Testing AI tools with a Java Deluge emulator and proposing potential bug fixes and features developed with AI agents....

[ludoch]

Porting the Deluge to Modern Java emulator & Auditing the C++ Firmware

Hi everyone,
I’m a Java developer with 30 years of experience and a lifelong passion for computer music. When I first encountered the Synthstrom Deluge, I was immediately struck by its UI/UX masterclass—the grid-centric, clip-based workflow is infinitely more cohesive and intuitive than many traditional alternatives.
Over the last three months, I have been working on two software passion projects to explore Deluge's architecture: a Java emulation of the engine and a deep-dive audit of the C++ firmware.

I want to be transparent about my development process: both projects were created in close collaboration with AI agents, including Gemini, Claude, DeepSeek, GitHub Copilot, and the Google Antigravity CLI. As a veteran developer, I have adopted this agentic workflow to rapidly analyze complex real-time audio code, implement assembly fixes, and build a custom Java audio engine. While highly efficient, this process requires careful oversight; I treat these agents like junior programmers, constantly cross-verifying their work to mitigate hallucinations. I am continuing to refine this process, using these tools to audit development history, correct errors, and steadily advance the Java emulation toward achieving perfect sound parity with the original Deluge hardware.

I’d like to share these projects with you and open a discussion on Deluge's software architecture, memory safety, and future possibilities.

Project 1: The Java Deluge Engine emulator

As a Java engineer, I’ve long heard the industry dogma that Java is "too slow" or "too unpredictable" for real-time, low-latency audio DSP. With modern Java features (Virtual Threads, Vector API, and ZGC), I wanted to challenge this. I set out to write Java Deluge—a line-for-line object model emulation of the Deluge engine.

• The Result: Modern Java is capable of low-latency real-time DSP. The object model translates beautifully, and using modern concurrency primitives, we can run highly parallel voice rendering.
• The Current Catch: After three months, we do not have perfect sound parity yet. There are still timeline-shifting drifts and subtle timbre-curve variations compared to the physical hardware. The core engine architecture, however, is solid.

You can explore the project codebase on GitHub, which includes nearly 1,500 commits—almost entirely AI-generated. The project’s User Guide was also developed using AI.
This project is pure java, no native deps and using rtmidijava pure java MIDI library and runs anywhere a JDK is running.

Project Technical Overview

This project is pure Java, requiring no native dependencies. It utilizes the rtmidijava pure Java MIDI library and is designed to run on any platform supporting a JDK.

Project 2: Auditing the C++ Firmware (14 Fix Proposals, 4 Feature Proposals)

While building the Java emulation, my system programming background kicked in. I audited via AI tools the official open-source C++ firmware for memory safety, concurrency, and DSP stability. I have compiled and pushed 14 separate bug-fix branches to my fork that I would like to propose to the community. These fixes address several long-standing issues, including:

1. Low-Frequency Filter Instability: Resolving float-to-int cast overflows causing filter explosions.
2. Scheduler Deadlocks: Replacing CPU-starving busy-waits with cooperative yields during SD card reads.
3. NEON Vector Loop Buffer Overflows: Fixing pointer overshoots in SIMD wave renderers.
4. Reverb Model vtable Corruption: Fixing stale base-class pointer updates.
5. Unison FM Volume Suppression: Replacing destructive global amplitude resets with local tracking.
6. Deserializer Stack/Heap Overflows: Resizing and protecting the copy buffers for song preview image blocks.

You can view these proposed fixes here: DelugeFirmware GitHub Repository

All these fixes have been written via Claude Code and reviewed by Antigravity CLI.

Fix Name	Description	Pull Request Link
Low-Frequency Filter Instability	Clamps cutoff frequency curves to prevent float-to-int cast overflows at extremely low cutoff frequencies.	fix/filter-instability
Scheduler Deadlocks	Replaces CPU-starving busy-waits with cooperative yields when reading the SD card inside the audio loop scheduler.	fix/scheduler-busy-wait-deadlock
Audio Log Buffer Overflow	Replaces unsafe strcpy calls with bounds-enforced strncpy in logging routines.	fix/log-buffer-overflow
DX7 Envelope Safety and Bounds Checks	Adds validation checks for envelope rates and levels to prevent negative array indexing or out-of-bounds parameter reads.	fix/dx7-envelope-bounds-and-safety
Delay Buffer Heap Overflow	Caps elements cleared in DelayBuffer::clear to the allocated buffer size, preventing heap corruption.	fix/delay-buffer-overflow
Reverb Model Vtable Corruption	Corrected a stale base-class pointer alignment issue in setModel, preventing vtable corruption.	fix/reverb-model-corruption
Unison FM Volume Suppression	Replaces global unison carrier amplitude resets with local active flag tracking to prevent volume drops.	fix/unison-fm-volume-suppression
Voice Culling Logic Flaw	Ensures the first active voice is evaluated rather than skipped from fast-release culling.	fix/voice-culling-logic
Vector Loop Buffer Overflow	Constrains NEON vector loops to multiples of 4 with copy buffers for remainder samples to prevent pointer overshoots.	fix/vector-loop-buffer-overflow
Granular Processor Safety Risks	Deletes default copy assignment to prevent shallow copies and adds bounds checking to reverb feed.	fix/granular-processor-safety
DX7 Importer Span Safety	Adds explicit span size verification inside the sysex importer to prevent memory overflows.	fix/dx7-cartridge-span-safety
ModFX Processor Safety Risks	Deletes default copy assignment to prevent double-free risks of the delay line buffer.	fix/modfx-processor-safety
DSP Filter Loop Boundary Safety	Adds explicit bounds checks to prevent do-while body execution on empty slices.	fix/filter-loop-safety
Deserializer Buffer Overflow Safety	Resizes buffers to 1024 bytes and adds copy limits to prevent corruption when loading large song preview images.	fix/deserializer-buffer-overflow

Community-Requested Features

All these feature branches have been written by Antigravity CLI and reviewed by Claude code.

Branch: feat/iteration-skip-condition

Discussion Reference: Discussion #4449 (Iteration Skip / Play Only N of M conditions)
Description: Implements the inverse condition (!N of M) for every iterance preset. When selected, it prints dynamically as "!N of M" on the OLED screen and skips the note event on cycle N, playing on all other cycles.

Branch: feat/midi-track-delay-offset

Discussion Reference: Discussion #1516 (MIDI Track Delay/Latency Offset)
Description: Adds a track-level offset parameter (ranging from -100ms to +100ms) to calibrate timing on external hardware relative to the internal Deluge sequencer clock.

• Adds delayMs directly to the Instrument base class with XML serialization.
• Exposes the offset parameter as "OFFSET" in the MIDI track options menu.
• Dynamically converts millisecond delay to tick offsets based on song tempo.

Branch: feat/song-snapshots

Discussion Reference: Discussion #3382 (Performance Song Snapshots)
Description: Captures and instantly recalls track mute and clip launch states during live sets without dynamic memory allocation.

• Adds snapshot variables to Clip and Output classes.
• Implements takeSnapshot() and recallSnapshot() routines.
• New Shortcuts: Hold Learn + Song to take a snapshot; Hold Shift + Song to recall.

Branch: feat/latching-stutter

Discussion Reference: Discussion #4196 (Latching Stutter Mode)
Description: Adds a latching mode to the stutter effect.

• Extends StutterConfig with a latch boolean field.
• Exposes "Latch" under the Stutter settings menu.
• Toggles stutter state on button press and ignores release when enabled.

I am aware that many open-source projects have policies against accepting AI-driven pull requests, and I completely respect those community guidelines.

Live Hardware DSP-Tap Connection

To use the live DSP-tap feature, build and flash the feat/dsp-buffer-dump firmware branch from this repository. This provides a low-level debug tap for raw audio analysis.

For setup instructions in the Java application, refer to the usage documentation.

The Firmware Branch: feat/dsp-buffer-dump

This branch introduces a custom developer-only SysEx debugging protocol inside the Deluge firmware, allowing the desktop workstation to capture raw audio samples directly from the physical hardware's output rendering loop:

• DSP Tap module (dsp_tap.h, dsp_tap.cpp): Allocates a 16 KB SRAM buffer on the Deluge CPU to capture a 4096-sample (~93 ms) window of the final master output.
• Final Render Tap (audio_engine.cpp): Taps the final audio output buffer directly inside the core rendering loop (renderAudio).
• Auto-Arm on Note On (voice.cpp): Automatically starts recording the buffer the exact millisecond a note triggers (Voice::noteOn).
• USB SysEx Command Protocol (sysex.cpp): Implements manual arming, note-onset triggering, and data read-back commands.

How to Fetch, Build, and Flash the Firmware

1. Fetch the Branch
In your local C++ firmware repository directory, fetch and checkout the remote branch:

git fetch origin
git checkout -b feat/dsp-buffer-dump origin/feat/dsp-buffer-dump

2. Build the Firmware
Compile the firmware binary using the Deluge build tools:

dbt release

3. Flash the Deluge

1. Turn off your Deluge and insert the SD card into your computer.
2. Copy deluge.bin to the root folder of the SD card.
3. Re-insert the SD card, hold Shift, and turn on the Deluge.

USB Integration Status Report: Deluge Firmware & Workstation Client

We have successfully designed, built, and tested the core layers for Phase 1 (Non-blocking file browsing & reading) and Phase 2 (Real-time synthesizer parameter editing) over a single composite USB MIDI + CDC Virtual Serial interface.
Technical Roadmap: usb_integration_roadmap.md

1 C++ Firmware (Deluge Device Side)

• GitHub Repository: trudaine/DelugeFirmware
• Branch: feature/tinyusb
• Latest Commit: d8ecb432

Codebase Links

• Core Packet & Transfer Logic: usb_sync.cpp
  • Packet Parsing State Machine: processIncomingCdcData
  • Phase 1 - Directory Listing: sendDirectoryListing
  • Phase 1 - Chunked File Streamer: processOutgoingFileTransfer
  • Phase 2 - Parameter Edit Senders/Receivers: handleReceivedPacket
• USB Task Runner: usbSyncTask (called every 10ms to schedule tasks without blocking the audio thread).

2 Java Workstation (Host/Desktop Side)

• GitHub Repository: trudaine/deluje
• Branch: main
• Latest Commit: 51881e02

Codebase Links

• Serial Service Protocol Class: DelugeUsbSyncService.java
  • Packet Transmitter: sendPacket
  • Command Dispatcher: runLoop
  • Phase 1 Senders & Listeners: Lines 28-42
  • Phase 2 Senders & Listeners: Lines 75-102
• Technical Roadmap: usb_integration_roadmap.md

3 Local Verification & Diagnostic Utilities

We added two Java test utilities to verify the integration, runnable with a physical Deluge connected:

Diagnostic 1: SD Card File List & Download

• Class: HardwareUsbFileTransferDiagnostic.java
• Behavior: Connects to the Deluge serial port, lists files in the /SONGS folder, and automatically downloads the first .XML song file over USB to a local file (downloaded_song.xml).

Diagnostic 2: Real-time Parameter Edit

• Class: HardwareUsbParameterDiagnostic.java
• Behavior: Connects to the active Synth slot on the Deluge, reads the active LPF Resonance value, modifies it in real-time, reads it back, and reports verification success.

Discussion Points

The Java Deluge application has only been tested by me and AI agents. It is WIP and we are still closing DSP parity gaps. Feel free to try it, file bugs, send feedback, and propose PRs.

For the Deluge fixes and new features, I can create PRs if needed, and of course happy to discuss/improve all of them.

[seangoodvibes]

Hey welcome

Just fyi, as I see you're using the tinyusb branch of the firmware - that branch is significantly out of date and not in line with the current main branch

[ludoch]

Thanks for the welcome and the information. First attempt to revive USB work and new features now in https://github.com/trudaine/DelugeFirmware/tree/feat/usb-comm

Will start testing this soon. Thanks!

[seangoodvibes]

Thanks for the welcome and the information. First attempt to revive USB work and new features now in https://github.com/trudaine/DelugeFirmware/tree/feat/usb-comm

Will start testing this soon. Thanks!

No problem!

By the way just so you are aware, we aren't actually proceeding with the tiny usb implementation in the future. That branch was more for exploration purposes and I believe we've landed on embassy usb in rust now.

But if tiny usb works for you, go for it. You'll just need to keep that in mind in the future whenever we transition the codebase from c/c++ to c/rust/c++.

Cheers

[ludoch]

Thanks again for more details. At the end, whatever new firmware USB features could be used on the Java client side ie usb audio streaming and more/better file system sync, remote control, etc.
Rust migration has started? Very happy such plans exist and I will adapt to whatever firmware new directions.

[lopho]

A bare metal deluge emulator using qemu running firmware images is already available here.
https://github.com/gramster/delugemu

[ludoch]

Hi,
Yes, this emulator is amazing and a true sound rendering of the original synth.

On my side, I was exploring as well a pure Java implementation:

• 100% Pure Java Architecture: The entire DSP sound engine and firmware logic is ported directly into Pure Java with zero C++ or JNI code required.
• True Cross-Platform Portability: Runs instantly on macOS, Windows, and Linux on any standard JVM ( JDK 27+ )—no native compilation, toolchains, or platform libraries ( .dll / .dylib / .so ) including all the MIDI layer. I wanted to see how far the latest Java Platform could behave with such demanding requirements for threading and memory management and garbage collection.

Also, wanted to explore UX Extensions (Beyond the Hardware Box)

• Expanded Multi-Resolution Grid: Dynamic visual grids ( 8×16 , 16×16 , 24×16 ) with gold Shift-shortcut overlays rendered directly on pads.
• Widescreen Synth Configurator: 8 tabbed visual editors (OSC/Filter, DX7 6-Op FM, Envelopes, LFOs, 24-Slot Mod Matrix, Arp, Unison, FX/EQ) extending 1-line OLED menu diving.
• 3D Wavetable Laboratory: Interactive 3D wireframe waterfall plots for inspecting and morphing wavetables.
• Sample & Waveform Slicer: High-resolution visual waveform editor with transient slice detection and loop crossfading.
• Dedicated Creative Generators: Built-in Drone Lab (ambient soundscapes) and Algorithmic Euclidean Generator (polyrhythms).
• Full Desktop DAW Editors: Multi-octave Piano Roll, continuous Automation Lanes, and Microtonal (.scl/.kbm) Tuning Editor.
• Modern DAW Workflow: 64-level Undo/Redo ( Ctrl+Z ), Multichannel WAV Audio Export, SysEx communication with real Deluge, Standard MIDI File (.mid) Import/Export, and USB SD-Card File Sync.
• And more as I explore features that would produce perfect Deluge XML files with advanced desktop level tooling.