Custom operations

dspk.custom runs your own Python function over a recording of any length. dsp-kitchen reads the windows, the next one on another thread while your function runs. Your function only ever sees a NumPy window (channels, samples), float32.

Measure: one row of results per window

import numpy as np
import dspk

rec = dspk.io.open("session.zarr")
table = dspk.custom.measure(rec, lambda x: {"rms": np.sqrt((x ** 2).mean(axis=1))}, window_s=1.0)

table["start_s"], table["stop_s"], table["rms"]   # rms is (windows, channels)

Transform: a processed copy

Tell it how many samples your function looks away, the halo. Windows are read with that much context on each side and trimmed back, so the result equals your function on the whole array.

@dspk.custom.transform(halo=12)
def smooth(x):
    k = np.ones(25, dtype=np.float32) / 25
    return np.stack([np.convolve(c, k, mode="same") for c in x])

y = smooth.apply(rec, start=0, stop=30_000)        # one array in memory
smooth.write(rec, "smoothed.zarr")                 # the whole recording, window by window

Caution

A halo that is too small gives wrong samples at the window edges, with no error. Set it to at least how far your function reaches.

If your function needs to know where it is, take a second argument: fn(x, w) gets the Window, with its sample range and time.

Your function runs on the calling thread, one window at a time, in order, so it can keep state.

Full details: dspk.custom.