Peacockery/citrinet-256-phoneme-en

Citrinet-256 Phoneme EN

This repository stores the full P003 Citrinet-256 phoneme model artifact trained on the train_clean_100 ARPABET-41 setup.

Training Run

• Base model: nvidia/stt_en_citrinet_256_ls
• Fine-tuning target: ARPABET-41 plus tokenizer control tokens
• Tokenizer size: 44
• Full training budget: 3 epochs
• Final report: report.json
• Primary artifact: artifacts/citrinet_256_p2b_trainclean100_local_full_s3.nemo

Result Snapshot

• Global step: 5353
• Seed: 17
• Validation WER: 0.2281
• Validation loss: 65.32
• W&B run: x6zaxexu

Intended Use

This artifact is intended to be consumed by the P003 scoring pipeline through the nemo: backend contract. It is not an HF AutoModelForCTC checkpoint.

Example backend string:

nemo:Peacockery/citrinet-256-phoneme-en