KudioEnhance workflow

Three stages, three re-runs

Each stage reads what the one before it wrote, so you can re-train without re-mixing and re-evaluate without re-training
run does all three in order; synth, train and evaluate also stand on their own
Once a run exists, denoise enhances a file or a whole folder with it, and curves says whether the fit actually converged

What the config decides

snr_db, rt60 and channel set how far the mixture is degraded before the model ever sees it
reverberant_target chooses what the model is asked to produce: the dry clean file, or the same speech with the room left on

Five architectures, two targets

ddae is frame-wise; blstm, conv_ae, crn and unet take sequences, so build_dataset shapes the examples to match
crn is the causal one: its output for a frame never waits for a later frame
A spectrum target is standardised, a mask target is not: it already sits in [0, 1], where the sigmoid output can reach it