utils.py

from datasets import load_metric

cer_metric = load_metric("cer")

def compute_metrics(pred):
    labels_ids = pred.label_ids
    pred_ids = pred.predictions

    pred_str = processor.batch_decode(pred_ids, skip_special_tokens=True)
    labels_ids[labels_ids == -100] = processor.tokenizer.pad_token_id
    label_str = processor.batch_decode(labels_ids, skip_special_tokens=True)

    cer = cer_metric.compute(predictions=pred_str, references=label_str)

    return {"cer": cer}