Fully Differentiable Neural Forced Alignment via Soft Dynamic Programming
Paper • 2606.25460 • Published
FALCON — pretrained checkpoints
Pretrained checkpoints for FALCON (Forced Alignment through Contrastive Optimization Networks), the neural forced aligner from the paper "Fully Differentiable Neural Forced Alignment via Soft Dynamic Programming" (arXiv:2606.25460).
| File | Trained on | Best for |
|---|---|---|
falcon_timit_english.pt |
TIMIT (read English) | English phoneme alignment |
falcon_buckeye_english.pt |
Buckeye (spontaneous English) | Spontaneous / conversational English |
falcon_joint_multilingual.pt |
Joint TIMIT+Buckeye | Cross-lingual / multilingual zero-shot alignment (Dutch, German, Hebrew, …) at phoneme and word level |
- Interactive demo: https://huggingface.co/spaces/MLSpeech/FALCON
- Code: https://github.com/MLSpeech/FALCON
Each checkpoint is a PyTorch state dict with the model hparams and the dill-serialized
peak-detection parameters. Load them with the FALCON code (predict.py / app.py), which
auto-downloads from this repo when HF_MODEL_REPO is set.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support