Project C: Silent speech
Surface EMG on the jaw, cheek, and throat while a person mouths words without sound. Decode a small vocabulary. A hot research direction with hobby-accessible hardware, and a communication aid for people who have lost their voice but not their articulators.
You are skimming: the title, the first figure, and the short version. Switch to Read in the header for the full page, or Deep to open every deep dive.
Speech is muscle. Even when you mouth a word without making a sound, the tongue, lips, jaw, and larynx move, and surface electrodes over them pick up the EMGElectromyography (EMG)Recording the electrical activity of muscles, hundreds of microvolts to millivolts, much larger than EEG. Glossary entry. A classifier can learn to tell “yes” from “no” from “help” from “water” from those patterns. This is a silent speech interfaceSilent speech interfaceDecoding intended words from muscle activity of the face and throat while the person mouths or subvocalizes without sound. Glossary entry: what a laryngectomy patient could use, what a person in a loud or quiet place could use, and what several companies are now building. With eight channels and the armband’s pipeline, a ten-word vocabulary is a month’s project.
Which is harder for a silent speech decoder: telling 'yes' from 'no', or telling 'pat' from 'bat'?
Pat versus bat. They use the same lip, jaw, and tongue movements; the only difference is voicing, a laryngeal event that is weak on the skin surface. Yes and no differ in almost every articulator. This is why silent speech systems start with vocabularies of distinct words, and why the research frontier is phoneme-level decoding with many more electrodes.
Beyond the amplifier
| Part | Where | Qty | Approx. |
|---|---|---|---|
| Disposable snap ECG electrodes, pack of 50, and snap leads For EMG and as a quick ground/reference. Snap leads with alligator or 3.5 mm ends. | Amazon | 1 | $12 |
| ESP32 DevKit (or Raspberry Pi Pico) For streaming from the ADS1299 board over USB later. Either board works; pick one ecosystem and learn it. | Adafruit, SparkFun, Amazon | 1 | $8 |
| Total (prices drift; treat as a ceiling) | $20 | ||
Electrode placement
Bipolar pairs at: the masseter (jaw, in front of the ear), the orbicularis oris (around the lips, above and below), the mentalis and digastric (under the chin, for the tongue), the zygomaticus (cheek), and the throat over the larynx. Eight channels, reference on the mastoid, bias on the collarbone. Photograph the placement and mark it; cross-session shift is worse on the face than on the arm because the skin moves.
Data
Ten words that matter to a user: yes, no, help, water, pain, stop, more, hot, cold, thank you. Twenty repetitions each per session, mouthed silently, with a marker at the start of each. Three sessions on different days. Add a “rest” class. Do not skip the rest class; a system that fires on chewing is unusable.
Features and classifier
Windows of 100 ms with 50 ms hop over the one-second utterance, the same features as the armband (RMS, waveform length, zero crossings, plus a few spectral bins), concatenated across the utterance into one vector, or fed to a small recurrent network if you have the data. LDA or a Riemannian classifier first. Grouped cross-validation by session, always.
Evaluation
Accuracy on the eleven classes by held-out session. Confusion matrix: which words confuse (hot and cold will). Then the real test: a live session where the person mouths random words from the list and the system speaks its guess aloud, and you count. Then a session where they mouth words not in the list and you count false positives.
Where this goes
More electrodes and finer placement move toward phonemes. Combining with a language model turns phoneme guesses into words. Companies (and a well-known 2023 paper) reach conversational rates with surface EMG on the face using dozens of channels and deep models trained on hours per person. Your ten words are the first rung.
- Place eight channels as described on yourself. Verify each channel responds to the intended articulator.
- Record three sessions of the eleven classes with markers.
- Extract features, train, evaluate by held-out session. Plot the confusion matrix.
- Run live with spoken feedback. Count hits, misses, and false positives on out-of-vocabulary words.
- Recruit one other person under your protocol and repeat. Report both.