iFLYTEK SparkAudio 1.0 speech model supports 99 languages and 202 dialects, pairs 0.65B encoder with 30B-A3B MoE
iFLYTEK detailed SparkAudio 1.0 at its 2026 Global 1024 Developer Festival. The speech foundation model combines a 0.65-billion-parameter audio encoder with a 30B-A3B mixture-of-experts language model, trained on 13 million hours of audio. iFLYTEK also launched Robot Super Brain Platform 3.0 for embodied robots.
iFLYTEK has detailed SparkAudio 1.0, a speech foundation model that combines a 0.65-billion-parameter dense audio encoder with a 30B-A3B mixture-of-experts language model and supports speech recognition in 99 languages and 202 dialects. The company introduced it at its 2026 Global 1024 Developer Festival alongside Robot Super Brain Platform 3.0, according to Sina Tech, IT Home and a transcript of chairman Liu Qingfeng's keynote that iFLYTEK posted on Tencent News, as summarized by Pandaily.
SparkAudio 1.0: specifications
- Architecture: 0.65B dense audio encoder plus a 30B-A3B mixture-of-experts language model
- Training data: 13 million hours of audio, plus large text corpora
- Language coverage: speech recognition in 99 languages and 202 dialects
- Training hardware: domestic accelerators, including Huawei Ascend 910B and Hygon BW1000 chips
- Context window, pricing, API availability: pricing not yet disclosed; no context window stated
SparkAudio 1.0 brings machine translation, speech recognition and semantic understanding into a single base model, so the tasks can learn from each other. Beyond recognition, simultaneous interpretation and voice interaction, iFLYTEK says it handles real-time speaker separation, environmental sound understanding, audio understanding and emotion recognition. As an example, the company says the model can pick out horns, vendors and tire noise on a noisy street, separate overlapping voices and infer the surroundings from sound.
Claimed technical gains
Liu named three technical changes. All figures are iFLYTEK's own:
- A semi-supervised audio decoder that raises accuracy by 20%
- Dynamic segmented end-to-end modeling that cuts response time by 15%
- Dual-paradigm collaborative decoding that lowers cost by 60%
The technical report states that instruction following and complex audio reasoning still need work. No benchmark scores were reported.
Deployment
SparkAudio 1.0 is already live in iFLYTEK Input Method 16.0, iFLYTEK Hearing, and the company's smart notebooks and translators. iFLYTEK also showed an AI recording ring that attaches magnetically to the back of a phone and can transcribe two hours of audio in two minutes.
Robot Super Brain Platform 3.0
For robots, iFLYTEK introduced Spark embodied multimodal interaction technology. It targets:
- Dynamic noise reduction in complex spaces
- Recognizing which of several people is addressing the robot
- Planning speech and motion together, such as waving while saying goodbye
- Starting conversations when the camera spots something relevant
Robot Super Brain Platform 3.0 packages this work. The hardware side includes the Spark super brain board, a multimodal interaction module and an embodied AI development kit. The software side includes an embodied multimodal interaction harness that coordinates device and cloud, and front-end and back-end tasks, so a robot can keep chatting while it carries out jobs in the background. The platform supports modular, configurable customization for different industries. iFLYTEK demonstrated it in an unstaffed, 24-hour robot-run store at the event.
What this means
SparkAudio 1.0 is built as one model that covers recognition, translation, semantic understanding and audio scene analysis, rather than separate systems for each task. The 30B-A3B design is the architecture behind that, with the 0.65B encoder feeding the language model.
The hardware disclosure matters for anyone tracking compute supply: iFLYTEK says the model was trained on Huawei Ascend 910B and Hygon BW1000 chips. The accuracy, latency and cost improvements are company claims without published benchmark comparisons, so independent evaluation is needed before they can be treated as established. The report's own caveat about instruction following and complex audio reasoning marks where the model is still limited.
The robot platform extends the same speech stack into hardware and software for devices, with the unstaffed store as the demonstrated use case.
Related Articles
Primotion's Paimon-0 cuts action-chunk inference from about 3,000 ms to 300 ms, company says
Primotion, a Shenzhen startup founded in July 2026 to build embodied-AI chips, has unveiled Paimon-0, an embodied foundation model. The company says it cuts single action-chunk inference from about 3,000 ms to 300 ms and ranks first on several public robot benchmarks. All results are self-reported.
iFLYTEK launches Agent OS, APA and XFYUN CLI, putting 700-plus AI capabilities behind one command
At its 2026 Global 1024 Developer Festival in Beijing on October 22, iFLYTEK's open platform introduced Agent OS, APA, the XFYUN CLI and Xingchen TaaS (Token as a Solution). The XFYUN CLI lets one command call more than 700 atomic AI capabilities and more than 200 agent adaptation capabilities. All performance figures are company-reported.
iFLYTEK Healthcare Upgrades Spark Medical Model to V3.7, Claims 99.6 on MedBench 5.0
iFLYTEK Healthcare announced on October 22 a full upgrade of its Spark medical large model to V3.7, and reports a score of 99.6 on MedBench 5.0. The company says the model now combines speech, medical imaging and language understanding for clinical agent use. The benchmark and usage figures are company-reported.
DeepSeek V4.1 Flash 1017: $0.15/$0.6 per 1M tokens, 1M context, 67% hallucination rate on non-answers
DeepSeek has upgraded its open-weight V4.1 Flash to revision 1017, a 552B-parameter MoE model with 16B active parameters and a 1 million token context window. Off-peak API pricing is $0.15 per 1M input tokens and $0.6 per 1M output tokens. Artificial Analysis data shows strong intelligence scores but weak results on hallucination.
Comments
Loading...