvidia launches Nemotron 3.5 Transcribe, an open‑weight multilingual transcription model

Nvidia launches Nemotron 3.5 Transcribe, an open‑weight multilingual transcription model

Lead (5W1H) On Wednesday, Nvidia released Nemotron 3.5 Transcribe, a free, open‑weight speech‑to‑text model that can transcribe audio in up to 40 languages, translate 31 languages to and from English, and differentiate up to eight speakers in recordings as long as an hour. The launch, reported by TechCrunch, also includes an updated version of Nvidia’s Granary speech dataset and is aimed at developers building business‑meeting and medical‑transcription applications.


New open‑weight model and dataset

According to TechCrunch, the centerpiece of Nvidia’s announcement is Nemotron 3.5 Transcribe, a 30‑billion‑parameter model that is freely available for developers. The model is described as “open‑weight,” meaning its parameters can be inspected and fine‑tuned without licensing restrictions.

In tandem with the model, Nvidia released Granary v2, the second iteration of its open‑source speech dataset. TechCrunch notes that the dataset is intended to help developers customize the transcription engine for industry‑specific jargon across the supported languages.

Multilingual capabilities and technical specs

As reported by TechCrunch, Nemotron 3.5 Transcribe supports transcription in up to 40 languages. In addition, the model can translate 31 languages to and from English, positioning it as both a transcription and translation tool.

The model is engineered to handle recordings up to an hour long and can distinguish up to eight separate speakers within a single audio file. TechCrunch adds that the model outperforms competitors on benchmark tests, including the Open ASR Leaderboard hosted by Hugging Face.

Hardware flexibility

According to TechCrunch, Nemotron 3.5 Transcribe’s 30‑billion‑parameter architecture is optimized to run on a single chip. The company cites several hardware platforms capable of executing the model, ranging from the Jetson AGX Thor robotics controller and the DGX Spark chip used in new laptops, to the Blackwell datacenter processors.

The announcement also references Nvidia’s broader portfolio of open‑weight voice models, highlighting the Parakeet model, which contains 1.1 billion parameters and is described as “super small yet powerful.”

Target applications and developer tools

TechCrunch indicates that Nvidia is positioning the new model for business‑meeting transcription and medical transcription, two of the most established use cases for voice AI in the workplace. Developers are encouraged to combine Nemotron 3.5 Transcribe with the updated Granary dataset and Nvidia’s NeMo Speech toolkit to fine‑tune the system for industry‑specific terminology in each supported language.

A limitation noted by TechCrunch is that Transcribe works only with recorded audio, not with live streaming inputs.

Market context

The release arrives during a period of heightened activity in the voice‑AI sector. According to TechCrunch, Mistral shipped Voxtral 3 earlier in the month, ElevenLabs introduced Realtime for instant speech generation in voice agents, and Salesforce launched Launa‑Omni, another voice model. Nvidia’s entry adds a free, open‑weight option to a competitive landscape dominated by proprietary solutions.


Brief analysis (opinion)

In my view, Nvidia’s decision to make a high‑capacity multilingual transcription model openly available could accelerate innovation among smaller developers who lack the resources to train comparable systems from scratch. By bundling the model with an updated speech dataset and a flexible toolkit, Nvidia lowers the barrier to entry for niche applications such as specialized medical dictation or multilingual corporate meetings. However, the restriction to pre‑recorded audio may limit its appeal for real‑time transcription services, a segment where competitors like ElevenLabs are already making strides.


Conclusion

Nvidia’s launch of Nemotron 3.5 Transcribe and Granary v2 represents a significant addition to the open‑source voice‑AI ecosystem, offering extensive multilingual support, speaker separation, and hardware versatility—all without licensing fees. As the voice‑AI market continues to expand, the availability of a free, high‑capacity model may influence how developers approach transcription and translation solutions across various industries.

Frequently Asked Questions

What languages can Nemotron 3.5 Transcribe handle?

According to *TechCrunch*, the model can transcribe audio in up to **40 languages** and translate **31 languages** to and from English.

Can the model process live audio streams?

*TechCrunch* notes that **Transcribe works only with recorded audio**, not with live streaming inputs.

Which hardware platforms are capable of running Nemotron 3.5 Transcribe?

As reported by *TechCrunch*, the model can run on a single chip, including Nvidia’s **Jetson AGX Thor**, **DGX Spark**, and **Blackwell** datacenter processors.

Source: TechCrunch — Read more

💻 Related Articles