Machine Learning
AudioFatigueDetection
Classifying vocal fatigue through mel-spectrograms.
Completed PrototypeMachine Learning / Audio ClassificationMachine Learning Developer
The friction?
1Audio Samples2Normalisation3Mel-Spectrogram4CNN Classification5Evaluation
Judgement, made visible.
Why this needed to exist.
The project explored whether voice recordings could be transformed into frequency-domain representations and used to distinguish energetic speech from tired-sounding speech.
- Context
- School Project
- Platform
- Python · Jupyter Notebooks
Complexity, given a direction.
- 1Audio Samples
- 2Normalisation
- 3Mel-Spectrogram
- 4CNN Classification
- 5Evaluation
The work behind the interface.
Mel-spectrogram input
Audio waveforms are converted into fixed-size mel-frequency representations before training.
The CNN receives a consistent visual representation of time and frequency information.
Separated notebook workflow
Data preparation, training, evaluation and inference are divided into separate notebooks and reusable utility functions.
Each stage can be inspected and repeated independently.
Evidence, not theatre.
Completed an end-to-end educational audio-classification workflow from preprocessing to inference.
Multi-format audio loading
CNN training and checkpointing
ROC, confusion-matrix and threshold evaluation pipeline