iFLYTEK details SparkAudio 1.0, a unified speech foundation model for 99 languages trained on Huawei and Hygon chips, and launches Robot Super Brain Platform 3.0
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 26 October 2026
Your AI and this story
- GPT-6.2 Astra148 days after its cutoff
- Claude Opus 5.5118 days after its cutoff
- Gemini 4 Hydrogen209 days after its cutoff
- Grok 4.7148 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 118 days before it.
Key facts
- SparkAudio 1.0 architecture (technical report, via Pandaily): 0.65B-parameter dense audio encoder + 30B-A3B mixture-of-experts language model; trained on 13M hours of audio plus text
- Coverage: speech recognition in 99 languages and 202 dialects; also real-time speaker separation, environmental-sound and audio understanding, emotion recognition
- Compute: trained on domestic accelerators including Huawei Ascend 910B and Hygon BW1000
- Chairman Liu Qingfeng’s claimed gains: semi-supervised audio decoder +20% accuracy; dynamic segmented end-to-end modelling −15% response time; dual-paradigm collaborative decoding −60% cost (company figures; the report says instruction following and complex audio reasoning still need work)
- Shipping in iFLYTEK Input Method 16.0, iFLYTEK Hearing, smart notebooks and translators; an AI recording ring said to transcribe two hours of audio in two minutes
- Robot Super Brain Platform 3.0: Spark super-brain board, multimodal interaction module, embodied AI dev kit, and a device–cloud interaction harness so a robot can keep talking while doing tasks; demoed in an unstaffed 24-hour robot-run store
- Date: Pandaily’s report is dated Oct 26, 2026; the exact keynote day was not confirmed
What happened
iFLYTEK, China’s largest speech-technology company, used its annual developer festival to lay out a path from voice interaction to multimodal and then embodied interaction for robots. Pandaily drew on Sina Tech, IT Home and the transcript of Liu Qingfeng’s keynote. We did not find the SparkAudio technical report itself; the architecture numbers come from Pandaily’s reading of it.
Why it matters
A large speech model trained entirely on Chinese accelerators (Ascend and Hygon) is one more data point that export controls have not stopped domestic training of production models, at least at this scale.
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 press
Press
Changes
- Filed. Technical report not located