greblus/solitito-ai

Solitito Guitar Chord Model (v0.1.0)

This is a lightweight, real-time guitar chord recognition model optimized for CPU inference. It serves as the core engine for Solitito - an open-source Guitar Chord Trainer & Arpeggio Practice tool written in Rust.

• Repository: https://github.com/greblus/solitito
• Format: ONNX
• License: MIT

Model Description

The model is a Transformer/CNN hybrid trained to recognize jazz guitar chords from a raw audio signal converted into a specific feature set. It is designed to run efficiently on consumer hardware (laptops) without needing a dedicated GPU.

Classes

The model recognizes:

• Roots: C, C#, D, D#, E, F, F#, G, G#, A, A#, B
• Qualities: Major, Minor, 7, Maj7, m7, dim7, m7b5, 9, 13 (and "Note" for single notes).

Usage (Preprocessing Required!)

Important: This model does not accept raw audio waveforms. You must preprocess the audio exactly as defined in dsp_weights.json.

Input Specification

• Input Shape: [Batch, 32, 156]
  • 32: Context frames (approx. 0.5s of audio history).
  • 156: Feature vector size per frame.

DSP Pipeline (Rust/Python)

1. Sample Rate: Resample audio to 16,000 Hz.
2. FFT: Size 8192, Hop Length 256.
3. CQT (Constant-Q Transform): 144 bins (6 octaves, 24 bins/octave).
4. Chroma: 12 bins derived from CQT.
5. Normalization: Logarithmic (dB) normalization + scaling to 0.0-1.0 range.
6. Stack: The input vector is [CQT (144) + Chroma (12)] = 156 float32 values.

Note: The dsp_weights.json file included in this repository contains the pre-calculated complex weights for the CQT matrix multiplication, allowing for fast Rust implementation.

Training Data

The model was trained on a combination of custom datasets and the GuitarSet dataset.

• Xi, Q., Bittner, R. M., Ye, X., & Bello, J. P. (2018). GuitarSet: A Dataset for Guitar Transcription. In ISMIR.

How to use in Rust

See https://github.com/greblus/solitito for a full implementation.