iFlytek trains a speech model on domestic chips
On 22 October 2026 the iFlytek Spark Audio Team posted the technical report of Spark-Audio 1.0, a speech foundation model that couples a 0.65-billion-parameter audio encoder with a 30B-A3B mixture-of-experts language model. The report says it was trained on 13 million hours of audio, on Huawei Ascend 910B and Hygon BW1000 accelerators from encoder pretraining through post-training, and that it supports 99 languages and 202 dialects.
Why it matters
A company reports that a competitive speech model can be trained end to end on accelerators made in China instead of on the imported GPUs most such models use. The claim is the authors' and rests on their own tests, including in-house sets; the report compares itself with Qwen3.5-Omni-Flash and says it used about a tenth of that model's audio data.
What the report says. The model takes text and speech as input and generates text. It is built for speech recognition, speech translation, audio understanding and speech- or text-based dialogue. The audio encoder is pretrained with representation-enhanced BEST-RQ and refined by multi-task supervised training; the audio-LLM is trained in stages, then post-trained with supervised fine-tuning and on-policy distillation. The authors state "comparable performance" with Qwen3.5-Omni-Flash on most audio and text tasks, better multilingual recognition, spoken dialogue and application scenarios, and a performance that "approaches" larger closed-source models; they name instruction following and complex audio reasoning as weaker. The abstract and the introduction say the training ran on Huawei Ascend 910B and Hygon BW1000 chips "from audio encoder pretraining through audio-LLM pretraining to post-training". What the record does not claim. No weights, code or licence are announced in the text, which points to a demonstration page of iFlytek. The report gives no number of chips, no training time and no cost, so a reader cannot weigh how much the domestic hardware limited the work. The comparisons are the authors' own. The page read is the first version of the preprint (v1, 22 October 2026, 13:55 UTC); the PDF is dated 23 October on its first page.