Auto-ACD: A Large-Scale Audio-Text Dataset with Long Descriptions

Auto-ACD is a large-scale, high-quality dataset containing over 1.9 million audio-text pairs. The dataset builds on the audio-visual correspondence found in existing video datasets like VGGSound and AudioSet. It was created by Loie and is hosted on Hugging Face, with a paper published in 2023.

Use Cases

• Training audio-language models based on the described audio-text pairs.
• Benchmarking audio captioning systems based on the long text descriptions.
• Studying audio-visual correspondence by leveraging the dataset's foundation on video datasets.
• Fine-tuning models for audio retrieval using natural language queries.

Strengths

• Contains over 1.9 million audio-text pairs, indicating a large scale.
• Text descriptions are noted to be long, averaging 18 words per description.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Loie (author), hosted on Hugging Face.

Collection Method

Likely derived or built upon existing video datasets VGGSound and AudioSet.

Freshness

Last updated 2025-02-06 02:50:16; freshness should be verified.