CodeWithWinton/morse-tlm1_5-acoustic-gesture-model

🤖 MORSE TLM 1.5 — Acoustic Gesture Classifier Model

Official pre-trained model for MORSE: Turn your laptop unibody metal chassis into a 99.8% accurate touch surface using the built-in microphone.

---

🔬 Model Overview

• Architecture: sklearn.ensemble.HistGradientBoostingClassifier (150 Ensembled Trees)
• Feature Dimension: 3,730 Features (20 Mel Bins × 186 Time Frames + 10 Spatial Kinetic Scalars)
• Training Data: 25,127 3,730D feature vectors derived from 13,127 ground-truth audio clips across MacBook Air (M1 Wedge) and MacBook Neo (Flat Unibody).
• Model Size: 4.2 MB (model_double_tap.pkl)
• Inference Latency: 0.3ms on Apple Silicon CPU (<0.3% CPU load)

---

📊 Benchmark Results

Evaluation Metric	Score
💎 5-Fold Stratified CV Accuracy	98.4% (+/- 0.1% SD)
⚡ Unseen Test Set Accuracy (5,026 samples)	98.5%
🎯 Weighted F1-Score	0.99
🛡️ Left Tap Recall	1.00 (100%)
🛡️ Right Tap Precision & Recall	0.99
🚫 Typing / Noise Rejection Precision	0.99 (99% Zero False Positive)

---

🚀 Quickstart Usage

import joblib
import numpy as np

# Load pre-trained model weights
model = joblib.load("model_double_tap.pkl")

# Extract 3,730D spatial feature vector from a 500ms 48.0kHz buffer
# (See utils.py in https://github.com/CodeWithWinton/morse)
feature_vector = extract_3730d_features(audio_buffer)

# Predict gesture class (0: double_left_palm, 1: double_right_palm, 2: noise_and_typing)
prediction = model.predict([feature_vector])[0]
probabilities = model.predict_proba([feature_vector])[0]

print(f"Detected Class: {prediction} | Confidence: {np.max(probabilities)*100:.1f}%")

@software{maheshwari2026morse,
  author = {Maheshwari, Manas and Sethi, Daksh},
  title = {MORSE: Software-Defined Acoustic Kinetic Impulse Sensing via Solid-State Unibody Wave Dispersion},
  url = {https://github.com/CodeWithWinton/morse},
  year = {2026}
}