Nvidia has introduced a new open-weight model for turning recorded speech into text. Nemotron 3.5 Transcribe is free and supports audio in up to 40 languages, with features aimed at developers building tools for settings such as business meetings and medical transcription.
What the model can do
Nvidia says Nemotron 3.5 can translate 31 languages to and from English. It can also process recordings up to an hour long and distinguish up to eight speakers, capabilities that could help make longer conversations easier to transcribe and review.
The company says the model beats rivals on benchmarks, including Hugging Face’s Open ASR Leaderboard. Those performance claims come from Nvidia; the source article does not provide benchmark scores or a detailed comparison with other models.
Designed for specialized workplace tools
Business meetings and medical transcription are two established workplace uses for AI voice technology. In these settings, a transcript may need to reflect specialized terms as well as the language being spoken.
Nvidia says developers can combine Nemotron 3.5, the updated Granary speech dataset, and tools such as its NeMo Speech model to customize applications. The goal is to help an app recognize an industry’s jargon in each language it supports. That customization path gives developers a way to adapt transcription tools to particular fields, according to the company.
Recorded audio, with a range of hardware options
There is a clear limit: Nemotron 3.5 Transcribe works with recorded audio, not live streams. That makes it suited to processing existing recordings, while applications that need live speech transcription would need a different approach.
Nvidia says the 30 billion parameter model can run on a single chip. The company names the Jetson AGX Thor robotics controller, the DGX Spark chip powering new laptops, and Blackwell datacenter chips as hardware options. The stated range spans several kinds of devices, although the source does not give additional details about speed or resource requirements.
Nemotron 3.5 joins Nvidia’s open-weight voice models, including Parakeet, which has 1.1 billion parameters. The new model is larger, while Nvidia’s portfolio also includes this much smaller option.
A crowded field for voice AI
Nvidia’s release arrives amid a busy period for voice products. Mistral shipped Voxtral 3 earlier this month, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released its Launa-Omni voice model.
For developers, the practical choice depends on the task. Nemotron 3.5 is presented as a tool for recorded audio, with multilingual transcription, translation, and speaker separation. Its lack of live-stream support is a key consideration for teams building applications around conversations as they happen.