iFlytek Xinghuo Speech Base Large Model Released: 0.65B Encoder with 30B MoE, Entirely Domestic Computing Power Trained

iFLYTEK announced the release of Spark-Audio-1.0-Preview, a speech foundation large model trained with fully domestic computing power.

Previously, processing audio with large models mostly used a cascading approach: first converting speech to text, then handing it over to the large model for understanding. This approach has two obvious shortcomings: one is information loss—text can only record what was said, but it loses key information such as tone, emotion, and background sounds in the voice, leading to incomplete scene judgment by the model; the second is fragmented workflow—capabilities such as speech transcription, speaker identification, and speech translation are split into independent modules that run in series, resulting in breakpoints between modules, which can accumulate errors and cause delays and lag in user experience.

No longer just transcription, but directly "understanding" speech

The speech foundation large model no longer just transcribes speech into text, but directly understands speech: it can comprehend human voices, ambient sounds, music, and also grasp the semantics, emotions, and context within the sound.

The initial release of Spark-Audio-1.0-Preview uses an audio encoder with 0.65B (Dense structure) and a large language model with 30B-A3B (MoE structure). It was trained using 13 million hours of audio and a large amount of text data, based on a fully domestic computing cluster. The same model can accept both text and audio modalities, supporting tasks such as speech transcription, multilingual translation, multi-dialect recognition, ambient sound recognition, speaker recognition, sentiment analysis, and complex audio Q&A, enabling machines to move from "hearing clearly" to "understanding." Similar to the "1+N" system of iFLYTEK's Spark large model, fine-tuning can also be performed on the speech foundation model to develop specialized models or systems for speech recognition, real-time speech translation, and speech interaction.

Supports 99 languages, 202 dialects, with significant advantages in complex scenarios

According to the official statement, Spark-Audio-1.0-Preview performs overall comparably and exceeds in some areas in various speech evaluation tasks, especially in challenging scenarios such as high noise levels and speaking softly—tasks that are more realistic in application. Compared to larger closed-source speech foundation models, its performance is also very close.

This model supports the recognition of 99 languages and 202 dialects. In multiple tasks, it shows an advantage in average performance for multilingual and multi-dialect speech recognition compared to Qwen3.5-omni-flash (35B-A3B) of the same size, and is comparable to Qwen3.5-omni-plus, which is an order of magnitude larger in size. In evaluations such as Fleurs, Kespeech, and LibriSpeech for Chinese-English, multilingual, and multi-dialect speech recognition, it scores higher than Gemini-3.1 Pro and achieves SOTA in the Fleurs Chinese test set.

iFLYTEK also admitted that this version does not show a significant drop in general knowledge, math, and code tasks, but still has room for improvement in tasks such as instruction following, dialogue, and audio understanding. In the next step, it will consolidate its strengths while addressing its weaknesses. Currently, Spark-Audio-1.0-Preview is officially open for experience, and the API will be launched on the iFLYTEK Open Platform later.