AudioX-IFcaps: 7 Million Instruction-Following Audio Samples with Timestamps

AudioX-IFcaps contains over 7 million audio samples with instruction-following captions, developed by HKUSTAudio for ICLR 2026. The dataset provides structured annotations for audio and music generation, focusing on sound event categories, counts, and temporal ordering.

Use Cases

• Training text-to-audio models to generate specific sound event counts based on text prompts
• Developing music generation systems that utilize temporal ordering annotations for structured compositions
• Fine-tuning multimodal models for timestamp-accurate audio captioning and event detection

Strengths

• Scale of over 7 million audio-text pairs
• Structured annotations including sound event categories and counts
• Inclusion of temporal ordering and precise timestamps for audio events

Limitations

• CC BY-NC-ND 4.0 license prohibits commercial use and the creation of derivative works
• High storage and compute requirements due to the 7 million sample scale

Provenance

Source

HKUSTAudio (ICLR 2026 publication Arxiv 2503.10522)

Freshness

Last updated February 2026.