G-SPIN: Graph-based ASR correction cuts phonetic errors by combining GNNs and LLMs

A new paper from Singh, Zaki, Mukkamala, and Wasnik—accepted at ACL Industry Track 2026—introduces G-SPIN, a framework that corrects residual phonetic errors in automatic speech recognition (ASR) output without modifying the underlying ASR model. Current ASR systems, despite low overall word error rates, systematically misrecognize semantically critical tokens such as named entities, negations, and sentiment-bearing words due to phonetic similarity. G-SPIN addresses this by combining a graph neural network (GNN) that constructs a neighborhood of acoustically plausible alternatives for each flagged token, restricting the search space to phonetic variants. A masked language model then scores these candidates based on local context, and an instruction-tuned large language model performs a final context-aware reranking over this compact set.

By decoupling structured phonetic reasoning from contextual semantic selection, G-SPIN avoids the pitfalls of unconstrained generation while improving correction accuracy. The entire pipeline operates at inference time, is modular, and requires no retraining. The paper demonstrates that this graph-based approach significantly outperforms naive token-level corrections and fully open-ended LLM-based corrections, especially on domain-specific named entities. For professionals deploying ASR in production, G-SPIN offers a practical, plug-and-play solution to reduce high-cost errors in transcription pipelines, call center analytics, and voice assistants.