This project proposes an approach for word-level stress identification from speech using prosodic features, assuming that the corresponding transcribed text is available.
The dataset consists of a CSV file with the following columns:
- 📌 Audio Path – Path to the audio file
- 📜 Transcribed Text – Manually transcribed speech
- 🔤 Stress Labels – Word-level stress annotations (e.g., stressed/unstressed)
You can access the dataset here:
👉 Raw audio files for training
| File / Directory | Description |
|---|---|
setup_env.sh |
Shell script to set up the development and training environment |
config.py |
Contains all configuration parameters (paths, hyperparameters, etc.) |
dataset.py |
Defines a custom PyTorch-compatible dataset class for loading and preprocessing audio data in NeMo-ASR-compatible format |
model.py |
Contains the model architecture for stress classification |
train_test.py |
Includes PyTorch training and evaluation loop logic |
utils.py |
Utility functions for audio loading and prosodic feature extraction |
stress_classification_model.ipynb |
Jupyter Notebook entry point to train and test the stress classifier |
-
Clone the repository:
git clone <repo_url>cd <repo_directory>
-
Set up the environment:
- create a new
condaenv withpython-3.10version chmod +x setup_nv.sh./setup_env.sh
- create a new
-
Run the notebook:
- Open and Run
stress_classification_model.ipynbto train the model.
- Open and Run