Time-Varying Emotion Perception Annotations for Live Music Performances

Note: This dataset is a work in progress and will be continuously updated as the full study is completed.

Introduction: 

Typically, music emotion research is conducted with participants who share similar characteristics, or
it relies on retrospective summative judgments regarding the perceived emotions of music. However, there is a lack of emotion-annotated data collected from multiple participants (N>10) assessing audio material in real-time. This dataset addresses this gap, featuring a substantial number of participants and a diverse selection of piano music performance excerpts.

Dataset Details:

• Participants: A total of 128 participants, representing diverse demographics including first languages, genders, and levels of musical instrument-playing experience.
• Audio Material:The dataset comprises 51 1-minute unique international award winning piano performances from the Western canon spanning various musical eras, with a specific focus on perceived emotion throughout each piece's duration.
• Annotation Platform: Participants used a web-based platform developed for this study, enabling time-varying emotion annotations and survey completion. The platform offers universal access via standard web browsers. 
• Annotation Method: The platform employs a Valence-Arousal (VA) model and provides guide emotion tags to facilitate emotion rating throughout the audio excerpts.

Emotion Ratings:

• A total of 133,477 emotion VA ratings were collected from all 128 participants across all 51 clips over time.
•  On average, there are 20.5 emotion VA ratings per one-minute clip per participant, with a standard deviation of 29.3.

File Summaries:

1. Raw_ratingspoints_nodup_128p_51samples.csv: Contains all VA points collected from the 51 audio samples annotated by 128 participants.
2. Cleaned_ratingspoints_nodup_128p_51samples.csv: Presents updated VA rating points for the 51 music samples from 128 participants, with the most recent ratings retained.
3. subscale_score_sum.csv: Provides background survey scores (18 distinct measures) for each of the 128 participants.
4. 51samples_url_meta.csv: Includes metadata about the 51 music clips chosen from the MEASTRO dataset for this study.
5. clean-up-function-220608.ipynb: A Jupyter notebook detailing how to retain only the most recent ratings for participants who may rewind and re-rate clips or provide multiple ratings on duplicate audio locations. It includes a comprehensive explanation of the updating process.
6. test-nodup.csv: A mock sample of VA ratings used to test the feasibility of the clean-up function.