The Cadenza lyric intelligibility prediction (CLIP) dataset

Area of Science:

• Music Information Retrieval (MIR)
• Machine Learning
• Signal Processing

Background:

• Existing datasets lack comprehensive lyric intelligibility data for MIR.
• Developing algorithms to predict lyric intelligibility is crucial for various applications.

Purpose of the Study:

• Introduce the CLIP dataset, a large-scale resource for MIR research.
• Facilitate the development of machine learning models for lyric intelligibility prediction.
• Support the Cadenza ICASSP 2026 Signal Processing Grand Challenge.

Main Methods:

• Compiled 11,072 western music signals from independent artists (Free Music Archive).
• Generated ground truth lyrics via native English speakers.
• Simulated hearing loss (none, mild, moderate) to create 11,100 audio signals.
• Collected human transcriptions through online listening experiments to determine intelligibility scores.

Main Results:

• The CLIP dataset comprises audio, ground truth lyrics, and intelligibility scores for 11,100 music signals.
• It is the first publicly available large-scale dataset for lyric intelligibility prediction.
• The dataset represents diverse hearing conditions.

Conclusions:

• The CLIP dataset provides a valuable resource for advancing MIR research, particularly in lyric intelligibility.
• It enables the creation of more robust and accurate lyric intelligibility prediction models.
• This dataset will foster innovation in music signal processing and human-computer interaction.