Oh, that’s fire! Picture a tiny neural net peeking into the spectrogram, nudging each bin toward its ideal pitch. We could start with a 128‑band STFT, feed the magnitudes into a lightweight Conv‑1D, then output a correction vector for the next hop. Training on a synthetic library of detuned notes will let it learn the “clean” manifold. Let’s prototype it in PyTorch, hook it up to a real‑time VST, and watch the auto‑toner dance. Ready to hit the code!

Alright, here’s a bare‑bones sketch you can drop into a cell and tweak: import torch, torch.nn as nn import torchaudio import numpy as np class ConvAutoCorrelate(nn.Module): def __init__(self, n_bins=128, hidden=64): super().__init__() self.encoder = nn.Sequential( nn.Conv1d(1, hidden, kernel_size=3, padding=1), nn.ReLU(), nn.Conv1d(hidden, hidden, kernel_size=3, padding=1), nn.ReLU() ) self.decoder = nn.Sequential( nn.Conv1d(hidden, hidden, kernel_size=3, padding=1), nn.ReLU(), nn.Conv1d(hidden, 1, kernel_size=3, padding=1) ) def forward(self, x): # x shape [B, 1, N_bins] h = self.encoder(x) out = self.decoder(h) return out def synth_detuned(freq, sr=44100, duration=0.1, detune=0.005): t = torch.linspace(0, duration, int(sr*duration)) phase = torch.cumsum(2*torch.pi*freq*(1+detune*torch.randn_like(t)), dim=0) return torch.sin(phase) # training loop placeholder model = ConvAutoCorrelate() optimizer = torch.optim.Adam(model.parameters(), lr=1e-3) criterion = nn.MSELoss() for epoch in range(200): # generate a batch of detuned sine bursts batch = [synth_detuned(440, detune=0.02) for _ in range(32)] batch = torch.stack(batch) # [32, samples] # STFT spec = torch.stft(batch, n_fft=256, hop_length=128, return_complex=False) mag = spec.pow(2).sum(-1).sqrt() # magnitude # target is the clean magnitude of a 440 Hz sine clean = torch.stft(torch.sin(torch.linspace(0, 0.1, 4410)), n_fft=256, hop_length=128, return_complex=False).pow(2).sum(-1).sqrt() mag = mag.unsqueeze(1) # [B, 1, N_bins] target = clean.unsqueeze(0).repeat(mag.size(0),1,1) pred = model(mag) loss = criterion(pred, target) optimizer.zero_grad(); loss.backward(); optimizer.step() # after training, you can hook model.eval() into the VST’s DSP chain: # for each incoming frame, compute its magnitude STFT, run model, apply corrections, overlap‑add the inverse STFT. # that’s the prototype—time to tweak learning rates, add a residual skip, maybe even let the net predict phase corrections next!