Voice AI in vogue: Nvidia releases new open transcription model

You can now get better, more accurate transcriptions for business meetings and medical dictation at no cost.

Nvidia has released Nemotron 3.5 Transcribe, a new open-weight model designed for audio transcription across 40 languages. The model can translate 31 languages to and from English, process audio up to an hour in length, and distinguish up to eight speakers. Nvidia claims it outperforms existing benchmarks, including the Hugging Face Open ASR Leaderboard. Alongside Transcribe, Nvidia also updated Granary, its open-source speech dataset, to version 2. These tools are aimed at developers creating applications for business meetings and medical transcription, offering customization capabilities for industry-specific terminology in multiple languages. Nemotron 3.5 Transcribe operates on recorded audio, not live streams, and at 30 billion parameters, it is designed to run on a single chip, from embedded systems like the Jetson AGX Thor to datacenter hardware. This release coincides with a period of heightened activity in voice AI, with other companies like Mistral, ElevenLabs, and Salesforce also launching new voice models.

Nvidia's release of Nemotron 3.5 Transcribe and the updated Granary dataset signals a push towards democratizing sophisticated voice AI capabilities. By offering these tools as open-weight models, Nvidia aims to accelerate the development and adoption of specialized voice applications, particularly in high-value enterprise sectors like business communications and healthcare. This move challenges proprietary solutions and potentially lowers the barrier to entry for smaller developers and startups looking to integrate advanced speech-to-text and translation functionalities into their products.

The company's emphasis on running the 30 billion parameter model on single chips, from edge devices to datacenter hardware, suggests a focus on efficiency and broad applicability. This strategy could enable more powerful AI voice features to be deployed in diverse environments, from on-device processing to scalable cloud-based services. Nvidia's aim to allow customization for industry jargon in multiple languages further underscores its intent to capture significant market share in specialized AI voice markets.