DroneClassifier — Real-Time Drone Audio Detection

Model. A lightweight CNN utilizing depthwise convolutions and mean+max pooling over the time dimension; no GRU is used. Recurrence is unnecessary for a one-second window, and omitting it simplifies export to ONNX and C++.

Data. The fit() function reads all shards from data/*.parquet once, decodes the audio into 16 kHz mono int16 format, and stores it in a memory-mapped cache. This cache is built once and does not require loading all 39 shards into RAM. Before execution, the script displays the mapping for label values ​​(if the Parquet files follow the HuggingFace format). The specific value corresponding to "engine" is defined via the --positive-label parameter.

Augmentation. Speed ​​variation: ±12%. Low-pass and high-pass filtering, and reverberation. Mixing in real "other" (background/non-target) clips or colored noise at an SNR of 0–25 dB. Random gain adjustment: −15 to +10 dB. Inherent microphone noise. SpecAugment (frequency and time masking) applied to batches.

Training. Balanced class sampling, AdamW optimizer, OneCycle learning rate schedule, AMP (Automatic Mixed Precision) on CUDA, and early stopping. A checkpoint is saved based on the best validation AUC, with a threshold determined by the F1 score.

Detection. detect_file() processes data in batches, smooths the output, applies hysteresis, and returns detected events with their respective start and end times.

Export. export_onnx() generates engine.onnx, mel_fb.f32, and engine_meta.txt. The Mel-filter matrix is ​​derived directly from torchaudio, ensuring the C++ implementation does not rely on a custom implementation of the filter bank.