Unified Tag System For Speech Music And Environmental Sounds

UTS provides a unified tag vocabulary bridging speech, music, and environmental sounds derived from high-fidelity audio captions. It was created by AudenAI using Qwen3-Omni-Captioner and Qwen2.5-7B-Instruct models on a subset of CaptionStew. The dataset was last updated on March 11, 2026.

Use Cases

• Training audio classification models on the unified tag vocabulary spanning speech, music, and environmental sounds.
• Developing multi-modal models using the link between audio samples and their structured text captions.
• Benchmarking model performance on a large-scale, data-driven label system derived from 400K audio caption subset.
• Pre-training audio foundation models on a dataset that integrates multiple audio domains under a single taxonomy.

Strengths

• Derived from a 400,000-sample subset of the CaptionStew dataset.
• Unified label vocabulary created by parsing high-fidelity captions with the Qwen2.5-7B-Instruct model.

Limitations

• Specific row count, column details, and file formats are unknown.
• Geographic and temporal coverage of the source audio data is unspecified.
• Potential label noise inherited from the automated caption parsing process.

Provenance

Source

CaptionStew 400K-subset, processed by AudenAI.

Collection Method

Audio captions generated by Qwen3-Omni-Captioner, parsed into structured tags using Qwen2.5-7B-Instruct.

Freshness

Last updated March 11, 2026.