Audio-FLAN Dataset

Summary

Audio-FLAN is a large-scale instruction-tuning dataset for unified audio understanding and generation, integrating 50+ source datasets across speech, music, and general audio domains. It covers tasks like speech recognition, text-to-speech, acoustic scene classification, audio event recognition, and music generation, organized with structured metadata (instruction, input, output, task type) for training audio-language models.

Keywords

audio speech music sound instruction-tuning text-to-speech asr audio-generation

Topics

Audio / Speech / Music

Research notes

• Requires sharing contact information for access. Audio included only for datasets with permissive licenses (CC-BY, Apache-2.0, MIT); metadata-only for restrictive licenses. Paper: arxiv 2502.16584.