Keyclick Detector

Keyclick-Detector is a specialized utility designed to identify and isolate mechanical noise—specifically keyboard and mouse clicks—at the start and end of voice-training recordings. By accurately locating these transients, the tool allows developers and audio engineers to trim unwanted artifacts from datasets without affecting the underlying speech. It helps address the inconsistencies of fixed-duration trimming, ensuring that recordings are cleaned effectively regardless of varying human reaction times.

The project utilizes a 1-D dilated convolutional neural network trained on a custom dataset of raw, untrimmed audio takes. The model processes log-mel spectrograms to identify click events with high precision, mapping them to specific timestamps. The repository includes an integrated review pipeline that enables manual verification and refinement of model predictions, facilitating a human-in-the-loop approach to dataset curation.

This tool is particularly useful for maintainers of speech recognition or text-to-speech datasets who rely on recorder apps that capture start and stop button clicks. It provides a structured workflow for auditing and cleaning audio files, making it a valuable asset for anyone working with high-fidelity voice-training data.