Open Source · 16 Voices + 8 FX · 46/46 Tests

txt2vst

Text → VST in One Command

Generate VST3/AU instruments from natural language.
16 voice engines. 8 FX processors. 23 UI themes. Zero friction.

txt2vst 
# From text to VST3 in seconds
$ python3 prompt2spec.py "drum machine acid 4 channels"
$ python3 forge.py spec.json output/AcidBox
$ cmake -B build && cmake --build build
  [100%] Built target AcidBox_VST3
  Installed: ~/Library/Audio/Plug-Ins/VST3/AcidBox.vst3
16+8 DSP Archetypes
10M+ Unique Sounds
2.7B+ Visual Combos
~1500 LOC Generated
View on GitHub 🤗 Try on HuggingFace Explore Archetypes

How It Works

Four stages. Fully deterministic. No LLM in the audio path.

1
Natural Language
"acid drum machine 4ch"
2
spec.json
Channels, voices, features
3
Forge + Router
22 C++ files generated
4
VST3 / AU
Installed & ready

16 Voice Archetypes

Each voice gets a purpose-built synthesis engine — not a generic fallback.

Kick
Dual envelope + pitch sweep + sub + drive
16ns/sample · 0.07% CPU
Snare
Body tone + bandpass noise + snap transient
9ns/sample · 0.04% CPU
Hats
6-oscillator ring modulation + noise blend
3ns/sample · 0.01% CPU
Tom
Pitched sine + exponential pitch sweep
4ns/sample · 0.02% CPU
Perc/FM
FM carrier/modulator + index envelope
8ns/sample · 0.03% CPU
Clap
4 micro-bursts + diffuse tail + bandpass
4ns/sample · 0.02% CPU
Bass Acid
PolyBLEP + 4-pole diode ladder filter
67ns/sample · 0.30% CPU
Lead
PolyBLEP pulse/PWM + SVF + portamento
8ns/sample · 0.03% CPU
Pad
Detuned saws + sub + LP filter + ADSR
8ns/sample · 0.03% CPU
Pluck
Karplus-Strong physical model + damping
5ns/sample · 0.02% CPU
Organ
9-drawbar harmonics + rotary speaker sim
24ns/sample · 0.10% CPU
FM Synth
2-op FM with self-feedback + index env
8ns/sample · 0.03% CPU
Noise
White/pink + SVF filter + Paul Kellet
6ns/sample · 0.03% CPU
String
4 detuned saws + pitch drift + SVF LP
14ns/sample · 0.06% CPU
Brass
Pulse + PWM + attack swell + DC blocker
8ns/sample · 0.04% CPU
Sub Bass
Sine + harmonics + soft saturation
4ns/sample · 0.02% CPU

Occam Guardrails

Every archetype passes a 6-point quality gate. Default and extreme parameters. No exceptions.

No NaN
0 occurrences
Numerical stability
No Inf
0 occurrences
Filter stability
Peak amplitude
< 1.5
Headroom before clipping
DC offset
< 0.01
No speaker damage
CPU usage
< 5% per voice
Real-time safety
Deactivation
Must stop
No infinite tails
test results 
  40/40 All Occam guardrails satisfied!
# cd dsplib && g++ -std=c++17 -O2 -I voices \
#   -o test tests/test_voices.cpp && ./test

8 FX Processors

Master bus effects, routed per-channel or globally. All Occam-safe.

Delay
Stereo ping-pong, tempo sync, LP feedback
Reverb
Schroeder 4-comb + 2-allpass, damping
Chorus
Dual LFO modulated delay, stereo
Compressor
Feed-forward, attack/release, makeup
Distortion
Asymmetric tube-style soft clip
Phaser
4-stage stereo allpass + LFO sweep
EQ
3-band: low/mid-para/high shelf
Gate
Envelope follower + threshold + hold

24 UI Themes

Set "theme": "acid" in your spec and the generated plugin gets a matching color palette.

Midnight
Void
Obsidian
Acid
Neon
Glow
Strobe
Matrix
Ember
Solar
Copper
Candy
Frost
Chrome
Arctic
Vapor
Industrial
Terminal
Hologram
White
Cream
Blood
Lavender

Forge Your Plugin

Configure your instrument visually. Download the spec, run two commands, done.

midnight
bypass
4
spec.json preview

Ready to forge?

Clone. Describe. Build. Ship.

Get Started on GitHub 🤗 Try Online — No Install