Analog

GuSynth – Ultra-Low-Latency Polyphonic Guitar Synth Pedal

GuSynth is a research + hardware project building a real-time polyphonic guitar synth pedal capable of transforming a guitar into other instruments with ultra-low latency.

The project combines:

• Transient-guided neural note detection
• Polyphonic inference for guitar input
• Direct neural timbre transfer
• Embedded deployment on microcontroller hardware
• Analog tone shaping and post-processing

The goal is a standalone guitar pedal that lets a guitarist play naturally while hearing the output as another instrument.

---

Overview

Most pitch detection and guitar synthesis systems rely on longer analysis windows, which adds noticeable delay. GuSynth instead focuses on extracting useful information from the earliest part of the note, especially the pick attack, so the system can react quickly enough for live performance.

The project currently explores two connected ML stages:

1. First neural network: fast polyphonic note detection from very short windows
2. Second neural network: direct timbre transfer from guitar to another instrument

A harmonic LUT-based pipeline and analog processing stages support the overall system.

---

Signal Path

This is the full signal path of the pedal, from instrument input through conversion, inference/synthesis, and output shaping.

---

Overall Pedal / System Schematic

This diagram shows the higher-level pedal architecture and how the DSP / neural / synthesis stages connect together.

---

424 / Analog Circuit Reference

This section documents one of the analog reference directions used in the project for tone shaping / overdrive exploration.

---

Core System Architecture

1. First Neural Network – Polyphonic Note Detection

The first model is designed to recognize notes from very short guitar input windows, with a strong focus on the transient / pick attack region.

It is intended to:

• detect note activity quickly
• work under polyphonic conditions
• provide a MIDI-style or note-based representation
• operate under strict latency constraints suitable for embedded systems

Audio Demo – First NN

After (first NN output / note-detection-driven result):

2. Second Neural Network – Timbre Transfer

The second model takes guitar input and transforms it toward the sound of another instrument.

Current work is focused on guitar → piano-style transfer, with the long-term goal of supporting more instruments and cleaner real-time conversion.

It is intended to:

• preserve performance timing and expression
• reshape the raw guitar waveform into a target timbre
• run in a streaming low-latency pipeline
• support eventual embedded deployment

Video Demo – Second NN

Harmonic LUT Support Pipeline

Alongside the neural models, the project includes a harmonic fingerprint LUT approach used for analysis and refinement.

This system helps by:

• storing harmonic note templates
• comparing incoming spectra to expected note structure
• refining note hypotheses
• supporting interpretable DSP-guided analysis

Rather than treating the project as purely neural or purely classical DSP, GuSynth explores a hybrid approach.

---

Embedded Hardware Target

The long-term target is a compact embedded pedal platform.

Current embedded direction

• Daisy Seed / Cortex-M7 class hardware
• 48 kHz real-time audio
• small-window streaming inference
• ADC / DAC audio pipeline
• analog front-end and output shaping

Embedded constraints

• low total latency
• predictable real-time processing
• limited memory / compute budget
• models must be compact enough for sustained streaming

---

Analog Processing

The pedal is not only digital / ML based. Analog stages are also important for:

• input conditioning
• gain staging
• distortion / overdrive
• EQ shaping
• reducing unpleasant artifacts in synthesized output

This is especially useful when blending guitar and synthetic instrument textures.

---

Project Status

• ✅ First neural network prototype for transient-guided note detection
• ✅ Second neural network prototype for timbre transfer
• ✅ Harmonic LUT research pipeline
• ✅ System-level pedal architecture defined
• ⚠️ Embedded optimization still in progress
• ⚠️ More training data and refinement needed for cleaner output quality

---

Repository Layout

• Documentation/images/
  Diagrams, schematics, and system figures

• Documentation/audio/
  WAV demos for model outputs

• Documentation/video/
  Video demonstrations of the system

• LUT - In Development/
  Harmonic fingerprint / template detection work

• DAISY - In Development/
  Embedded hardware and firmware work

• JUCE - In Development/
  Desktop or plugin prototyping

---

Long-Term Goal

A standalone guitar pedal that:

• detects polyphonic playing with minimal delay
• converts guitar into the sound of other instruments
• runs on embedded hardware
• combines neural models, DSP, and analog circuitry
• remains responsive enough for real live performance

---

Summary

GuSynth is a hybrid guitar-synthesis system exploring how far real-time note inference and timbre transfer can be pushed under embedded, performance-critical latency constraints.

It is not just a guitar effect — it is an attempt to turn the guitar into a real-time expressive controller for other instruments.