Reproduce “Correcting Automated and Manual Speech Transcription Errors using Warped Language Models” on OpenResearch

About this paper

Mahdi Namazifar, John Malik, Li Erran Li, Gokhan Tur, Dilek Hakkani Tür

Amazon Alexa AI developed WLM Sentence Correction (WLM-SC), an adaptation of Warped Language Models, to explicitly correct errors in both automated speech recognition output and human transcriptions. The system achieved a 4.6% relative word error rate reduction for ASR output and a 10.6% relative reduction for human transcripts by integrating ASR n-best hypotheses.

Key insights

• Adapting Warped Language Models (WLMs) with an explicit operation prediction head transforms them from noise-robust understanding models into effective correction systems.
• Incorporating ASR n-best hypotheses provides valuable acoustic and linguistic context, proving particularly beneficial for correcting human transcriptions.
• The quality of ASR hypotheses directly impacts the effectiveness of human transcription correction, with high-confidence ASR outputs yielding the largest gains.

Summary by alphaXiv.