Graph-Based Phonetic Error Correction of Noisy ASR

Short summary

Automatic speech recognition systems frequently misrecognize critical tokens due to structured phonetic errors rather than random noise. G-SPIN addresses this through a modular framework combining graph neural networks for phonetically plausible candidates, masked language models for context, and instruction-tuned LLMs for semantic re-ranking. The approach improves accuracy without unconstrained generation and operates entirely at inference time.

• •Proposes G-SPIN framework to fix structured phonetic errors in ASR that affect named entities, negations, and sentiment-bearing words
• •Combines GNNs for candidate generation, MLMs for contextual scoring, and instruction-tuned LLMs for final re-ranking
• •Lightweight, modular approach that avoids unconstrained generation and operates at inference time for practical deployment

Generated with AI, which can make mistakes.