NVIDIA is democratizing access to AI by releasing a suite of open-source language tools aimed at empowering developers to create high-quality speech AI applications for 25 European languages. Recognizing that AI currently operates effectively in only a limited number of languages, NVIDIA’s initiative seeks to bridge this gap and bring the benefits of AI to a wider global audience.
The newly released tools include Granary, a massive library encompassing a million hours of human speech, and two innovative AI models: Canary-1b-v2, designed for precise transcription and translation, and Parakeet-tdt-0.6b-v3, optimized for real-time applications. These resources enable developers to build multilingual voice-powered tools, such as chatbots and translation services, with increased efficiency and reduced development time.
NVIDIA leverages an automated pipeline, built on its NeMo toolkit, to streamline the conversion of raw audio into structured data suitable for AI training. This automated approach significantly lowers the time and cost typically associated with traditional, human-driven annotation processes. The Granary dataset has demonstrated its effectiveness, achieving comparable accuracy with half the amount of data compared to other datasets.
Canary delivers translation and transcription performance on par with significantly larger models, while offering substantial speed improvements. Parakeet excels in real-time processing, capable of analyzing a 24-minute meeting recording in a single pass, identifying the language spoken and automatically adding punctuation, capitalization, and word-level timestamps. The dataset and models are readily available through Hugging Face.