Low-Latency Spell Correction for Japanese Music Search Queries

The researchers proposed a compact sequence-to-sequence model based on the BART model, aimed at addressing the issue of low-latency spelling correction in Japanese music search. The model consists of three encoder layers and three decoder layers, with its core innovation being the construction of a script-aware synthetic spelling generation pipeline. This pipeline combines keyboard layout models, speech confusion priors, and specific error patterns to generate real training data. Experimental results show that on the curated evaluation set, the model achieved an accuracy of 41.09% in accurate matches and a character error rate of 11.62%, outperforming baseline methods based on edit distances and achieving the lowest character error rate. Additionally, the model’s inference latency under a single GPU environment was less than 4ms, effectively meeting the requirements for low-latency performance.