macDrum

Turn Your Macbook Into A Drum

One day I was scrolling on reels when I saw this reel:
Reel

While watching this reel I was tapping my table like Its a drum, and then I decided to make My laptop detect where I am tapping it, using the Inbuilt accelerometer that this reel's app slapmac uses, (I am guessing)

So, I reversed-engineered the concept and built macDrum—a tiny, low-latency, machine-learning-powered drum kit that turns different zones of your MacBook chassis into live MIDI/audio triggers.

How it Works

Instead of using the microphone or camera, macDrum intercepts raw physical vibrations traveling through the aluminum unibody of the Mac. Here is the engineering breakdown of the pipeline:

1. Telemetry & Data Ingestion

MacBooks have an internal IMU (Inertial Measurement Unit) to detect sudden drops (historically to park hard drive heads). Using the macimu bridge, this project hooks directly into Apple's internal Shared Memory (SHM) segment. This bypasses high-overhead OS layers to read raw, high-frequency linear acceleration across three spatial dimensions: $X$, $Y$, and $Z$.

2. Signal Thresholding & Edge Triggering

The system idle-polls gravity (resting at roughly $-1.0g$ on the Z-axis). When an impact occurs, the sudden shockwave causes an acute spike in amplitude. The script uses an asymmetric delta threshold mechanism to catch this spike. The moment $\Delta Z > 0.06g$, an edge-trigger fires, and the script instantly captures a rolling spatial-temporal window of the next 40 hardware clock ticks.

3. Spatial Matrix Flattening

A single slap yields a 2D matrix of shape (40, 3)—representing 40 sequential moments in time, each with 3 spatial axes. To feed this into a classical classifier, the matrix is programmatically linearized into a 1D Feature Vector of 120 dimensions:

$$[X_1, Y_1, Z_1, X_2, Y_2, Z_2, \dots, X_{40}, Y_{40}, Z_{40}]$$

4. Machine Learning Inference (The Brain).

The core classification is handled by a Supervised Random Forest Classifier (scikit-learn).

• Rather than relying on rigid mathematical formulas, the model uses an ensemble of 100 independent Decision Trees.
• During training, trees evaluate splits using Gini Impurity to discover exactly how a shockwave dissipates through the chassis.
• When you hit a zone, the 100 trees run parallel evaluations and hold a majority vote to instantly classify the vibration signature into one of five categories: top_left, top_right, bottom_left, bottom_right, or none.

---

Repo Blueprint

• collect_data.py: The data acquisition script. It prompts you for a label, registers the slaps, flattens the vectors, and appends them to the training set.
• train_model.py: The training engine. It ingests your raw data, splits it into an 80/20 train/test distribution, prints an evaluation report, and serializes the model.
• play_drums.py: The real-time live engine. It loads the compiled brain, continuously polls the IMU, runs real-time inference, and triggers local audio files.
• drum_data.csv: The local dataset containing your raw, flattened 120-feature vibration fingerprints.
• drum_model.pkl: The compiled, pickled binary representation of your trained Random Forest model.

---

Next Steps

• Virtual MIDI Loopback: Replace native afplay execution with a virtual MIDI port driver (mido), allowing macDrum to drive professional DAWs like Logic Pro, Ableton Live, or GarageBand as a physical hardware controller.
• CoreML Compilation: Export the trained Random Forest directly into Apple's native .mlmodel format to leverage the Apple Neural Engine (ANE) for hardware-accelerated inference.
• Dynamic Calibration Loop: Implement a 5-second automatic threshold calibration routine upon script startup to read ambient environment noise and adjust DELTA_THRESHOLD algorithmically.

---