Clustering Spoken Audio Samples by Delivery Style using Acoustic Features

Portland State University – Data Clustering Course Final Project This project clusters 200 spoken samples of the word “backward” from the Google Speech Commands dataset. Using acoustic features like MFCCs, pitch, tempo, and energy, the goal is to group samples by how the word is spoken - focusing on delivery style rather than content. The word is kept constant to isolate stylistic differences in speech.

Project Setup

Note: This project uses a Hugging Face token to authenticate and stream data from the Google Speech Commands dataset. To reproduce the code, you'll need to create a Hugging Face account and store your token in a .env file as described in the auth.py script.

Step 1: Install dependencies

pip install -r requirements.txt

Step 2: Run the dataset loading script

python load_dataset.py

This script

• Authenticates with Hugging Face Hub using a token stored in your .env file
• Streams the Google Speech Commands dataset
• Filters and loads 200 samples labeled "backward"
• Extracts acoustic features for each audio sample:
  • MFCCs (Mel-frequency cepstral coefficients)
  • Pitch
  • Tempo
  • Energy
• Saves the extracted features to backward_features.csv for clustering and analysis in the notebook

Step 3:

Open and run the cells in demo-notebook.ipynb to:

• Perform KMeans, DBSCAN, and Agglomerative clustering
• Analyze cluster patterns
• Visualize results
• Play example audio clips from each cluster