EN / 中文

Brain-Computer Interface for Speech Restoration: West Lake Lingxi Unveils Progress on Chinese Language Neuroprosthetics

by xiangzhi·November 18, 2026

Human-computer interaction has evolved over the past century. Initially, programmers punched instructions onto paper tapes to feed machines. Later, keyboards emerged, followed by touchscreens that made interaction more intuitive. Today, people interact with devices through voice. At the 5th Nandujiang Smart Medical and Rehabilitation Industry Summit Forum held on November 13, Yang Jie, Co-founder of West Lake Lingxi Technology (Beijing) Co., Ltd. and Researcher at Westlake University, shared the company's recent progress in brain-computer interface (BCI) language neuroprosthetic technology.

"The next revolution is brain-computer interfaces," said Yang Jie. The evolution of interaction methods aims to make machine syntax increasingly aligned with natural human expression.

While technical visions are ambitious, reality remains challenging. Yang Jie acknowledged that the BCI technology chain has not yet converged. Signal acquisition involves multiple pathways—invasive, semi-invasive, and non-invasive—while core chips and electronic systems remain immature. Additionally, direct machine-neural interaction raises safety and legal regulatory concerns. The industry has progressed through trials and errors, ultimately converging on a practical path: implementation as medical devices in clinical settings. This explains the growing connection between BCIs and smart medical forums.

The academic community has spent two decades exploring motor decoding, establishing a laboratory technology chain of "signal acquisition—decoding—stimulation feedback." Consequently, motor function reconstruction has become the most prominent focus in BCI research. Language decoding started later, with key breakthroughs concentrated after 2020. However, rapid progress is being made both domestically and internationally. Westlake Lingxi chose to enter the field through this point.

Why focus on language? The demand is evident: approximately 30% of stroke patients cannot regain language function, meaning language neuroprosthetics address a vast medical market. Meanwhile, medical scenarios allow for the continuous accumulation of high-quality data, laying the foundation for future novel human-computer interactions. "Whether aiming for the ultimate goal in healthcare or in interaction, infrastructure is indispensable," said Yang Jie. Lingxi's strategy is to build BCI infrastructure.

The first component of infrastructure is chips. Westlake Lingxi's product line covers three areas: neural signal acquisition, neural signal stimulation, and neuromorphic computing. In 2026, the team will continue iterating on 256-channel and 1024-channel solutions. Notably, the 1024-channel sampling chip has been published at ISSCC (International Solid-State Circuits Conference). Regarding neural stimulation, current efforts involve dual-channel acquisition, with integrated closed-loop systems planned for next year. A highlight is the neuromorphic computing chip—the team believes the essence of BCI signals is neural pulse encoding. Using neuromorphic approaches allows for direct acquisition of neuronal spike discharges, achieving low-power, high-efficiency information transmission and processing. Related work has continued since 2021, resulting in a series of papers in top international journals. This year, the team developed a language decoding chip based on a neuromorphic architecture, with related reports expected to be released within the next one to two months.

The second component is invasive data, currently the scarcest resource in the entire industry. Westlake Lingxi, in collaboration with Westlake University, has built an invasive BCI Chinese language database. It includes 20 subjects, all using high-density electrodes to collect invasive EEG signals covering language-related brain regions. The total task duration is 50 hours, comprising 14,000 characters and 4,200 sentences. It includes both single-character tasks and continuous sentence reading paradigms, accompanied by high-density actual audio recordings, as well as annotations for start/end times, Pinyin, tones, phonemes, and electrode channel coordinates. "This is a very comprehensive database," introduced Yang Jie. The data continues to expand and will be fully open-sourced to society in 2027.

The third component is the Chinese language decoding framework. In early results, the team utilized tonal information for language decoding, achieving a character accuracy rate of 71%. The system displays sentences on the screen through single-character decoding combined with language model error correction, with an overall processing latency of less than 20 milliseconds. The goal for 2027 is to achieve completely free conversation without vocabulary limitations.

Yang Jie also provided an industry timeline: In the next one to two years, signal acquisition technology will mature—the barrier is not as high as externally perceived, with only one or two foreign entities currently capable; in the next three to four years, the industry, especially application vendors, will focus on chip transmission technology; five to six years from now, once acquisition and transmission issues are resolved, the focus will shift to decoding massive amounts of data. He predicts that BCI chips will eventually evolve into mature ASICs or SoCs (System on Chip). "Our product layout follows this trend," he stated.

Yang Jie emphasized that the logic behind the company's BCI infrastructure is simultaneous advancement across three layers: chips, data, and algorithms. By 2027, they aim to launch three mass-produced chips, open-source the Chinese language database, and advance real-time continuous decoding. "The continuous evolution and integration of infrastructure are driving language decoding toward real-time, continuous, and natural interaction—addressing clinical needs while serving as a key direction for the next generation of human-computer interaction."