Converting a generated log-spectrogram to audio requires denormalizing, de-logging to amplitude, then an iSTFT

Neural audio models typically train on min-max normalized log-magnitude spectrograms, and add a channel dimension for convolutional layers. To turn a generated spectrogram back into a waveform, reverse each preprocessing step in order: (1) strip the extra channel dimension to restore a 2D (frequency × time) array; (2) denormalize using the per-sample original min and max stored during preprocessing; (3) de-log from decibels back to linear amplitude (librosa.db_to_amplitude); (4) apply the inverse STFT — here Griffin-Lim — passing the same hop_length used to analyse. Skipping the denormalize or de-log step scales amplitudes wrongly; the stored per-sample min/max values are therefore essential and must be persisted from preprocessing.

Examples

log_spec = spec[:, :, 0] (drop channel) → denorm = normalizer.denormalize(log_spec, mn, mx) → lin = librosa.db_to_amplitude(denorm) → signal = librosa.istft(lin, hop_length=hop_length).

Assessment

List, in order, the four steps to turn a normalized log-spectrogram (with a channel axis) into a playable waveform, and say why the original min/max must be stored during preprocessing.