Resampler

Overview

Grows a small folder of audio samples into a much larger one. Point it at a few dozen sounds and Resampler trains a small neural network on them — entirely on your Mac — then writes out new sounds that blend, drift from and mutate what it learned. Sliders and presets from Faithful to Extreme push the results from close relatives of the originals to deliberately destroyed extremes, and a seed reproduces a batch exactly.

How it works

The model

Resampler trains a convolutional variational autoencoder on log-mel spectrograms of your samples. The encoder squeezes each sample into a 64-number latent code; the decoder turns any code back into a spectrogram. Training and generation both run on the Mac's GPU through Metal, and the trained model is cached and reused until the folder changes.

Analysis that adapts to the corpus

Short analysis windows keep drum attacks sharp but smear the partials of a sustained note into noise; long windows do the opposite. Resampler measures the spectral flatness of the corpus and picks a percussive profile (1024-point FFT, 96 mel bands) or a tonal one (4096-point FFT, 384 mel bands). You can override the choice.

Making a variation

1. Pick a number of corpus samples and blend their latent codes.
2. Add noise: Drift moves along the directions in which the corpus itself varies; Mutation moves along directions it never uses.
3. Optionally Push beyond the corpus average, then decode to a spectrogram.
4. Shape the spectrogram — pitch, time-stretch, brightness tilt, contrast, smear, grit — then rebuild its phase with fast Griffin–Lim (few passes sound robotic, on purpose), followed by decay, drive, bit and sample-rate crushing and optional reversal.

Limits worth knowing

With a small corpus the network mostly interpolates: variations resemble blends of your samples, not a general model of the instrument. Output is mono; the model learns at most four seconds of each sample; and Griffin–Lim phase reconstruction softens transients and very high partials a little.