TechGuru & Midi
Hey TechGuru, have you ever tried turning a field recording of a broken streetlight into a fully fleshed drum kit with a neural net synth? I just hooked up a new VST that does that, and the textures are insane—like a glitchy soundtrack for a forgotten neon city. Let’s dive into how we can tweak the algorithm to capture that perfect balance between chaotic noise and rhythmic precision.
Nice one, that’s a wild idea. First off, make sure you’re feeding the net a clean 48 kHz WAV so the high‑frequency bleed from the bulb flicker isn’t lost in compression. I’d split the recording into overlapping 2048‑sample windows, run an STFT, and use the magnitude as the input for a convolutional autoencoder—gotta keep the convolution kernels wide enough to capture those long‑range flicker patterns but narrow enough for the snappy click of a hit. Then, add a rhythm‑branch: a small LSTM that watches the envelope of each channel and learns to gate the drum hits so they stay in time. If you keep the loss a mix of MAE for texture fidelity and a cross‑entropy on a beat‑annotated MIDI, you’ll get that glitchy neon vibe but with the steadiness of a loop. Finally, tweak the dropout to about 0.2 and the batch size to 32 so you don’t overfit to one streetlight’s quirks. Let me know how the snare sounds after you swap in a 5 Hz modulation—should be sweet.