iFlytek Launches Spark-Audio-1.0-Preview End-to-End Audio Model Trained on Domestic Compute
iFlytek has released Spark-Audio-1.0-Preview, an end-to-end audio foundation model trained entirely on domestic Chinese compute hardware. Combining a 0.65B dense audio encoder with a 30B Mixture-of-Experts (MoE) language model, it directly understands speech, emotion, tone, and ambient acoustics without relying on text transcription cascades. This development demonstrates China's growing capability to train competitive multimodal AI models using native computing infrastructure amid export controls. By removing cascaded text conversion, native audio models eliminate processing latency and acoustic information loss, setting a higher benchmark for real-time speech interaction. Trained on 13 million hours of audio and large-scale text datasets, Spark-Audio-1.0-Preview supports speech recognition across 99 languages and 202 dialects. Benchmarks show it achieved state-of-the-art results on the Fleurs Chinese test set and outperformed Gemini-3.1 Pro on multi-language speech recognition tasks, showing strong resilience in high-noise and low-volume environments.
## BACKGROUND
Traditional voice processing uses a cascaded architecture where spoken audio is first converted to text via automated speech recognition before being sent to an LLM, causing latency and losing rich non-verbal context like emotional tone or background audio. Additionally, Mixture-of-Experts (MoE) models like the 30B-A3B architecture route inputs to specialized sub-networks, activating only a small subset of total parameters (e.g., 3B per token) to optimize compute efficiency.