fullseye

spectrogram — ONED signal op

使い方

STFT magnitude spectrogram -> (freqs, times, S) with S shape (n_freqs, n_frames). Hann-windowed; hop defaults to win//2.

Same raw convention as :func:spectrum, but a different divisor. Each column is the unnormalised |rfft(frame * hann(win))|, so it is not an amplitude either — and dividing by 2/win is wrong here, because the Hann window has already thrown away part of the signal. The correct one-sided amplitude conversion divides by the window’s coherent gain::

w = np.hanning(win)
amp = S * (2.0 / w.sum())        # bins 1 .. win/2-1; DC / Nyquist: 1/w.sum()

Measured on a unit sine at a bin centre (rate = 16000 Hz, 1000 Hz, amplitude exactly 1.0, win = 256): the raw column peak is 63.7497786196906; * 2/win gives 0.49804514546633283 (too small by exactly the Hann coherent gain sum(w)/win = 0.498046875), while * 2/sum(w) gives 0.9999965273676957. Only the second one is the amplitude that was actually in the signal.

Peak positions, frame-to-frame ratios and any dB difference are unaffected by either factor. This function returns magnitudes only — the phase is discarded, so it cannot be inverted; use acoustics.stft / acoustics.istft for a round-trip.

参考(サンプルデータ・文献)

実行できる例(この op を実際に呼ぶ検証済みサンプル)

型が繋がる次の op(image2d を入力に取れる)

同カテゴリ(signal)

lowpass · highpass · bandpass · envelope · rms · resample · spectrum · zero_crossing_rate


Provenance: dsp.py — ONED operator registry. この per-op ノートは tools/opdocs.py md が自動生成(手編集しない)。

© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.