Crowd Sound Project - Crowd Sound Project

Dataset Description¶

Name: Crowd Sound Project
Version: 1.0.0
Authors: Oceane Mourgues
License: MIT
DOI: N/A

Resume¶

Here is the resume of the project:

This project focuses on the organization, visualization, and exploration of crowd sound stimuli. Crowd vocalizations are widely used to investigate emotion perception, multisensory integration, and social behavior, but current crowd-audio stimuli are often custom-made, difficult to reproduce, poorly standardized, and limited in experimental control.

The objective of this project is to develop an interactive and reproducible framework for exploring crowd sounds through parameter-based organization, audio visualization, and structured metadata. The system aims to support the comparison and analysis of auditory stimuli using spectrograms, waveforms, and acoustic parameter displays.

The project organizes sounds according to several perceptual and acoustic dimensions, including temporal, dynamic, vocal, spectral, and spatial parameters. This framework is intended to facilitate the creation, organization, and reuse of auditory stimuli in experimental research contexts.

The project also emphasizes reproducibility and open science practices through the use of Python, Jupyter notebooks, GitHub, and structured workflows for audio processing and metadata organization. Final deliverables will include reproducible scripts, interactive audio exploration tools, audio visualizations, structured datasets, and documentation to support reuse and long-term accessibility.

Project Structure¶

crowd_sound_project/
│
├── README.md
├── myst.yml                        # Jupyter Book configuration
├── environment.yml
│
├── data/
│   └── sounds/
│       ├── A1.wav ... A26.wav      # Group A (26 files)
│       └── J1.wav ... J26.wav      # Group J (26 files)
│
├── notebooks/
│   ├── 01_extraction.ipynb         # Feature extraction
│   ├── 02_visualization.ipynb      # Visualization and analysis
│   └── 03_interactive.ipynb        # Interactive visualizations, UMAP, │
│
├── src/
│   ├── crowd_analysis.py
│   └── visualization.py
│
├── figures/
│   ├── boxplots_by_group.png
│   ├── pca_mfcc_individual.png
│   └── kmeans_mfcc.png
│
└── outputs/
    ├── A1/results.json ... 
    └── feature_inspection.json

This structure is designed to support reproducibility, organization, and open science practices throughout the project development process.

Presentation link¶

Click here to view the slides: https://docs.google.com/presentation/d/1v9t9jJARwTjZ2fG6t1Csj8lXJtg4WSRqbLEC1cxJR6A/edit?usp=sharing

Site link¶

Site : https://brainhack-school2026.github.io/Mourgues_Project/

Procedure¶

Step 1 – Audio Feature Extraction¶

Notebook: `notebooks/01_extraction.ipynb`

File: src/crowd_analysis.py

Objective¶

Extract acoustic features from each .wav audio file using librosa.

Extracted Features¶

duration
global intensity (RMS)
fundamental frequency (F0)
spectral centroid
spectral bandwidth
spectral rolloff
MFCC (13 coefficients)

Output Each audio file generates a JSON file in outputs/<sound_name>/results.json.

Step 2 – Visualization and Analysis¶

Notebook: notebooks/02_visualization.ipynb File: src/visualization.py

Objective Compare audio signals between groups (A vs J) and analyze their acoustic structure.

2.1 Feature Inspection Ranks sounds by each feature (highest/lowest F0, centroid, bandwidth, rolloff). Output saved to outputs/feature_inspection.json.

2.2 Boxplots

F0 (pitch)
Spectral centroid (brightness / noise differences)

2.3 PCA (dimensionality reduction) Reduces MFCC features from 13D → 2D using PCA(n_components=2).

pca_mfcc_individual.png → individual labeled points

2.4 K-Means clustering Checks if the algorithm can recover groups without labels (k=2).

kmeans_mfcc.png

Step 3 – Interactive Exploration¶

Notebook: notebooks/03_interactive.ipynb

Objective Explore sounds interactively with Plotly and ipywidgets.

Features

PCA and UMAP projections with toggle selector (group filter)
Click any point → play the corresponding audio
Waveform + spectrogram display per sound (Plotly)
Acoustic parameters table per sound
Free feature scatter plot (X/Y dropdowns) with audio playback on click

Statistical comparison (Group A vs J)

Boxplots for all features (F0, RMS, centroid, bandwidth, rolloff, duration)
Mann-Whitney U tests with p-values

Key results

F0 mean: p=0.0001 *** — Group J significantly higher pitch (300 Hz vs 204 Hz)
Spectral centroid: p=0.047 * — Group J slightly brighter
Other features: no significant difference

Step 4 – Feature Interpretation¶

Feature	High value	Low value
F0	High-pitched sounds	Low-pitched sounds
Spectral centroid	Noisy / mixed sounds	Stable / pure sounds
PCA separation	Distinct classes	Acoustic similarity
K-Means clusters	Good separability	High variability

Step 5 – Run the project¶

jupyter nbconvert --to notebook --execute notebooks/01_extraction.ipynb --inplace
jupyter nbconvert --to notebook --execute notebooks/02_visualization.ipynb --inplace
jupyter nbconvert --to notebook --execute notebooks/03_interactive.ipynb --inplace

Or run the original scripts directly:

python src/crowd_analysis.py
python src/visualization.py

Conclusion¶

This project allows:

extraction of acoustic features from 52 crowd sound stimuli
statistical comparison between groups A and J (Mann-Whitney U tests)
visualization of dataset structure via PCA, UMAP, and K-Means
interactive exploration with audio playback, waveform, and spectrogram
testing class separability
exploration of acoustic variability