SpikeInterface 0.105.0 release notes

Sept 21st 2026

Main changes:

• Added dartsort [Boussard2023] as an external sorter, run with run_sorter("dartsort") (#2125)

• Generalized the chunking machinery: the recording-specific chunking code is now a reusable TimeSeries base class, plus TimeSeriesChunkExecutor, so that any object chunkable along the time axis reuses the same parallelization and time handling (#4472, #4533)

• main_channel_indices and the main_channel_id sorting property replace peak_sign across postprocessing and metrics. The main channel is now set once on the analyzer, and sorters such as Kilosort and dartsort set it directly (#4624, #4672, #4733, #4725)

• SortingAnalyzer gained a lazy mode, which keeps the spike vector and extension data as memmaps / Zarr arrays and defers loading until first access, and an independent read_only mode (#4623, #4709, #4713) (see: Core module)

• Node pipelines can gather results straight to disk with GatherToZarr, so pipeline node extension data no longer has to fit in memory (#4703)

• Reworked probe handling around ProbeGroup: set_probe and set_probegroup now act in place, while the new select_channels_with_probe and select_channels_with_probegroup return a new recording (#4465)

• Added the detect_and_remove_artifacts preprocessor, with "envelope" and "saturation" detection, a separate recording_to_detect, and "zeros" / "noise" / "apodization" silencing modes (#4539)

• Added antialiased decimation (decimate(..., antialias=True)) and support for non-integer resampling rates (#4621)

• Added SortingAnalyzer.split_by() to split an aggregated analyzer into per-group analyzers (#4660), and read_kilosort4_motion to build a Motion object from a Kilosort4 output folder (#4518)

• Significant speedups in multi-segment synchrony metrics (#4770), duplicated spike removal (#4614), amplitude scalings (#4764, #4759, #4786), PCA over all spikes (#4488), NWB sorting loading (#4662), spike vector handling (#4618, #4695, #4581), and import time (#4641, #4644, #4650, #4651)

API and breaking changes:

Despite this long list of breaking changes, we have worked hard to try and preserve backwards compatibility for any objects saved using older versions of SpikeInterface. If you cannot load an object saved with an older version of SpikeInterface, please raise an issue.

• Probe handling is reworked around ProbeGroup, and the internal "contact_vector" property is gone (#4465). Previously saved objects and folders still load.

  • set_probe() and set_probegroup() always act in place and return None. Use select_channels_with_probe() / select_channels_with_probegroup() to obtain a new recording (see: Core module)

  • set_probe() accepts a single Probe only; pass a ProbeGroup to set_probegroup()

  • reset_probe() –> remove_probe()

  • has_3d_locations() –> is_probe_3d(), and has_channel_location() is now equivalent to has_probe()

  • get_channel_locations() raises when no probe is attached, instead of falling back to a bare "location" property

• The main channel of each unit is fixed when the analyzer is created and stored as the main_channel_id sorting property. peak_sign is also decided when the analyzer is created, and so is removed from the API (#4624)

  • get_template_extremum_channel() –> SortingAnalyzer.get_main_channels(), get_template_extremum_channel_peak_shift() –> get_template_peak_shift_on_main_channel(), get_template_extremum_amplitude() –> get_template_amplitude_on_main_channel(). The old names are deprecated shims

  • peak_sign is no longer a valid parameter of compute_quality_metrics, compute_template_metrics, compute_spike_amplitudes, compute_spike_locations and compute_unit_locations. It is removed from the saved params of older analyzers on load

  • compute_snrs(sorting_analyzer, unit_ids, peak_sign, peak_mode, ...) –> compute_snrs(sorting_analyzer, unit_ids, method), and the snr metric params peak_sign/peak_mode –> method

  • get_template_amplitudes(..., mode=...) –> peak_mode=...

  • compute_sparsity() and the ChannelSparsity.from_* constructors take peak_sign=None / amplitude_mode=None by default. With an analyzer they must stay None; with a Templates object they are now mandatory. amplitude_mode also moved earlier in the positional order of compute_sparsity()

  • estimate_sparsity() now defaults to peak_sign="both" instead of "neg", which changes results silently

  • peak_sign is removed from remove_redundant_units(), export_report() and plot_unit_depths()

  • sorting_to_peaks(sorting, extremum_channel_inds, dtype) –> sorting_to_peaks(sorting, main_channel_indices, dtype), taking an array rather than a dict, and SpikeRetriever now requires the sorting to carry main_channel_id

  • BaseSorting.to_spike_vector() gained main_channel_indices as its third positional parameter

• The chunking and node pipeline machinery is generalized to any time series (#4472, #4533)

  • ChunkRecordingExecutor –> TimeSeriesChunkExecutor, whose first argument is time_series rather than recording, and whose run(recording_slices=...) –> run(slices=...)

  • PipelineNode.get_trace_margin() –> get_margin(), PipelineNode.compute(traces, ...) –> compute(chunk, ...), and PipelineNode takes time_series rather than recording

  • run_node_pipeline(recording, ..., recording_slices=...) –> run_node_pipeline(time_series, ..., slices=...), and check_for_peak_source now defaults to False

  • divide_recording_into_chunks –> divide_time_series_into_chunks, write_binary_recording_file_handle –> write_binary_file_handle, get_random_recording_slices –> get_random_sample_slices

  • BaseRecordingSegment now derives from TimeSeriesSegment, whose time attributes are private (#4701): use get_times_kwargs(), get_start_time(), set_times() and shift_times() instead of segment.t_start, segment.time_vector and segment.sampling_frequency

• Saving and serialization are unified on the concrete classes (#4671, #4661, #4664)

  • BaseExtractor.save() is abstract, and BaseExtractor._save() is removed, so third-party extractors overriding _save no longer take part in save(). Use Recording.save(), Sorting.save() and Snippets.save()

  • folder_metadata is removed from to_dict(), dump(), dump_to_json() and dump_to_pickle(), and load_metadata_from_folder() / save_metadata_to_folder() are removed. to_dict() also gained include_extra_metadata in its third positional slot

  • check_if_memory_serializable(), check_if_json_serializable() and check_if_pickle_serializable() –> check_serializability(type)

  • BaseExtractor.load_from_folder() is removed; use si.load()

  • ZarrRecordingExtractor.write_recording(recording, zarr_path=...) –> folder_path=..., and add_traces_to_zarr() is removed

  • _get_t_starts() –> get_segment_t_starts()

  • si.load() forwards unknown keyword arguments instead of dropping them, so a bad argument now raises (#4709)

• PreprocessingPipeline takes a list of {"name": ..., "params": ...} steps instead of a dictionary (#4755) (see: Preprocessing module)

  • PreprocessingPipeline.preprocessor_dict –> .preprocessor_list

  • get_preprocessing_dict_from_analyzer / get_preprocessing_dict_from_file –> get_preprocessing_list_from_analyzer / get_preprocessing_list_from_file

  • apply_preprocessing_pipeline(recording, pipeline_or_dict, ...) –> apply_preprocessing_pipeline(recording_or_dict, pipeline, ...). Passing a dictionary still works, but is deprecated

• Other signature and naming changes

  • parent_recording –> recording in ChannelSliceRecording, FrameSliceRecording, depth_order, detect_and_remove_bad_channels and average_across_direction (#4779)

  • AnalyzerExtension._select_extension_data() –> _select_units_extension_data(), with a new _select_channels_extension_data(). This breaks custom AnalyzerExtension subclasses (#4648)

  • create_sorting_analyzer() gained several parameters between folder and sparse, so sparse and sparsity must now be passed by keyword (#4624, #4623)

  • NoiseGeneratorRecording and noise_generator_recording moved from spikeinterface.core to spikeinterface.generation (#4520). The lightweight MockRecording replaces the old core class for testing

  • resample(): margin_ms now defaults to None rather than 100.0, and skip_checks and check_nyquist() are removed (#4621, #4499)

  • read_kilosort_as_analyzer() gained recording as its second positional parameter (#4529), and model_based_label_units() gained unit_ids as its second positional parameter (#4716)

  • run_sorter raises an error when channel locations are not unique, suggesting run_sorter_by_property instead

  • Python packaging: numpy>=2.0 is required on all supported Python versions, the minimum neo and probeinterface versions are raised to 0.14.5 and 0.4.0 (#4593), the test, test_core, test_extractors and test_preprocessing extras are replaced by PEP 735 dependency groups, sonpy is dropped, and the pandas<3 (#4691) and scikit-learn<1.8 (#4780) pins are relaxed

• Changed results and defaults

  • sliding_rp_violation values change: the metric now reuses compute_correlograms and derives refractory-period bins from the actual correlogram bins (#4324), the expected-violations formula is corrected to the [Llobet] formulation, which previously overestimated it, and the tested contamination range now includes 35% (#4682)

  • Multi-segment synchrony_metrics were incorrectly grouping equal sample indices across segments, and are now computed per segment (#4770)

  • estimate_sparsity() defaults to peak_sign="both" (#4624)

  • amplitude_cutoff now prefers amplitude_scalings over spike_amplitudes (#4513)

  • num_depth_bins for MEDiCINe motion estimation uses MEDiCINe’s own recommended default (#4686)

  • Preprocessed traces are rounded, rather than truncated, before being cast to integer dtypes (#4653)

  • Amplitudes are scaled to uV using channel offsets in addition to gains (#4765)

  • get_metrics_extension_data() drops duplicated columns when two extensions provide a metric of the same name (#4609)

  • Template metrics no longer raise on 3D channel locations; they warn and use the first two dimensions (#4572)

  • decimate emits a FutureWarning when antialias is not set explicitly: the default will become True in a future release (#4621)

• Removal of previously deprecated features

  • curation.auto_label_units() –> model_based_label_units() (#4716)

  • the load_sync_channel argument of OpenEphysBinaryRecordingExtractor –> stream_name / stream_id (#4466)

  • the num_chan argument of BinaryRecordingExtractor –> num_channels (#4666)

• New deprecations, which will be removed in 0.106.0, and which now raise FutureWarning rather than DeprecationWarning so that they are visible by default (#4667, #4676):

  • BaseSnippets

  • the in_place argument of set_probe, as well as its return value

  • set_channel_locations() and clear_channel_locations()

  • get_template_extremum_channel() –> analyzer.get_main_channels()

  • get_template_extremum_channel_peak_shift() –> get_template_peak_shift_on_main_channel()

  • get_template_extremum_amplitude() –> get_template_amplitude_on_main_channel()

  • save_to_memory(), save_to_folder() and save_to_zarr() –> save(format=...)

  • the "npz_folder" save format and NpzFolderSorting.write_sorting() –> "numpy_folder" (reading npz_folder is still supported)

  • the extremum_channel_inds argument of SpikeRetriever –> main_channel_ids from the sorting

  • passing sparsity and job arguments as loose keyword arguments to create_sorting_analyzer –> sparsity_kwargs and job_kwargs dictionaries

  • the input_data argument of model_based_label_units –> metrics

  • split_non_somatic_good_mua and label_non_somatic in Bombcell curation –> split_non_somatic

  • computing amplitude_cutoff without the amplitude_scalings extension (#4513)

• Passing a dictionary to apply_preprocessing_pipeline is deprecated and will be removed in 0.107.0 (#4755)

• Items deprecated during the 0.104 cycle, such as return_scaled, the spikeinterface.qualitymetrics and spikeinterface.postprocessing.template_metrics modules, CurationModel, get_potential_auto_merge(), get_default_qm_params, get_default_tm_params, the sorting and sorting_analyzer arguments of several widgets, are removed.

core:

• Add GatherToZarr and save pipeline node extension data to disk (#4703)

• Respect requested unit order for templates and sparsity in SortingAnalyzer.select_units (#4791)

• Fix main_channel_id donor selection when merged units disagree (#4773)

• Keep spikeinterface importable when zarr-python v3 is installed (#4751)

• Run the kwargs backward compatibility hook regardless of the saved version (#4749)

• Propagate shared annotations in aggregate_channels (#4731)

• Support a SortingAnalyzer with no unit in the core extensions (#4729)

• Fix n_jobs=-1 oversubscribing cores under CPU affinity restrictions (#4724)

• fix: preserve caller-owned inputs (#4721)

• Fix is_merged/is_split properties for different strategies (#4718)

• SortingAnalyzer: add read_only mode (different than lazy) (#4713)

• Fix channel order bug and make select_channels private (#4712)

• Fix cache spike vector in base sorting (#4710)

• Propagate lazy arg to from si.load() (#4709)

• Fixes a couple of minor warnings (#4707)

• Create SortingAnalyzer object directly during create_sorting_analyzer (#4705)

• Make time attributes private in TimeSeriesSegment (#4701)

• Fix and speed up to_reordered_spike_vector (#4695)

• Clean up code around create_sorting_analyzer and create/load_* functions in sortinganalyzer.py (#4680)

• Add backwards compatibility when passing sparsity kwargs to create_sorting_analyzer (#4677)

• Move the Base.save to Sorting.save() and Recording.save() (#4671)

• Add copy_sorting option to SortingAnalyzer.create_memory (#4668)

• Remove load_from_folder from base (#4664)

• Cleanup Serializability checking in base.py (#4661)

• Add SortingAnalyzer.split() function for aggregated analyzers (#4660)

• Fix template estimate sparsity (#4657)

• Parameter pass through for class -> function wrapper (#4656)

• Improve sparsity estimation for aggregated analyzers (#4649)

• Add SortingAnalyzer.select_channels() function (#4648)

• Remove peak_sign in favour of main_channel_indices (#4624)

• Lazy-mode for loading Sortinganalyzer (#4623)

• write_binary: None dtype should be replaced by actual dtype (#4622)

• Avoid materializing entire time vector from zarr for time_to_sample_index (#4620)

• Remove some full lexsort of spikes when not necesary. (#4618)

• Remove deduplicated columns in get_metrics_extension_data (#4609)

• Use np.load(…, mmap_mode=”r”) for time_vector in BaseRecording._extra_metadata_from_folder (#4608)

• Fix file size calculation in write_binary() (#4607)

• SortingAnalyzer: intepret any remote path as zarr (#4583)

• Faster unit selection sorting. (#4581)

• Handle failure in dumping sorting provenance to Zarr (#4577)

• Add SelectSegmentEvent (#4575)

• Improve backward compatibility for dartsort after the chunkable PR (#4559)

• Add shift_times to BaseSorting (#4551)

• Improvements over chuking refactoring (#4533)

• Add num_channels to BaseRecordingSegment (#4526)

• Add get_start_time, get_end_time, and get_last_spike_frame to BaseSorting (#4525)

• Add t_starts parameter to generate_recording and generate_sorting (#4523)

• Add regression test for GH-4508 (deprecated metric names in legacy WE folders) (#4514)

• Improve get_times function to slice timestamps (#4512)

• Delete a failed SortingAnalyzer extension (#4511)

• Replace int() truncation with shared ms_to_samples() (#4484)

• Implement TimeSeries class (#4472)

• Add zarr_class_info attribute to zarr files for unambiguous and extendible zarr read (#4467)

• Use ProbeGroup object instead of contact_vector property and set_probe/probegroup from select_channels_with_probe/probegroup (#4465)

• Centralize add_segment and segments to BaseExtractor (#4462)

extractors:

• Add main_channel_id when running Kilosort (#4725)

• Add ZarrArrayExtractor to read “plain” Zarr arrays (#4760)

• Let get_streams work from the path alone on Neuralynx, Biocam and Blackrock (#4727)

• Add option to export spike waveforms subset to phy (mimic phy extract-waveforms) (#4699)

• Fix backward compatibility for binary recording/openephys/spikeglx (#4666)

• Build the NwbSortingExtractor spike vector with a single bulk read (#4662)

• Read the ragged index before the data when loading NWB table properties (#4655)

• Map NWB sorting spikes to samples via the ElectricalSeries timestamps (#4654)

• Fix NwbSortingExtractor unit-property loading for DynamicTableRegion columns (#4647)

• Infer sampling_frequency from Units.resolution in NwbSortingExtractor (#4646)

• Separate NWB reading from extractor construction (#4630)

• Materialize electrode-table columns before indexing in the NWB recording extractor (#4627)

• Fix read_mclust returning zero units for .t files (#4626)

• Use probeinterface’s renamed Neuropixels readers and new probe detectors (#4593)

• Implement cache spike vector for phy sorting reader (#4579)

• Fix the IBL Cbin extractor so that it works with 4 split shanks as well (#4567)

preprocessing:

• Accept recording= consistently in detect_and_remove_bad_channels, depth_order and average_across_direction (#4779)

• Extend preprocessing pipeline by defining a BasePipeline (#4755)

• Parallelize coherence-based bad channel detection (#4750)

• Handle empty annulus in local common reference (#4728)

• Cast not_enough_channels as str before ‘join’ (#4690)

• Fix remove_artifacts sparsity handling for median/average modes (#3290) (#4685)

• Fix test_pipeline_equiv_to_step (#4675)

• Fix double setting of operator in self (#4659)

• Round before casting preprocessed traces to integer dtypes (#4653)

• Add antialiased decimation and allow non-integer resampling rates (#4621)

• Add per-group ref_channel_ids to common_reference (#4601)

• Extend artifact detection: DetectAndRemoveArtifacts preprocessor and signed saturation (#4539)

• Fix false-positive margin warning in forkserver workers (#4506)

• Fix resampling of gapped recordings. (#4499)

• Allow overwrite argument when saving motion object (#4464)

• Refactor artifacts detection + add saturation detection (#4297)

sorters:

• Save recording_attributes before run_sorter() (#4795)

• Only use nonempty main channel indices when making an analyzer from templates (#4763)

• Fix encoding error in shellscript on Windows (#4720)

• Add read_kilosort4_motion function (#4518)

• Initial dartsort wrapper implementation. (#2125)

postprocessing:

• Speed up collision-aware amplitude scaling (#4786)

• Fix ComputeACG3D soft merge indexing bug that swaps kept units’ data (#4776)

• Materialize templates array before amp scalings computation for lazy mode (#4769)

• BUG: preserve correlogram unit axes across segments (#4767)

• Use channel offsets, not gains, when scaling amplitudes to uV (#4765)

• perf(postprocessing): compute only the amplitude scaling coefficients (#4764)

• Scale only the waveforms amplitude_scalings reads (#4759)

• Remove unused full-chunk scaling in spike_amplitudes (#4758)

• Fix PCA metric recomputation when “soft” merging (#4748)

• Expose seed for spike_locations (grid) (#4717)

• Remove peak_sign in favour of main_channel_indices (#4624)

• “Allow” for 3D channel locations for compute unit locs and template metrics (#4572)

• Vectorize per-channel PCA transform in run_for_all_spikes (#4488)

metrics:

• Correct and speed up multi-segment synchrony metrics (#4770)

• Fix PCA metric recomputation when “soft” merging (#4748)

• Add template_operator to template metrics backwards compatibility (#4714)

• Deepcopy of metric params (#4688)

• Fix Llobet expected-violations formula in sliding RP metric (#4682)

• QualityMetrics: store use_valid_periods param (#4596)

• Amplitude cutoff should favor amplitude_scalings extension (#4513)

• Harmonization of the way we are computing sliding_rp_violations (#4324)

curation:

• Fix mismatch in vairable name in mergeunitssorting.py (#4743)

• Fix is_merged/is_split properties for different strategies (#4718)

• Model_based_curation + unitrefine prediction: add unit_ids input (#4716)

• Vectorize _find_duplicated_spikes_keep_first/last_iterative kernels. (#4614)

• Bombcell extension (#4494)

• Change to using upset-bombcell package (#4493)

• Make sure default automerge params reflect their type (#4483)

widgets:

• Fix unit presence time_range (#4571)

• Add sort_by_depth kwargs to RasterWidget (#4544)

generators:

• Fix amplitude being ignored for static hybrid recordings (#4674)

• Add interpolation_kwargs to interpolate_templates (#4665)

• Small change to add amplitude jitter in drifting_generator (#4592)

• Use mock instead of generate in tests (#4587)

• Generate recording with probe from library (#4568)

• generate: Add power decay spatial profile (#4489)

sortingcomponents:

• Only use nonempty main channel indices when making an analyzer from templates (#4763)

• Enhance single channel denoiser and WaveformNode (#4673)

• Add thin_plate interpolation to interpolate_templates (#4640)

• Add install check for components sorters (#4574)

• Small fix in nodepipeline that prevent tdc peeler in drift-aware mode (#4557)

• Patch to improve the nearest peeler. (#4516)

motion correction:

• Allow Dredge AP to use online motion resolution method (#4746)

• Use MEDiCINe’s own recommended default for num_depth_bins (#4686)

documentation:

• Update copyright in conf.py (#4642)

• Update release dates and add 0.104.4 (#4613)

• Fix doc typo on home page (#4492)

continuous integration:

• Use a single dataset cache for all the runner OSes (#4772)

• Skip pruning of macOS uv cache (#4678)

• Add cross-version serialization tests (#4636)

• Update actions versions (#4585)

• Modularize CI test and runtime dependencies per module (#4541)

• Fix numpy and scipy failures on the CI (#4534)

• Loosen import-time gate on Windows and unpin pytest (#4532)

• Use uv and caching in CI (#4515)

packaging:

• Relax scikit-learn<1.8 pin (#4780)

• Dev version should be 0.105.0 (#4696)

• Relax pandas<3 pin (#4691)

• Fix imports in sortingview (#4652)

• Cleanup importing spikeinterface rather than inspecting it during curation (#4651)

• Update checking for editable mode of spikeinterface (#4650)

• Another round of import edits to reduce memory burden and increase speed of startup (#4644)

• Updating Imports (details in comment) (#4641)

• Use probeinterface and neo source when using spikeinterface source (#4615)

• Use legacy numpy versions for ks 4.1.1-4.1.6 tests (#4591)

• Add install check for components sorters (#4574)

• Fix module level test groups (#4573)

• Modularize CI test and runtime dependencies per module (#4541)

testing:

• Fix failing tests: decimate in chunks and skops (#4792)

• Use a single dataset cache for all the runner OSes (#4772)

• Fix interpolate tests after probeinterface to_numpy change (#4637)

• Fix curation tests for missing kachery dependency (#4599)

• Use legacy numpy versions for ks 4.1.1-4.1.6 tests (#4591)

• Fix module level test groups (#4573)

• Modularize CI test and runtime dependencies per module (#4541)

• Speed up test_sorting_s3_nwb_zarr (again) (#4538)

• Fix tests paths (#4490)

deprecations:

• Accept recording= consistently in detect_and_remove_bad_channels, depth_order and average_across_direction (#4779)

• Deprecation -> Future (#4667)

• Fix backward compatibility for binary recording/openephys/spikeglx (#4666)

• Segment backward compatibility (#4542)

Contributors:

• @4iar

• @Antorhythms

• @AtomicGlance

• @CodeBlackwell

• @JESUSROYETH

• @Julie-Fabre

• @Leonard013

• @RobertoDF

• @SZNRR

• @adityasingh2400

• @alejoe91

• @chrishalcrow

• @cwindolf

• @ecobost

• @galenlynch

• @grahamfindlay

• @h-mayorquin

• @hayleybounds

• @lochhh

• @nwatters01

• @oliche

• @rai-pranav

• @samuelgarcia

• @stanbot8

• @tayheau

• @yger

• @zm711