Dinhopjsa/guitar-audio-transcriber-ai
🎸 Transform guitar audio into note predictions and tablature using deep learning and DSP in this research-focused project.
What's novel
🎸 Transform guitar audio into note predictions and tablature using deep learning and DSP in this research-focused project.
Code Analysis
9 files read · 3 roundsA prototyping project that classifies single guitar notes from audio clips using an ensemble of a small MLP (on MFCC features) and a small CNN (on mel-spectrograms), with onset-based audio slicing and YIN pitch detection as auxiliary features.
Strengths
The training pipeline is well-structured with proper use of modern PyTorch practices (AMP, gradient clipping, label smoothing, ReduceLROnPlateau, Kaiming init), and the audio preprocessing pipeline (DB gating → RMS gating → onset detection → slicing) is a reasonable multi-stage approach. The config system using dataclass inheritance is clean and practical.
Weaknesses
The 'DSP algorithms' and 'advanced deep learning' claims are misleading — the DSP is entirely librosa wrappers and the models are small standard architectures with no novel contribution. There is zero test coverage, significant code duplication between MLPTrainer and CNNTrainer (~70% shared logic), and the README describes a polished user-facing application that doesn't exist (just a tkinter file dialog and print statements).
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
28
Contributors
2
Files
36
Active weeks
5
Repository
Language
Python
Stars
1
Forks
1
License
—