VisemIQ

A visual speech recognition experiment that tries to turn speech into text from silent video, by looking only at lip movements.

Question
Can speech be read from video alone, without any audio?
Status
Prototype
Started
September 2026
Tools
  • Visual speech recognition
  • Large language models
  • Python
  • Claude

The problem

Understanding what is being said in videos where the audio is missing or unusable. Lip reading is a limited and exhausting skill for people; I wanted to see how far AI can go with it.

What I built

The first version reads English speech from lip movements, with a Turkish interface and optional Turkish translation. I have also put in place the language infrastructure and automatic language detection for Turkish videos; Turkish lip reading itself is the next research phase.

Early observations

  • Lip reading is a much harder problem than speech recognition from audio: the same lip shape (a viseme) can correspond to several different sounds. That is where the project’s name comes from.
  • Passing the raw output through a language model to put it in context looks like the most promising way to make it readable.

All experiments

View all
Experiment 01Prototype

CourtScribe

The court’s scorekeeper; an AI system that watches game video and keeps the player-by-player statistics table itself.

Can player-level statistics be extracted from an amateur basketball game video without anyone keeping score by hand?

  • YOLOv8
  • ByteTrack
  • EasyOCR
  • faster-whisper
  • Google Colab

June 2026

Experiment 02Prototype

Angel Lens

A tool for angel investors that turns a start-up pitch deck into a structured analysis with the discipline of an institutional investment committee.

How close can an angel investor get to the discipline of an institutional fund’s investment memo, starting from a single pitch deck?

  • Claude API
  • React
  • PDF.js
  • Vercel

March 2026