Make a whole new instrument with your voice!
Sing a note, teach a synth your tone, then play a melody with your voice-instrument ♡
Sing one steady note.
Hold your own steady sound for 2 seconds—a hum, a vowel, or an unusual tone. No perfect pitch needed.
Sing for two seconds. Handwired learns your sound, then the piano is yours to play.
Your recording stays in this tab. Reloading or choosing a sample clears it.
A synth that learns from you.
These waveforms and the fit score describe the harmonic model only. They do not measure how much the synth sounds like your original recording.
Your voice. All these notes.
Keep your tone as the notes change.
LEARNED WITH HANDWIREDWhat did it learn?
Record or train above to make this keyboard sound like your chosen tone.
Computer key guide ⇧ = HOLD SHIFT
Shape the soundOPTIONAL · 64 HARMONICS ＋
Drag a slider to change the tone. Click its number to lock it during training.
First, find your note. Rust detects the pitch you sang. We take short sections of your actual waveform, keeping the details that make vowels and textures different.
Then, teach your instrument. Handwired calculates how to adjust frequency-and-phase coefficients until they match those short sections. The learned texture can retain vowel differences and some noisy or rough detail from your voice.
Finally, make a melody. Audio processing reconstructs the sound from those learned coefficients and tunes it to each piano key. Short sections overlap to sustain while you hold a key, without replaying the full recorded phrase. The preview and piano use the same instrument; hear your original recording to compare.
The nerdy details ＋
For recordings, each of 16 four-cycle segments is resampled to 1,024 points. An ordinary FFT measures the target coefficients. A Handwired Circuit computes their squared error and its gradients; gradient descent learns 1,022 signed Fourier coefficients from zero. The loss is proportional to waveform mean squared error apart from DC and Nyquist. Inverse FFT reconstructs the segments from the learned weights. An AudioWorklet overlaps phase-aligned, windowed segments in a changing order, with high-frequency removal before pitch transposition. The model memorizes short textures; it is not a pretrained voice model.
A second Handwired Circuit learns optional tone-correction filters for up to 25 nearby notes. A filter is kept only if it improves a spectral-envelope comparison on a separate part of the recording. Applying that curve to reconstructed segments is an approximation, so use Keep my tone to compare by ear. Farther notes can still sound smaller, deeper, or less like you. The synth does not reproduce words or infer how you would sing a different note. The built-in vowel samples keep the simpler 64-harmonic Handwired demonstration. No audio is uploaded.