Kedaixinfly launches enhanced speech recognition model Spark-ASR-2.0 for input method and other products

On September 23, 2026, Kedaixinfly announced the release of its latest speech recognition model, Spark-ASR-2.0. The model incorporates key innovations such as non-autoregressive and LLM-enhanced autoregressive collaboration, Chinese-English mixed text and acoustic joint enhancement, and dynamic context injection, resulting in significant improvements in speech recognition performance. Notably, the model excels in general recognition, complex acoustic scenarios, contextual recognition, and text fluency. The inference cost increased by only 10% compared to Spark-ASR-1.0. According to Kedaixinfly, Spark-ASR-2.0 not only enhances accuracy but also produces more coherent and semantically clear text by refining redundant expressions and details. The model will be gradually integrated into the Kedaixinfly input method starting September 24, 2026, and will also be available via the Kedaixinfly Open Platform API. It will be deployed across multiple products, including Kedaixinfly AI Glasses, Smart Office Notebook, and Kedaixinfly Hear. Spark-ASR-2.0 outperforms industry benchmarks in dialect, high-noise, and low-volume environments, demonstrating strong capabilities in complex acoustic scenarios. The model’s performance in these areas is significantly better than the current industry best. Kedaixinfly emphasized that the new model represents a step forward from previous versions, which focused on accurately transcribing spoken content, to generating more natural and coherent text.