Conformer-CTC fine-tuned on FluencyBank Timestamped

Results on test set

Training details

• Base model: nvidia/stt_en_conformer_ctc_small
• Dataset: arielcerdap/TimeStamped-Splits (train=2,744 / val=342 / test=342)
• Target: verbatim transcription (disfluencies preserved)
• Optimizer: AdamW (lr=1e-5, weight_decay=1e-3)
• Scheduler: CosineAnnealing (warmup_ratio=0.05)
• Batch effective: 64
• Early stopping: patience=10 on val_wer