Borrison/jsalt26-csr-lab

JSALT 2026 — CSR Lab: Pre-trained Character ASR Model

Pre-trained weights for the JSALT 2026 Continuous Speech Recognition lab.

Model

• Encoder: facebook/hubert-base-ls960 (HuBERT-base, frozen for first 1000 steps)
• Output head: linear 768 → 33 (character-level vocabulary)
• Loss: CTC with blank token at index 0
• Tokenizer: CharacterTokenizer (33-class: <blk>, <pad>, <unk>, <bos>, <eos>, A–Z, space, apostrophe)
• Training data: LibriSpeech train-clean-5 (~5.4 h)

Files

File	Description
speech_encoder_char.pt	model.state_dict() — load directly with model.load_state_dict(...)

Usage

from huggingface_hub import hf_hub_download
import torch

weights_path = hf_hub_download(repo_id="Borrison/jsalt26-csr-lab",
                                filename="speech_encoder_char.pt")
state_dict = torch.load(weights_path, map_location=device)
model.load_state_dict(state_dict)