iFlytek Releases Spark-ASR-2.0 Speech Recognition Model with Only 10% More Inference Cost

Chinese AI company iFlytek has officially released its next-generation speech recognition model, Spark-ASR-2.0. The model uses new techniques like non-autoregressive and LLM-enhanced autoregressive collaboration, joint enhancement of mixed Chinese-English text and acoustics, and dynamic context injection, achieving major improvements in general recognition, complex acoustic scenarios, and text fluency.

Official data shows that Spark-ASR-2.0's overall inference cost is only 10% higher than the previous generation. In scenarios involving dialects, high noise, and low volume, the model outperforms the current industry best, with a clear drop in word error rate.

Spark-ASR-2.0 will gradually roll out on iFlytek Input and provide API services through the iFlytek Open Platform. The model will also be applied in products including iFlytek AI Glasses, iFlytek Smart Office Book, and iFlytek Hearing.