launchiflytek2026-09-16
iFlyTek releases Spark-Audio-1.0-Preview, voice foundation model on domestic compute
iFlyTek's Spark-Audio-1.0-Preview pairs a 0.65B audio encoder with a 30B-A3B MoE, trained on 13M hours of audio across 99 languages/202 dialects on Ascend 910B + Hygon compute.
科大讯飞 9 月 16 日发布语音基座大模型 Spark-Audio-1.0-Preview，采用 0.65B 音频编码器搭配 30B-A3B MoE 架构，基于 1300 万小时音频数据训练，在昇腾 910B 与海光等全国产算力集群上完成。
模型摒弃"先转写后理解"的级联方案，直接理解语气、情绪与环境音，支持 99 种语言、202 种方言，在 Fleurs 中文测试集取得 SOTA。该模型面向车载、会议、教育、客服等需要"听得懂上下文"的多模态场景。
语音国产算力spark-audio
Related News
10-22
launch
iFlytek releases Spark X2.6 end-cloud unified model at $0.09 CNY-equivalent per million input tokens
10-22
launch
Replit launches Monitoring Agent, an autonomous operations agent that can autofix and publish
10-22
launch
Canva releases its first in-house image generation model Canva Craft at up to 2K resolution