🎧 SFX Classifier — Sort your sound palette in one command!

A lightning‑fast CLI that sifts through a folder of miscellaneous audio samples and neatly places them into four musical drawers: Drums · Synths · Voice · Water.

---

✨ Why did I make this?

🔑 Feature	🚀 Benefit
📂 One‑liner sorting	Point to any directory — even thousands of files deep — and watch it organise itself.
🧠 Trained CNN under the hood	Uses our own SampleClassifier weights for studio‑grade accuracy.
⚡ GPU optional	Harness CUDA in a single flag, or run comfortably on CPU.
🤝 Non‑destructive mode	--copy keeps your originals untouched; perfect for archival libraries.
🔌 Minimal dependencies	torch, librosa, numpy, tqdm, soundfile. That’s it.
🛠 Easy to extend	Swap weights, add new classes, or embed in a bigger pipeline.

---

🖇️ Installation

# Create & activate a clean environment (optional but recommended)
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate

pip install torch librosa numpy tqdm soundfile

Clone or drop classify_samples.py + your weight file (audio_classifier_cnn.pth) into your project folder.

---

🚀 Quick Start

python classify_samples.py /path/to/unsorted_samples

Common flags

--weights  path/to/weights.pth   # use alternative weights file
--cuda                          # infer on the first available GPU
--copy                          # copy instead of move (keeps originals)

Full example:

python classify_samples.py ./previews --weights audio_classifier_cnn.pth --cuda --copy

After completion you’ll see:

Classifying: 100%|█████████████| 1472/1472 [01:12<00:00, 20.3it/s]
kick_018.wav          → drum_samples
splash_large.flac     → water_samples
...

Summary
────────
drum_samples   : 519
synth_samples  : 411
voice_samples  : 286
water_samples  : 256

…and four freshly minted sub‑folders inside your target directory.

---

🔍 How it works

1. Pre‑processing — stereo→mono, resample to 24 kHz, mel‑spectrogram (128 mels, 50 fps), zero‑pad/truncate to ~13 s.
2. CNN inference — architecture copied from the notebook you trained: two conv blocks → adaptive‑avg‑pool → dense.
3. Hard max — highest logit dictates the class.
4. File IO — sorted with shutil.move (or copy); already‑sorted files are skipped.

Swap in your own spectrogram routine or model by editing preprocess() and SampleClassifier.

---

🧩 Extending & Embedding

• Add classes: list new folder names in classes, adjust num_classes in the model, retrain & save new weights.
• Batch scripts: integrate into Ableton/FL Studio prep, run from CI, or wrap in a PyQt GUI — it’s just Python.
• Dataset building: pass --copy and keep originals tidy while you curate training data.

---

🤗 Contributing

Pull requests are welcome! Please open an issue for feature requests or bugs.

1. Fork the repo
2. Create a feature branch: git checkout -b cool_feature
3. Commit changes & push
4. Open a PR 🎉

---

📝 License

Released under the MIT License — free to use, modify, and distribute, with attribution.

---

🙏 Credits

Built upon the brilliant open‑source ecosystems of PyTorch, librosa, NumPy, and Freesound.

May your sample library always be organised and your creativity unblocked!