Common questions about the pipeline, formats, training, and behavior.
The backend uses torchaudio.load(), which supports WAV, MP3, FLAC, OGG, and other formats supported by the system backend (e.g. sox, ffmpeg). For best results use 16 kHz mono WAV or let the server resample automatically.
The model is trained on fixed-length inputs (e.g. 3 s at 16 kHz = 48,000 samples). Longer files are trimmed from the start; shorter ones are zero-padded. This keeps inference fast and consistent.
This app uses torchaudio for loading and resampling. Librosa is often used in research for the same steps (load, resample, mel spectrogram). RawNetLite uses raw waveform, so we don’t extract mel features; if you use a feature-based model, you’d add a librosa/torchaudio feature step. See Process for details.
Results are stored in your browser (localStorage). They are not sent to a separate server. Clear them anytime from the Results page. A future version could back results with an API.
The model outputs a single probability P(fake) in [0, 1]. Values above 0.5 are labeled “fake”, below 0.5 “real”. You can use the raw probability for thresholds or downstream logic.
The checkpoint name suggests: data augmentation, multiple (triple) domains, cross-domain evaluation, and focal loss. See the Model page for architecture and training methodology.
Yes. We provide API keys and a hosted endpoint (in progress) so you can integrate our service into your apps without running anything locally or on a VPS—just call our API with your key. The same API is also available if you self-host. See How to use the API for endpoints, auth, and code examples.
Clone the repo, set up a Python venv in backend/, install dependencies, place the .pt model file in backend/models/, and run the Flask server. Then run the Next.js app in web/. See Get started for step-by-step setup and download guidance.
See Deployment: it covers host locally vs host on VPS, where to get a VPS (DigitalOcean, Linode, Vultr, AWS Lightsail, Hetzner, etc.), preparing the server, deploying backend and frontend, Nginx, SSL (Let's Encrypt), and security (secrets, HTTPS, firewall, rate limiting).