- app.py inference was using wav2vec2-large-xlsr-53 but model was trained
on wav2vec2-lv-60-espeak-cv-ft embeddings -> always wrong predictions
- Dockerfile: pre-download correct model (lv-60-espeak-cv-ft)
- Language dropdown now optional for audio/URL classification
- Add model download endpoint for admin panel (ephemeral HF storage)
- Add Download Model button in admin results
Add admin panel for model training with SSE progress
