transform op• Data kinds: table → signal
• Call: import acoustics; acoustics.istft(transform) (or opsacoustics.get("istft"))
Invert :func:stft by weighted overlap-add — exactly.
Weighted overlap-add divides the synthesised sum by the overlap sum of the
*squared* window, which makes the reconstruction exact for any window and
hop satisfying NOLA, not only for the COLA pairs. :func:stft refuses the
NOLA violation up front, so if the transform was produced by it the inverse
cannot be lossy.
Measured round-trip error, `max |x - istft(stft(x))|` on 4096 samples of
white noise (float64, so 2.2e-16 is one ulp of the largest sample):
=============== ==== ==== ========= =========
window win hop max error nola_min
=============== ==== ==== ========= =========
hann 256 128 1.33e-15 0.5
hann 256 64 1.33e-15 1.5
hann 256 255 2.73e-12 2.27e-08
hamming 256 128 1.33e-15 0.5832
blackman 512 128 1.33e-15 1.206
flattop 256 64 1.33e-15 0.396
boxcar 256 128 8.88e-16 2.0
hann (nfft 512) 256 128 1.33e-15 0.5
=============== ==== ==== ========= =========
Read the third row's two columns together. `hop = 255` on a 256-sample
window overlaps by one sample, which breaks plain (unweighted) overlap-add
completely; weighted overlap-add still inverts it, but only to 2.7e-12
rather than 1.3e-15, because the squared-window overlap sum falls to
2.3e-08 and the reconstruction divides by it. NOLA is satisfied and the
result is four orders of magnitude less accurate than every other row —
which is why `nola_min` is *returned* and not merely checked. A NOLA
minimum that is small but positive is a conditioning warning, and there is
no threshold at which it stops being one, so no threshold is invented here.
Raises `ValueError: a dict missing any key :func:stft` writes, a
`spectra whose shape disagrees with the recorded nfft` / frame count,
or a non-complex `spectra`.
• acoustic_condition_monitoring family guide
• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).
• Operator provenance and references — the sources of the research/methods this op family came from.
• The canonical algorithm (author, year) and its uses are named in the family usage guide above.
• acoustic_condition_monitoring — py -3.11 examples/acoustic_condition_monitoring.py
signal as input)stft · envelope_spectrum · spectral_kurtosis · cepstrum · angular_resample · order_spectrum · octave_spectrum · weighting_response
transform)*Provenance: acoustics.py — ACOUSTICS operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.