Interview with GIAC AI Forum Lecturer Long Mingkang on Building AI Platforms, Speech Recognition Challenges, and Future AI Trends

From June 1‑2, the GIAC Global Internet Architecture Conference was held in Shenzhen, featuring experts from leading Chinese tech companies. In the lead‑up to the event, High‑Availability Architecture interviewed Long Mingkang, the lecturer of the GIAC AI Forum.

Long joined iFlytek in 2011, helped launch the speech cloud from zero to billions of daily page views, and led the development of AIUI (human‑computer interaction) and AIoT platforms. He emphasizes high‑performance, highly‑available components and a deep understanding of AI, HCI, instant messaging, and IoT.

He describes his role in moving AI research results onto a platform that lowers the barrier for developers, now supporting over 800,000 developers—the largest AI platform in China. He also built a C10M push system (XPush) and led the AIUI project, which defined a new generation of intelligent HCI standards.

When asked about speech‑recognition difficulties, Long explains that while ideal‑environment accuracy is high, real‑world noise, accents, far‑field capture, and domain‑specific vocabularies (e.g., medical, legal) remain challenging. iFlytek uses a proprietary CNN‑based acoustic model and Encoder‑Decoder approaches, and learned that offline optimization can mislead performance estimates, prompting a shift to online data‑driven iteration.

Regarding the motivation for building a private AI platform, Long notes that early AI services were B2B private‑cloud offerings. Growing demand and the rise of mobile made platformization and standardization essential, leading to the first generation AI platform integrated with iFlytek’s input method.

He stresses that the platform aims to reduce developers’ AI adoption cost, offering a wide range of services from speech to image, and provides a ready‑made solution (MoFei) for hardware integration.

Long explains basic AI terminology: AI as the overarching field, machine learning as its sub‑field for learning, deep learning as a further sub‑field, and common open‑source frameworks such as TensorFlow and MXNet. He compares them, noting TensorFlow’s comprehensive ecosystem and ease of use versus MXNet’s superior performance and memory efficiency, though MXNet focuses on training.

The interview also covers industry AI directions—autonomous driving, speech recognition, intelligent客服, robotics—and their distinct technical challenges, especially the need for knowledge graphs and context in Chinese semantic understanding.

On AutoML, Long describes Google’s project that uses neural networks to design neural networks, lowering the barrier for model selection and hyper‑parameter tuning. While AutoML can automate parts of the workflow, expert AI engineers remain essential for novel architecture design.

He reflects on the current AI landscape, identifying the third wave driven by big data and deep learning, and points out talent shortage as the main bottleneck. For engineers wanting to enter AI, he recommends understanding the AI pipeline (research, training, engine engineering, service), gaining distributed systems experience, and mastering fundamentals such as probability, graph theory, and optimization.

Finally, Long previews his GIAC talk, which will cover the evolution of iFlytek’s AI platform from zero to one, the architecture of the AIUI human‑computer interaction platform, and insights into AI link industry challenges.