Improving Readability for Automatic Speech Recognition Transcription

Modern Automatic Speech Recognition (ASR) systems can achieve high performance in terms of recognition accuracy. However, a perfectly accurate transcript still be challenging to read due grammatical errors, disfluency, and other noises common spoken communication. These readable issues introduced by speakers ASR will impair the downstream tasks understanding human readers. In this work, we present task called post-processing for readability (APR) formulate it as sequence-to-sequence text generation problem. The APR aims transform noisy output into humans while maintaining semantic meaning speakers. We further study from benchmark dataset, evaluation metrics, baseline models: First, address lack task-specific data, propose method construct dataset using data collected error correction. Second, utilize metrics adapted or borrowed similar evaluate model on task. Lastly, use several typical pre-trained models Furthermore, fine-tune constructed compare their with traditional pipeline proposed metrics. Experimental results show that all fine-tuned perform better than method, our RoBERTa outperforms 4.95 6.63 BLEU points two test sets, respectively. case reveal ability improve transcripts.