This model is a Sound Event Detection (SED) system designed to detect and extract specific audio events from long-form recordings. It is built on the PANNs (Pretrained Audio Neural Networks) architecture, specifically utilizing the CNN14 backbone with a Multiple Instance Learning (MIL) head for weak-label training and frame-level inference.
Model Details
- Architecture: CNN14 (PANNs) + MIL Head
- Task: Sound Event Detection (SED) / Audio Classification
- Input Sampling Rate: 32,000 Hz
- Primary Target: Specific sound event detection (e.g., sniffing, breathing, or vocalizations depending on the finetuning dataset).
- Time Resolution: ~10ms - 160ms depending on the hop size and pooling configuration.
Disclaimer
This repository contains the model weights only. It is not a standalone executable and requires a compatible PANNs-based inference script to function.
This model is part of a project called AccuSlice-SED that hasn't rolled out yet. I wanted to upload the weight to my private repository, but I was told "You've reached private storage quota. Upload it to public!" so I'm doing this reluctantly. :/