Flexible human-machine interfaces (HMIs) encounter several significant challenges, including intricate architectures, reliance on external power, and crosstalk. This study develops a flexible self-powered Keypad based on triboelectric nanogenerators (TENGs) and introduces an effective strategy to greatly reduce internal crosstalk based on separate cavity structure. A theoretical model is established to clarify the relationship among finger-tapping force, device's mechanical strain, and TENG output. Comparative experiments demonstrate that the crosstalk ratios in the upper, right, and upper-right units adjacent to the central unit are substantially reduced to 24%, 29%, and 15%, respectively. The flexible Keypad, as an HMI, exhibits superior response time (34 ms), outstanding durability (30 000 continuous tapping), and linear feedback in response to finger-tapping force (1–10 N). In Finger Tapping Test, the self-powered Keypad accurately detects subtle variations in finger-tapping patterns, both between individuals and across different fingers of the same individual, offering valuable reference for neuropsychological assessments. By integrating machine learning, the self-powered Keypad achieves high accuracy in individual identification (100%) and handwritten digit recognition (97%). A mixed-reality gaming control system based on the flexible self-powered Keypad is developed, enabling real-time precise control of game characters in virtual environment, thereby broadening the application prospects of self-powered HMIs.