Nvidia has released a new open-weight artificial intelligence model designed for audio transcription, adding another tool to its growing portfolio of speech-focused AI products. The model, called Nemotron 3.5 Transcribe, is available at no cost and supports transcription across up to 40 languages. Alongside the model, Nvidia also published version 2 of Granary, its open source speech dataset used for training and fine-tuning voice AI systems.

According to Nvidia, Nemotron 3.5 Transcribe is capable of translating speech between English and 31 other languages, processing audio recordings of up to one hour in length, and distinguishing between as many as eight individual speakers in a single recording. The company says the model outperforms competing systems on several industry benchmarks, including the Open ASR Leaderboard maintained by Hugging Face, which tracks the performance of automatic speech recognition systems.

Targeting Business and Medical Applications

The primary audience for these tools is developers building software for professional environments, particularly in sectors where voice transcription has already gained meaningful traction. Business meeting transcription and medical documentation are two of the more established use cases for AI-driven speech recognition, and Nvidia has positioned Nemotron 3.5 Transcribe with those workflows in mind.

Nvidia notes that developers can combine the model with the updated Granary dataset and its existing NeMo Speech framework to build applications capable of recognizing industry-specific terminology across multiple languages. This customization capability is particularly relevant in fields like healthcare and legal services, where specialized vocabulary is common and accurate transcription is critical.

One limitation worth noting is that Nemotron 3.5 Transcribe is designed to work with pre-recorded audio files rather than live audio streams. This means it is not suited for real-time transcription scenarios, which could be a factor for developers evaluating it against tools that support live input.

Hardware Compatibility and Model Scale

Despite its size of 30 billion parameters, Nvidia says the model is engineered to run on a single chip. The company highlights compatibility across a broad range of its own hardware, from the Jetson AGX Thor robotics controller and the DGX Spark chip found in new laptop-class devices, to Blackwell-architecture datacenter processors. This flexibility is intended to make the model accessible to a wider range of deployment environments, from edge computing scenarios to large-scale cloud infrastructure.

Nemotron 3.5 Transcribe joins a lineup of open-weight voice models that Nvidia has been building out in recent months. Among them is Parakeet, a significantly smaller model at 1.1 billion parameters that has drawn attention for delivering strong performance relative to its compact size.

A Crowded Moment for AI Voice Technology

Nvidia's release arrives during a notably active stretch of product launches across the AI voice space. Earlier this month, Mistral introduced Voxtral 3, its own speech-capable model. ElevenLabs rolled out a Realtime feature aimed at enabling instant speech generation within voice agent applications. And Salesforce released Launa-Omni, a voice model developed as part of its broader AI product strategy.

The flurry of releases reflects growing competition among AI companies to establish a foothold in voice and audio processing, a segment that is increasingly seen as a key interface layer for enterprise software, productivity tools, and autonomous systems. Nvidia's entry into this space with an open-weight, multi-language model signals the company's intent to extend its influence beyond hardware and into the software and model ecosystem that runs on top of it.