Madaha

Madaha is a clone of the Yamaha S-YXG50 / S-YXG2006 LE software synthesizer, written in Rust. It parses the original Yamaha .tbl instrument banks and drives a full XG audio pipeline, aiming to be a superset of S-YXG2006 LE: everything the 2006LE supports must work, plus features it lacks.

Status

Functional. The complete chain works and is covered by tests:

• MIDI input: ALSA Sequencer / JACK / PipeWire ports (real-time note/CC/SysEx)
• Sound engine: full XG tone generation (Oscillator → LPF → HPF → Amp → EQ → Pan, PEG/FEG/AEG envelopes, dual velocity-layer elements, drum setup)
• Effects: all 2006LE effect kernels reverse-engineered from the Mac binary (Reverb/Chorus/Delay/ER/Distortion/Phaser/RotarySP) + the full XG2.0 effect set (~100 Variation types incl. Harmony/Vocoder/Talking, serial chains, Dyna family) — see dev_docs/effect_dsp.md
• MIDI behavior: controllers, RPN/NRPN, SysEx (GM/GM2/XG/GS/Roland/Yamaha), Active Sensing watchdog, XG RAM emulation with GS address remap, voice stealing with configurable scoring
• Audio output: ALSA / PipeWire backends, 4 depths, DC-offset correction, soft-clip, master gain
• Tests: 123 passing (unit + 6 end-to-end through the real render chain)

Requirements

• Rust nightly (edition 2024 — ci-never 1.95+). Stable will not work.
• ALSA (needed for MIDI input). Audio backends: ALSA, PipeWire.
• The original Yamaha .tbl files (not included). Madaha reads the S-YXG50 family files (sxgbin41.tbl + sxgwave4.tbl); the 2006LE files (sxgbnw6l.tbl etc.) are not read yet (see docs/TODO.md).

A Nix flake provides all system dependencies (direnv allow or nix develop).

Quick start

# copy the TBL files (or point madaha.toml at them with absolute paths)
cp /path/to/sxgbin41.tbl /path/to/sxgwave4.tbl .

nix develop
cargo run -- -D                          # debug logs, reads madaha.toml
cargo run -- -C /path/to/madaha.toml     # custom config
MADAHA_CONFIG_FILE=/path cargo run -- -D # via env var

Connect a MIDI source (keyboard/DAW/aplaymidi):

aconnect <source> Madaha:0     # list with: aconnect -l

Configuration

The config file is TOML and gitignored — an example lives at the repo root as madaha.toml. Field names follow the config structs:

log_level = "info"

[sound_module]
module_type = "syxg50"          # auto / syxg50 / syxg2006le / wingroove
tbl_bin_file = "sxgbin41.tbl"   # instrument definition file (absolute path ok)
tbl_data_file = "sxgwave4.tbl"  # wave data file

[audio]
engine = "alsa"                 # alsa / pipewire
sample_rate = 44100             # 22050 / 44100 / 48000 / 96000 / 192000
depth = "s16"                   # u8 / s16 / s24 / f32
buffer_size = 64                # block size, power of two
master_volume = 1.0             # output gain (0.05..=4.0)
soft_clip = true                # tanh soft clipping on the master bus
dc_blocker = true               # DC offset correction (XG Spec)

[midi]
max_polyphony = 512             # multiple of 16, max 2048
poly_replicant = 150            # voice count = max_polyphony × replicant/100
device_id = 16                  # SysEx device id (>= 16)
master_tune = 440.0
input_engine = "alsa"           # alsa / pipewire
# scoring = { ... }             # voice-stealing weights (see src/config/midi.rs)

Architecture

src/
├── main.rs            — entrypoint: args → config → check → Synth::run
├── args.rs            — clap CLI (-D, -C / MADAHA_CONFIG_FILE)
├── synth.rs           — MIDI event loop + audio render thread
├── config/            — TOML deserialization & validation (audio/midi/sound_module)
├── audio/             — rendering
│   ├── audio_render.rs — master bus, effect chain, DC correction, modulation
│   ├── wrapper.rs      — AudioRender (voices, effects, sinks)
│   ├── tone_generator/ — XG voice chain (oscillator/LPF/HPF/amp/EQ/pan + EGs)
│   └── dsp/            — effects: reverb/chorus/delay/ER/distortion/phaser/
│                         rotary/harmony/vocoder/talking + XG2.0 misc (xg20_effects)
│                         + core (biquad/delay/dc_blocker/wsola)
├── midi/              — engine, part, ram (XG/GS), sysex (GM/XG/GS/Roland),
│                         source (alsa/pipewire), active_sensing
├── voice_manager/     — TBL sound bank loading, instrument cache, drum setup
├── lfo/               — LFO (DDS + 13 waveforms)
├── double_buffer.rs   — lock-free-ish parameter sharing (Mutex + Arc snapshot)
└── libmadaha/         — (separate crate) TBL file format parsers

Tests

nix develop --command cargo test -- --test-threads=1

123 tests: DSP kernels, effect parameter tables (verified against the 2006LE binary), GS address remap, MIDI/SysEx parsing, config validation, and 6 end-to-end tests driving the real render chain (NoteOn → voice → audio buffer).

Reference docs

• dev_docs/effect_dsp.md — 2006LE effect DSP reverse engineering (kernels, coefficient flow, type dispatch, parameter tables)
• dev_docs/xg20_params.md — XG2.0 type parameter layouts (XG Spec)
• dev_docs/wsola_harmony.md — WSOLA pitch shifter + harmony/vocoder design
• dev_docs/gm2.md — GM2 SysEx implementation
• dev_docs/note_opencode.md — S-YXG50 .tbl binary layout (Ghidra RE)
• dev_docs/XGSpec2.0.md — XG 2.0 spec translation
• docs/TODO.md — deferred features and decisions

License

GNU General Public License v3.0 or later. See LICENSE.