iFlytek Releases Spark-Audio-1.0-Preview Voice Foundation Model

iFlytek has announced the launch of Spark-Audio-1.0-Preview, a voice foundation model trained entirely on Chinese domestic computing power. The model uses a 0.65B audio encoder paired with a 30B-A3B large language model, trained on 13 million hours of audio data, and can process both text and speech inputs.

Unlike traditional cascaded systems, this model understands speech directly, recognizing environmental sounds, music, semantics, emotions, and scenes all at once. It supports speech transcription, multilingual translation, multi-dialect recognition, speaker identification, emotion analysis, and complex audio Q&A.

The model supports 99 languages and 202 dialects. It outperformed Gemini-3.1 Pro in Chinese, multilingual, and multi-dialect benchmarks, achieving SOTA on the Fleurs Chinese test set. It also shows clear advantages in noisy and low-volume real-world scenarios.

The model is now available for testing, with API access coming to the iFlytek Open Platform.