Detection

On a whole recording

import dspk

rec = dspk.io.open("session.imec0.ap.bin")
probe = dspk.synapse.io.probe_of("session.imec0.ap.bin")
sorting, summary = dspk.synapse.detection.detect_recording(rec, probe)

summary["spikes"], summary["sigmas_uv"]
sorting.spikes()["locations"]      # where each spike came from, in µm

It reads in place in batches: optional CAR, zero-phase band-pass, noise measured on chunks spread over the recording, threshold crossings, removing the same spike seen on neighbouring channels, snippets and templates. Chunked detection gives the same result as detecting on the whole recording.

The command line does the same:

dspk detect session.imec0.ap.bin --output detected.sorting.zarr

On data in memory

Detector	Finds
threshold	Local extrema beyond threshold * sigma on each channel
neo	Peaks of the Teager–Kaiser energy
matched_filter	Correlation with a spike shape

Noise is median(|x|) / 0.6745 (Quiroga et al., 2004) unless you give sigmas. Each returns samples, channels and amplitudes, sorted by time, and reports the spike’s extremum.

spacing="locally_exclusive" (the default) keeps a spike unless a larger one is nearer, which chunked detection reproduces exactly. spacing="scipy" follows scipy.signal.find_peaks.

Then deduplicate keeps one spike per event seen on several channels, and snippets cuts the waveforms.

Worked examples: Course: Detection.