Aero-1-Audio

Aero-1-Audio is a lightweight yet powerful audio-language model that delivers robust performance in speech recognition and audio understanding despite its compact 1.5B parameter size. Trained on 50K hours of high-quality data, it handles 16-minute continuous audio without segmentation—outperforming larger models in long-form ASR tasks. Features include real-time transcription, scene analysis, and audio instruction following.

Ideal for developers needing efficient audio processing, Aero-1-Audio achieves these results with just one day of training on 16 H100 GPUs. Its MIT license and sample-efficient design make it accessible for applications like meeting transcription, voice assistants, and audio analytics.