SparkAudio 1.0
iFLYTEK detailed SparkAudio 1.0, a speech foundation model with a 0.65B audio encoder and a 30B-A3B MoE language model, covering 99 languages and 202 dialects.
SparkAudio 1.0 puts machine translation, speech recognition and semantic understanding in one base model so the tasks can learn from each other. It also handles speaker separation, environmental sound understanding and emotion recognition. It was trained on 13 million hours of audio on domestic chips including Huawei Ascend 910B and Hygon BW1000. iFLYTEK launched Robot Super Brain Platform 3.0 alongside it, with a board, interaction module and development kit for robots.
- Date
- Monday 26 October 2026
- Lab
- iFLYTEK
- Kind
- model
- Access
- app only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Audio encoder size | 0.65B dense parameters paired with a 30B-A3B MoE language model | company |
| Training audio | 13 million hours plus large text corpora | company |
| Language coverage | 99 languages and 202 dialects speech recognition | company |
| Accuracy gain from semi-supervised audio decoder | 20% as stated by chairman Liu Qingfeng | company |
| Response time reduction from dynamic segmented end-to-end modeling | 15% | company |
| Cost reduction from dual-paradigm collaborative decoding | 60% | company |
Already live in iFLYTEK Input Method 16.0, iFLYTEK Hearing, smart notebooks and translators. The technical report says instruction following and complex audio reasoning still need work.
Sources
This record was checked against its sources on 26 October 2026. How we check