SilentSpeak reads facial and jaw muscle activity through four Grove EMG sensors mounted in headphone earmuffs. When you mouth words without sound, an ESP32-S3 streams 4-channel EMG data to a laptop, a 1D CNN classifies 3-second windows into one of 10 commands, and a companion app displays the result with text-to-speech.
Shipped a public demo so reviewers can actually try SilentSpeak without the hardware.Built a static page at luffy101-hash.github.io/silent-speak-emg with two things working end-to-end:
Live EMG scope: simulated 4-channel waveform (masseter + temporalis, left/right) animating at 1000 HzCommand picker, click any of the 10 silent commands, watch the classifier “detect” it with a confidence score, then hit Speak for browser TTS
Spent today getting the full SilentSpeak pipeline ready before the Grove EMG boards arrive.The firmware already streams 4-channel data at 1000 Hz over USB (timestamp_ms,ch0,ch1,ch2,ch3). On the Python side I finished the preprocessing pipeline (20–450 Hz bandpass), the 1D CNN classifier, and a Streamlit app that shows live predictions.
