Low-Latency Hand-Gesture Recognition for Real-Time Virtual MIDI Control Using MediaPipe and OpenCV on Consumer Apple Silicon

Physical Musical Instrument Digital Interface (MIDI) controllers frequently present cost and portability barriers for independent musicians and digital audio developers. While camera-based gesture control offers a hardware-free alternative, computer vision pipelines often suffer from high inference latency and frame-rate instability, rendering them impractical for real-time musical performance where sub-50ms responsiveness is critical. This paper presents the architecture and performance evaluation of AirMIDI, a lightweight, real-time virtual MIDI controller and audio modulation engine built in Python using OpenCV and Google's MediaPipe Hands framework. Running exclusively on consumer-grade Apple M1 hardware without dedicated external GPUs or depth-sensing cameras, the system captures 21-point 3D hand landmarks from a standard 720p webcam feed and maps five distinct spatial gestures—Pinch, Slide, Raise, Glide, and Tilt—to continuous MIDI Control Change (CC), Pitchbend, and Note On/Off messages. Experimental evaluation demonstrates that the proposed pipeline achieves an average throughput of 44.6 frames per second (FPS) and a mean end-to-end processing latency of 22.4 ms under Single-Hand Lite configuration. Furthermore, the implementation of an Exponential Moving Average (EMA) spatial filter reduces unintended MIDI jitter by 94.8% during static hand holds. These findings confirm that sub-35ms vision-based multi-parameter musical expression is viable on accessible consumer hardware.