Multiple overlapping sound sources can also be precisely parsed! iFlytek unveils speech foundation model SparkAudio 1.0

What happened: iFlytek unveiled SparkAudio 1.0, a speech foundation model trained on domestic compute, at its 2026 1024 Developer Festival. It handles transcription, translation, dialect recognition, speaker separation, emotion analysis and complex audio question answering. Key numbers: MOS score for Spark-LiveTranslation rose Why it matters: iFlytek says SparkAudio 1.0 outperforms several SOTA closed-source models on published benchmarks, with instruction following on par. The model is already deployed across iFlytek input method, hearing, translator, glasses and office notebook products.

iFlytek unveiled SparkAudio 1.0, a speech foundation model trained on domestic compute, at its 2026 1024 Developer Festival. It handles transcription, translation, dialect recognition, speaker separation, emotion analysis and complex audio question answering.

iFlytek says SparkAudio 1.0 outperforms several SOTA closed-source models on published benchmarks, with instruction following on par. The model is already deployed across iFlytek input method, hearing, translator, glasses and office notebook products.