fast-short-time-fft

Fast Short-Time FFT — 7× faster than scipy.signal.ShortTimeFFT, bit-identical.

Vectorized, CPU-only (pocketfft, no GPU). One batched rfft per signal vs per-slice Python loop.

from fast_short_time_fft import FastShortTimeFFT
import numpy as np

SFT = FastShortTimeFFT(win=np.hanning(256), hop=64, fs=8000)  # same API as scipy
Sx = SFT.stft(x)  # (129, p_num) — exact as scipy, 7× faster
# or
from fast_short_time_fft import stft
Sx = stft(x, win=np.hanning(256), hop=64, fs=8000)

For 1k wavs (8kHz, avg 2s, 250k frames): 0.66s vs scipy 3.3s (median, incl. I/O).

Repo: https://github.com/yonikremer/fast-short-time-fft

Install

pip install fast-short-time-fft

CLI

fast-stft sounds/*.wav --hop 64 --mfft 256 --workers 6

Why faster

Scipy loops for p in 250k: rfft(x_*win) in Python (10µs overhead per 256-pt FFT). This batches all p_num frames per file: pad → as_strided view → frames*win → rfft(workers=6).T — 1k Python calls vs 250k, one AVX *win, one batched C FFT with 6 threads. Same pad/win/rfft math → max err 0.

License

MIT