Validation balanced accuracy
400 validation clips
Classify guitar technique from one-second audio
A machine-learning model identifies how a guitar is being played from compact audio windows, using MFCC, RMS energy, zero-crossing rate, and spectral features.
background
chords
palm muted
single note
Upload a 1-second guitar clip
No WAV uploaded
Upload WAV to play it here
PredictionAwaiting WAVConfidence--Upload WAV to analyze
Class probabilities
background
0.00
chords
0.00
palm muted
0.00
single note
0.00
Model results
The baseline logistic-regression model is trained on session metadata splits and evaluated on a held-out recording session.
Test balanced accuracy
428 held-out test clips
Window size
16 kHz mono features
testvalidation
| actual / predicted | background | chords | palm muted | single note |
|---|---|---|---|---|
| background | 118 | 1 | 1 | 0 |
| chords | 1 | 89 | 3 | 8 |
| palm muted | 0 | 2 | 86 | 11 |
| single note | 2 | 5 | 10 | 91 |
Pipeline overview
From recordings to predictions, each step is reproducible from the repository scripts.
Audio ingestion
Split raw WAV sessions into one-second clips with metadata-controlled splits.
Quality control
Flag silence, clipping, duration drift, duplicate clips, and amplitude outliers.
Feature extraction
Resample to 16 kHz mono, then compute MFCC, RMS, zero-crossing, and spectral features.
Model inference
Run the balanced logistic-regression classifier and return label probabilities.
Feature set
- 13 MFCC means and standard deviations
- RMS mean, spread, max, and 90th percentile
- Zero-crossing rate
- Spectral centroid, bandwidth, rolloff, and flatness
Live recognition
The repository includes a live prediction script for a guitar interface on macOS.
$ python3 src/live_predict.py --device 3 --channels 1
single_note confidence=0.82 rms=0.04123 peak=0.512
single_note: 0.82 | palm_muted: 0.10 | chords: 0.06 | background: 0.02Outputsingle notesmoothed over 5 windows