feat: Add Sound Event Detection (SED) module using YAMNet backbone

[Alokkohli200]

Overview

I have implemented the Sound Event Detection (SED) module as part of the Intelligent CC Suggestion Tool. This module is designed to identify contextually significant non-speech audio events to enhance accessibility for regional content.

Technical Implementation

• Model: Utilized the YAMNet architecture via TensorFlow Hub, chosen for its high accuracy in environmental sound classification.
• Pipeline: Developed a robust extraction process that converts video audio to 16kHz mono waveforms as required by the model.
• Filtering Logic: Implemented a confidence-based filter (threshold 0.25) to ignore ambient noise (Silence, White noise) while capturing impactful events like footsteps, music, or sirens.
• Environment: Optimized and verified the pipeline for Apple Silicon (M4) using Python 3.12, with specific dependency pinning (protobuf<5) to ensure compatibility.

Demo & Verification

• Demo Video: https://www.youtube.com/watch?v=JVOTMxdUcbs
• Sample Output: The module successfully generates a timestamped table of events with confidence scores.

Future Improvements

• Implement temporal smoothing to merge consecutive short-window detections into single, continuous captions.
• Integration with the SRT generator module for automated caption insertion.