Sequence-to-Sequence-ASR-Error-Correction

Model Description

Pre-Training Settings:

166k samples from Common Voice 13.0 was recognized by Whisper tiny.en.

1,000 random samples was selected as the test set, and the rest for training and validation with an 80%-20% split

• Batch size: 256

• Initial learning rate: 1e-5

• Adam optimizer

• 30 epochs

• Cross-entropy loss

• Best checkpoint saved based on WER as the evaluation metric

• Decoding is performed using beam search with a size of 5

• S2S backbone model adopted from ''Exploring data augmentation for code generation tasks''.

Continue-Training Setting:

• 2 epochs for gold-gold to prevent the over-correction problem on ''Ted talk data''