iFlytek has officially launched its latest speech recognition model, Spark-ASR-2.0. The new version utilizes non-autoregressive and LLM-enhanced autoregressive coordination, alongside dynamic context injection, to improve recognition accuracy.
The model shows significant performance gains in handling mixed Chinese-English text, dialects, and complex acoustic environments such as high noise, low volume, and fast speech. It also focuses on generating more coherent and fluent text output compared to its predecessor.
Despite the performance improvements, the inference cost of Spark-ASR-2.0 is only 10 percent higher than that of the Spark-ASR-1.0 model. The company reports that the word error rate has been reduced across core recognition scenarios.
Starting tomorrow, the model will be integrated into the iFlytek input method and made available via the iFlytek open platform API. Future plans include deploying Spark-ASR-2.0 across various hardware products, including AI glasses, smart office notebooks, and transcription services.
