Here are the available CSV files for the FME-24 dataset. You can download them by clicking the respective download buttons.
The July 2026 CSV provides the latest files developed for the Film Music Semantic Emotion
Representation (FMSER) pipeline. This includes window-based acoustic fingerprints)
The March 2026 CSVs contain the main FME-24 regression features, including acoustic features,
Valence-Arousal annotations, timestamps, participant demographics and GoldMSI responses.
Earlier CSV versions are retained for reproducibility but are no longer recommended for new
analysis.
JULY 2026 Update: CSVs for FMSER - semantic emotion representation and window-based feature extraction
audio_fingerprints.csv
Audio fingerprint features extracted from the FME-24 film music excerpts.
This CSV contains timestamped acoustic feature windows used for modelling
relationships between musical characteristics and listener-derived emotion representations.
Each row represents a temporal audio window extracted from a film music excerpt.
The dataset includes low-level and high-level acoustic descriptors including MFCC
statistics, spectral features, RMS energy, zero-crossing rate, harmonic and
percussive energy ratios, pitch-related features, rhythm descriptors, onset density,
danceability, key, scale and tonal information.
These features were extracted using fixed temporal windows and are designed for
machine learning experiments involving temporal prediction of semantic emotion
representations. The CSV provides the acoustic input features used for FMSER
prediction experiments.
MARCH 2026 Update: CSV for continuous emotion analysis V-A (regression)
Main CSV: This CSV contains 1784 rows and 91 columns, including 78 acoustic, spectral, harmonic, and rhythm features extracted from 422 audio excerpts across 98 participants and 275 unique songs, providing a comprehensive dataset for audio-emotion and music analysis, with V-A annotations, timestamps, familiarity ratings and emotion sentences.
The dataset captures timestamped segments of audio annotated with emotional responses, familiarity, valence, and arousal. Each segment includes a rich set of features such as MFCC means and variances, Tonnetz descriptors, spectral centroids, zero-crossing rate, RMS, harmonic and percussive energy ratios, pitch statistics, chord information, key and scale, rhythm metrics including beats per minute and onset density, and other derived measures of timbre and structure. The data is organized by participant and file, allowing detailed exploration of both perceptual annotations and precise acoustic measurements for each excerpt.
fme-demographics-music-sophistication.csv
Participants info: GoldMSI and Demographics CSV. This CSV contains GoldMSI responses and demographic information about the participants.
Particiipant ID corresponding with demographics and Music Sophistication inputs per participant. 100 participants although 2 were removed in the march_2026_csv due to lack of V-A annotations and timestmaps. This CSV can be used to look aat certain subgroups within the main dataset, or for more information in demographics in the dataset.
per-song-fme.csv
This CSV contains the raw data in a different format, where each row is one of the excerpts which includes missing data and coordinates together as arrays rather than per column, this is how the data was extracted raw from the online experiment.
Raw Data directly extracted from the online survey. This data contains the missing information before cleaned for features extraction and pre-processing, this CSV doesn't need to be used but provides information on the data before it was cleaned.
film-emotion-music-datasheet.csv
This CSV contains metadata on the music/ films/ excerpts and quality of the film music used.
This datasheet contains information on the music used in the dataset. It includes film and music metadata, timestamps extracted from the full pieces with the ISRC numbers. With track length, size, sample rate and director.
APRIL 2026 Update: Emotion Categories from Free-Text Sentences
emotion_clean_dataset_29APRIL2026.csv
Emotion-labelled dataset: This CSV extends the main FME-24 dataset by adding structured emotion categories derived from participants’ free-text emotion sentences. Each row includes refined emotion phrases, cluster-based labels, and probabilistic emotion scores (top-1, top-2, top-3), alongside the original audio features, timestamps, and annotations.
This dataset captures how participants describe emotion in natural language and maps it into interpretable categories. Emotion sentences were cleaned, split into time-aligned clauses, and reduced to short “refined emotion clauses” using a language model. These were embedded using a Sentence Transformer and grouped using clustering methods. Final emotion labels are assigned probabilistically based on similarity to category centres, allowing each data point to reflect multiple emotional influences rather than a single fixed label.
Key columns include:
• refined_emotion_clause: short, standardised emotion phrase
• cluster_name: main category label
• top1_label / top2_label / top3_label: ranked emotion predictions
• confidence: strength of the top prediction
• ambiguity: how similar the top predictions are
• final_label_clean: simplified label (e.g. “ambiguous”)
This CSV is designed for categorical emotion analysis and complements the continuous Valence–Arousal data by capturing richer, language-based emotional descriptions.
ALL 2026 CSVs:
Copy_rubysFMEsurvey_excel.csv
all_timestamps.csv
film-emotion-music-dataset-excel.csv
librosa-full-features-MARCH2026-fme-dataset.csv
SEPT 2025 Update: CSVs for emotion category analysis (not V-A)
fme-24-features-emotions-after.csv
The after CSV contains extracted audio features, arousal-valence values, and annotated emotion categories for each 2-second segment following participants’ emotion-change clicks.
The after CSV provides a row-wise dataset where each row corresponds to a 2-second segment of a song immediately after a participant indicated an emotion change. It includes 179 audio feature columns capturing high-level acoustic and temporal properties, alongside arousal and valence coordinates reflecting perceived emotion. Each segment is annotated with emotion categories, both as names and numeric labels, and includes metadata such as file_path, timestamp, and timestamp_order. This also contains emotion sentences for each of the data points which are categorised using sentence transformers. This structure allows for detailed analysis of the temporal evolution of emotional responses to music across participants.
fme-24-features-emotions-before.csv
The before CSV contains extracted audio features, arousal-valence values, annotated emotion categories, and emotion sentences for each 2-second segment immediately preceding participants’ emotion-change clicks.
The before CSV provides a row-wise dataset where each row corresponds to a 2-second segment of a song immediately before a participant indicated an emotion change. It includes the same 179 audio feature columns as the after CSV, alongside arousal and valence coordinates reflecting perceived emotion, and metadata such as file_path, timestamp, and timestamp_order. Each segment is annotated with emotion categories, both as names and numeric labels, and also contains emotion sentences for each data point, which are categorized using sentence transformers to capture semantic content. This structure enables analysis of emotional responses prior to participants’ clicks and allows for comparisons between the “before” and “after” emotional states.
OLDER VERSIONS:
FME-24-Annotations.csv
CSV of emotion annotations, timings, valence/arousal changes, ISRC number and film/music metadata.
This file contains FME-24 dataset. Each row is an individual annotation - one click where one participant felt a change in emotion.
fme_emotion_with_correct_path_matches.csv
Best for individual marker annotations, one arousal/valence coordinate per row; does not include feature extraction.
annotations_with_path_matches.csv
CSV of emotion annotations, timings, valence/arousal changes, ISRC number and film/music metadata - row K matches the audio file path names. Contains fuzzy matching of audio file names to song titles.
fme-per-sample-per-participant-annotation.csv
This CSV file layout is 662 X 16. Each row shows the time-stamped and arousal-valence annotation per sample by one participant.
469-audio-samples-fme-24-metadata.csv
CSV file of 469 audio samples that were in the cleaned version of the data and the corresponding metadata.
FME-chord-extraction-before-after.csv
CSV file of the before and after chord extraction of the timestamps in the data.
FME_24-Dataset-Librosa-Feature-Extraction.csv
CSV file of librosa features extracted from before and after timestamps of the emotion annotation in the data.
FME-dataset-metadata.csv
CSV file of the metadata of the audio files in the dataset including length, ISRC number, and sample timestamps.
fme-93-matched_participants.csv
CSV file of 93 participants responses from the original 185.
FME-Survey-Details.csv
CSV file of the uncleaned survey results. Contains demographics of participants, blank fields, and optional information.