Inspiration
Sleep deprivation is often overlooked because people normalize exhaustion as part of daily life—especially students, healthcare workers, and professionals working long hours. We were inspired by the idea that something as simple and accessible as the human voice could reveal hidden physiological conditions. With the growing relevance of artificial intelligence in healthcare, we aimed to explore how machine learning could contribute to early detection and awareness of mild sleep deprivation, supporting preventive health practices aligned with SDG 3: Good Health and Well-being.
What it does
SleepSpec detects mild sleep deprivation by analyzing subtle changes in a user’s voice. The system segments recorded speech into 15-second intervals, extracts spectro-temporal modulation (STM) features, and classifies each segment using a Support Vector Machine (SVM) model. It provides overall classification results, prediction statistics such as confidence and decision scores, segment-level analysis, feature visualization, and evidence-based recommendations
How I built it
The project began with collecting and preprocessing voice recordings under sleep-deprived and non-sleep-deprived conditions. Each recording was segmented into 15-second samples, and STM features—specifically frequency-rate, frequency-scale, and scale-rate components—were extracted. A Support Vector Machine (SVM) classifier was trained and evaluated using standard performance metrics such as accuracy, precision, recall, F1-score, and balanced accuracy. The final model was integrated into a mobile application built with a client-server architecture, including background noise reduction methods such as Wiener Filtering and DeepFilterNet to enhance real-world performance.
Challenges I ran into
One of the major challenges was handling real-world audio variability, including background noise and inconsistent speech patterns. Feature extraction and parameter tuning for the SVM required careful experimentation to achieve balanced performance without overfitting. Ensuring stable generalization between training and validation datasets was also critical. Integrating the machine learning pipeline into a mobile environment while maintaining performance and usability presented additional technical complexity.
Accomplishments that I'm proud of
We successfully developed a system that achieves highly consistent performance metrics, demonstrating stable and balanced classification of mild sleep deprivation. Integrating advanced spectro-temporal modulation analysis into a functional mobile application is a significant achievement. The ability to visualize extracted features and provide interpretable prediction statistics adds transparency and usability to the model.
What I learned
Through this project, I learned how machine learning models can be applied beyond theoretical experiments and deployed into real-world applications. I gained deeper knowledge in signal processing, feature engineering, model evaluation, and system integration. Most importantly, I learned the importance of balancing technical accuracy with user-centered design.
What's next for sleepspec
Future improvements include expanding the dataset for stronger generalization, enhancing real-time processing efficiency, exploring hybrid or deep learning models, and conducting broader real-world validation studies. Long-term goals involve integrating wearable data, improving multilingual support, and further aligning the system with digital health monitoring solutions.
Log in or sign up for Devpost to join the conversation.