New survey maps 5 approaches to fix ASR errors without retraining models

Automatic Speech Recognition (ASR) systems are essential for voice assistants, transcription services, and accessibility tools, but they still struggle with accents, dialects, background noise, and domain-specific terminology. Redesigning an ASR model is expensive and time-consuming, so researchers have turned to non-intrusive refinement—techniques that improve transcription accuracy without altering the core model. A new comprehensive survey on arXiv (v3, May 2026) by Mohammad Reza Peyghan and colleagues systematically reviews these methods, grouping them into five distinct classes: fusion, re-scoring, correction, distillation, and training adjustment.

Fusion combines outputs from multiple ASR systems to reduce errors; re-scoring uses additional language models or context to rank hypotheses; correction directly edits erroneous transcriptions; distillation transfers knowledge from a large model to a smaller one; and training adjustment fine-tunes external components (like a separate error corrector) while keeping the ASR model frozen. The survey outlines each method's strengths, limitations, and ideal application scenarios—for example, correction works well for domain-specific jargon, while re-scoring excels in noisy environments.

Beyond classification, the paper surveys adaptation techniques for domain-specific ASR refinement (e.g., medical or legal terminology), reviews commonly used evaluation datasets and their construction, and proposes a standardized set of metrics to enable fair comparisons across studies. The authors also identify open research gaps, such as handling code-switching and scaling refinement to real-time streaming applications. By providing this structured overview, the survey equips researchers and practitioners with a clear foundation to build more robust, accurate ASR pipelines without expensive model overhauls.