learning_frame_token_responsibilities_via_optimal_transport

FTA Conformer CTC — LibriSpeech 100h

This repository contains a Conformer CTC acoustic model trained on the LibriSpeech 100-hour (train-clean-100) subset with FTA training.

Checkpoint

• File: pretrained-epoch40-avg10.pt
• Training endpoint: epoch 40
• Averaging: accumulated averaged model over epochs (30, 40]
• Framework: PyTorch + icefall
• Checkpoint format: torch.save({"model": state_dict, ...})

The checkpoint contains the FTA inference modules under encoder.*, ctc_head.*, blank_gate.*, and blank_prior.*. It is an icefall checkpoint, not a Transformers AutoModel.from_pretrained() package.

Loading

import torch

checkpoint = torch.load(
    "pretrained-epoch40-avg10.pt",
    map_location="cpu",
    weights_only=False,
)
state_dict = checkpoint["model"]

Use the model definition and decoding pipeline in conformer_ctc2/decode_vi_blank_v2.py from the accompanying icefall codebase.

Notes

FTA-specific training components require the matching model implementation. For reproducible decoding, use the same token inventory, SentencePiece model, and decoding configuration as the original experiment.