Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Listening, Imagining \& Refining: A Heuristic Optimized ASR Correction Framework with LLMs

Created by
  • Haebom

Author

Yutong Liu, Ziyue Zhang, Yongbin Yu, Xiangxiang Wang, Yuqing Cai, Nyima Tashi

Outline

This paper proposes LIR-ASR, an iterative error correction framework inspired by human auditory perception that leverages large-scale language models (LLMs). LIR-ASR generates phonetic variations using a "listen-imagine-refine" strategy and refines them based on context. To avoid local optima, heuristic optimization using a finite state machine (FSM) is employed, along with rule-based constraints to maintain semantic fidelity. Experimental results on English and Chinese ASR outputs demonstrate that LIR-ASR significantly improves transcription accuracy, reducing CER/WER by an average of 1.5 percentage points compared to baselines.

Takeaways, Limitations

•
Takeaways:
◦
We demonstrate that the accuracy of ASR systems can be improved by using an LLM-based iterative error correction framework.
◦
We propose that a “listen-imagine-refine” strategy inspired by human auditory perception is effective in correcting ASR errors.
◦
We demonstrate that FSM-based heuristic optimization and rule-based constraints can achieve performance improvements and semantic consistency.
◦
Experimental results for both English and Chinese suggest the generalizability of LIR-ASR.
•
Limitations:
◦
The performance improvements of the proposed method may be limited to specific datasets and models.
◦
The design of heuristic optimization and rule-based constraints may need to be tailored to specific languages or tasks.
◦
The computational cost and processing time of LLM may limit its practical application.
◦
Further research is needed on compatibility and scalability with various ASR systems.
👍
Made with Slashpage