Preprocessing¶

A one-dimensional signal is transformed using an STFT, then log1p, DC-bin removal, and percentile clipping. The defaults use n_fft=1024 and hop=128, matching the released model's training. Frames are centered at samples 0, hop, 2 × hop, and so on. With the default even FFT size, N samples give 1 + N // hop columns.

A two-dimensional input is treated as an existing spectrogram. Set log=True for linear-scale values to apply log1p; the default assumes the input is already log-scaled. Before inference, the whole image is standardized with (x - mean) / (std + 1e-6). Large inputs can be tiled after this global standardization.

When a sample rate is supplied, the axes are in hertz and seconds. Otherwise they are bin and frame indices. See the Python guide for configuration.