A 0 → 1 Capstone Project
By Carl Kho · Minerva Class of 2026
← BACK TO PORTFOLIO
▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞

Jump! Did you feel it? The split-second tension in your calves? The silent command in your throat? Your brain screamed "MOVE", but your body stayed still. The command was loud. The movement was silent. That silence is data. And for my last year at university, I am building the interface to hear it.

Watch Prototype Demo
Phone sensor diagram
FIG_001: PHONE SENSOR REPURPOSING

We carry supercomputers with accelerometers in our pockets, yet we tap on glass like cavemen. In Phase 1, I turned a phone into a sword. I used raw IMU data to control a video game character, proving that "motion" is just a stream of numbers waiting to be interpreted.

Watch Wearable Demo
Smartwatch disassembly
FIG_002: SMARTWATCH INTEGRATION

But who holds a phone while gaming? I needed it wearable. I strapped the sensor to my wrist and taught a machine to know the difference between a punch and a wave. I built a "Press-to-Label" system to solve the data collection problem, proving that good ML starts with good data engineering.

Though... motion is crude. It suffers from the "friction between having a thought and the physical action required to input it." I wanted the spark. Phase 3 began as a hands-free cyclist turn signal, detecting the specific flex of the Flexor Digitorum Profundus leading to the pinky finger. I benchmarked 18 architectures on a "Ladder of Abstraction" (detailed in Technical Narrative), from simple variance thresholds to combining Inception blocks, Bi-LSTMs, and Multi-Head Attention.

The surprise? Deep Learning failed (49.6%). Random Forest won (74.3% at 0.01ms latency). This validated a core Ubicomp thesis: on the edge, feature engineering > raw compute.

Why move at all? Speak without sound. Phase 4 replicates the MIT Media Lab's AlterEgo (Arnav Kapur, Pattie Maes). By placing electrodes on the jaw, I capture Neuromuscular Signals—the "echoes" of internal speech on the skin. This is a Peripheral Nerve Interface, detecting the intent to speak before sound leaves your mouth. The insight? Muscles sound like hearts. I hacked a simple heart sensor (AD8232) to capture the subvocal frequency range (1.3-50Hz), democratizing the Silent Speech Interface with off-the-shelf components.

The ultimate interface is silence.

CP194 Capstone · Minerva University · Spring 2026

SOMACH

A $40 dual-channel sEMG system for silent speech classification.
2 studies · 3 arXiv papers · 4,033 CSVs · open-source.

Carl Vincent Ladres Kho · Advisor: Prof. Patrick Watson · kho@uni.minerva.edu


51.8% Study A CV accuracy
3.1× Above chance
93.5% Ensemble + gated
$40 Total hardware cost
▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞
WHO

Carl Kho — CS + Cognitive Neuroscience undergrad at Minerva University, studying across 7 countries. Hardware-software integrator building cognitive prosthetics.

WHAT

A $40 silent speech interface that replicates MIT Media Lab's $1,250 AlterEgo headset. Think typing with your jaw — no sound, no movement, just the electrical echoes of internal speech.

WHEN

Fall 2025 Capstone Project. One semester. Four phases. 23 technical blog posts documenting every failure and breakthrough.

WHERE

Developed across Taipei, San Francisco, and Hyderabad during Minerva's global rotation. Debugged hardware in hotel rooms, coffee shops, and 12-hour flights.

WHY

To prove that expensive research hardware can be democratized. If a cheap cardiac sensor can read muscle signals, then brain-computer interfaces don't need to cost thousands. The best interface is the one anyone can build.

PHASE 1 Phone Motion as input
PHASE 2 Watch Wearable IMU
PHASE 3 Muscle EMG signals
PHASE 4 Mind Silent speech
Recreate it yourself
The Build Log [0 → 1]
[ POSTS: 23 · PHASES: 4 ]

1. KINETIC (PHASE 1-2)

2. BIOSIGNAL (PHASE 3-4)

3. FOUNDATIONS

OpenEMG Data Visualization
FIG_01: RAW EMG SPECTROGRAM // 8-CHANNEL
COMMON QUESTIONS
  • Is this an actual product?
  • How does it work?
  • Why a heart sensor?
  • Can I build one myself?
  • What's next for the project?
  • How can I contact you?
IN: WHY A HEART SENSOR?
OUT:

The AD8232 is a $5 ECG sensor designed to capture heart signals in the 0.5–40Hz range.

Subvocal muscle signals (sEMG) happen to fall in the 1.3–50Hz range—almost the same frequency band. The heart sensor's built-in bandpass filter and amplifier work perfectly for jaw muscles.

This was validated by MIT's AlterEgo research: their custom hardware explicitly filters to this exact range to isolate subvocal signals from noise. A $5 heart sensor does the same filtering by accident.

It's not designed for this use case, but physics doesn't care about product labels. Muscles sound like hearts.

Work Journal (Live Pomodoro Feed)

Loading Work Logs...
SIGNUP FOR UPDATES:
I promise to only email you cool shit. Hardware breakthroughs, research updates,
sneak peeks at the build process. Stuff like that.
CURRENT PROGRESS:
██████████▓▓▓▒▒▒░░░░░░░░
▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞ ▞
×