SPEECHMAX

Speech therapy is boring and people drop off. 77% of people have public speaking anxiety but professional coaching costs $200/hour. Only 8% ever seek help. Existing tools are clinical and feel like homework, not something people stick with.

Built a Duolingo-style progression system with real clinical speech therapy methods. Real-time webcam analysis using MediaPipe for eye contact and posture, Web Speech API for transcription, and Web Audio API for pitch detection. All processing runs client-side for privacy. Designed five targeted mini-games that each isolate a specific weakness.

SpeechMAX was built in 48 hours at UNIHACK 2026, a national university hackathon (183 projects, 800 participants). We placed 5th overall. The premise: 77% of people have public speaking anxiety, professional coaching costs $200/hour, and only 8% ever seek help. We built the speech coach that everyone can access for free.

The app runs entirely in the browser. Open it, grant camera and mic access, and start talking. Within seconds you get real-time feedback across five dimensions: voice clarity (filler words, WPM, pitch variation), eye contact quality (468 facial landmarks + iris tracking via MediaPipe), and body language (posture alignment, fidget detection via 33 body keypoints).

After a 30-second radar scan that scores you across all five axes, SpeechMAX recommends targeted mini-games to train your weakest areas. Filler Ninja catches your "ums" and "likes". Eye Lock trains sustained camera eye contact. Pace Racer keeps your WPM in the target zone. Pitch Surfer rewards vocal variety. Stage Presence coaches open body language and power zone gestures.

Mike, the AI speech coach powered by Gemini 2.5 Flash, sees your scores, game history, and progress and gives short, actionable coaching advice. All data syncs to Supabase with Google OAuth or anonymous guest mode.

Frontend is React 18 with TypeScript (strict mode), Vite, and Tailwind CSS 4. State management uses Zustand 5 with localStorage persistence and Supabase sync. Framer Motion handles all animations.

The analysis pipeline is entirely client-side. MediaPipe FaceLandmarker (WASM, GPU-accelerated) tracks 468 facial landmarks for eye contact quality using a 3-signal fusion approach: iris position, head pose estimation, and gaze direction. MediaPipe PoseLandmarker tracks 33 body keypoints for posture alignment and fidget detection.

Audio processing runs through Web Speech API for real-time transcription with noise suppression, through a DynamicsCompressorNode (threshold -40dB, 4:1 ratio), into an AnalyserNode for FFT pitch detection via autocorrelation. Filler word detection uses count-based pattern matching against a configurable dictionary.

Backend is Supabase: anonymous + Google OAuth authentication, PostgreSQL with RLS for profiles/scan_results/game_results tables, and a gemini-proxy Edge Function that routes Gemini API calls through the server so the API key never touches the client.