Le 2025 Dataset

Dataset Source: figshare 29203457

Companion code (paper analyses): melizalab/auditory-restoration (also archived at Zenodo 15864044).

Citation:

Le, B.; Bjoring, M. C.; Meliza, C. D. (2025). The zebra finch auditory cortex
reconstructs occluded syllables in conspecific song. Nature Communications,
16:8452. https://doi.org/10.1038/s41467-025-63182-y

Dataset Details

Population fitting: ✅

Description of Stimuli:

  • 8 conspecific song motifs (“natural-syntax”, nat8a / nat8b stim sets) plus 8 scrambled-syntax pseudo-motifs (synth8b); each motif 837–1200 ms, high-pass filtered at 500 Hz, RMS-normalized to −27 dBFS (cohorts 1 & 2) or −30 dBFS (cohort 3).

  • 2 critical intervals (CI1, CI2) per motif, 58–100 ms duration each.

  • Up to 7 variants per CI, probing the auditory restoration illusion:

    code

    name

    content

    cohorts

    C

    Continuous

    Unmodified motif (shared across both CIs)

    all

    G

    Gap

    CI replaced by silence

    all

    N

    Noise

    CI-duration noise burst in isolation

    all

    GB

    Gap + Burst

    CI replaced by noise within the motif (illusion)

    all

    CB

    Continuous + Burst

    Motif unchanged, noise added on top of CI

    all

    GM

    Gap-Masked

    Whole motif masked by noise, CI deleted

    nat8b / synth8b

    CM

    Continuous-Masked

    Whole motif masked, CI intact

    nat8b / synth8b

  • Wav files at native sample rate (40 kHz nat8a, 48 kHz nat8b, 44.1 kHz synth8b).

  • 10 presentation repeats per (motif, variant) per neuron; some trials curated out → effective R ranges 4–10.

Description of Neurons:

  • Extracellular single-unit recordings from anesthetized adult zebra finch auditory pallium (NCM/L3, L2a/b, CM/L1; some sites unlabelled).

  • Three cohorts × two stim sets = four (cohort, stim-set) pairings:

    sub-experiment

    response dirs

    cohort(s)

    birds

    units

    nat8a

    nat8a-{alpha,beta}-responses/

    1 + 2

    14 + 10

    407 + 387

    nat8b

    nat8b-responses/

    3

    ~8

    445

    synth8b

    synth8b-responses/

    3

    ~5

    476

    Cohort 1 (nat8a-alpha) had a familiarity manipulation: half the birds were socially housed with half of the 8 stimulus singers before recording. Cohorts 2 and 3 are entirely unfamiliar (singers had died by then).

Spike sorting:

  • Cohort 1: MountainSort 4 + phy curation.

  • Cohorts 2 & 3: Kilosort 2.5 + phy curation.

  • Spike times stored per-unit as JSON pprox files. Two pprox schemas coexist: cohort 1 uses the legacy spec:2/pprox (spike times in ms, condition field encodes the variant); everything else uses spec:2/stimtrial (spike times in seconds, stimulus.name carries the full filename stem). The dataset class dispatches on $schema and presents a uniform interface to callers.

Processing needed (Dataset class __init__):

  • Wav → gammatone spectrogram (50 log-spaced bands, 1–8 kHz, 2.5 ms window, hop = dt_ms, log(P+1) compression). Matches the paper’s gammatone package recipe (Methods p. 10).

  • pprox → (R, T) PSTH per (stim, neuron), aligned to stim onset, binned at dt_ms.

  • Optional 21 ms Hanning smoothing (Hsu / Borst / Theunissen 2004 convention).

  • Optional per-neuron Sahani-Linden signal_power / noise_power / snr, length-weighted across stims (∼1 min for nat8b; skip with compute_reliability=False).

Benchmark results

Quick baseline established from examples/le_2025_baseline.ipynb: nat8b, dt_ms=5, 50-band gammatone, train on 6 motifs, val/test 1 each, Continuous variant only (8 unoccluded motifs, no occlusion).

Area

Model backbone

Rank

Remarks

Params / nrn

Perfs
(CCraw / CCnorm) [%]

Paper (backbone)

nat8b (all areas)

StateNet (GRU)

🥇

pop, 6-train

1,030

18.3 / 46.5 (median)

this notebook

This is a no-tuning baseline — proper LOOCV across all 8 motifs and a hyperparameter sweep are expected to push CCnorm appreciably higher.

Setup

The figshare archive (zipped 105 MB; unpacks to ~720 MB) is auto-downloaded by the dataset class:

from deepSTRF.datasets.audio import Le2025Dataset

ds = Le2025Dataset(experiment='nat8b', download=True)

Default cache dir is platformdirs.user_cache_dir('deepSTRF') / 'Le_2025', overridable via $DEEPSTRF_DATA_DIR. download=True is idempotent — already-unpacked trees are reused.

Manual install:

  1. Download zebf-auditory-restoration-1.zip from figshare 29203457.

  2. Unzip — the archive expands to a zebf-auditory-restoration-1/ directory containing metadata/, *-responses/, *-stimuli/.

  3. ds = Le2025Dataset(path='/path/to/zebf-auditory-restoration-1', experiment='nat8b').

gammatone>=1.0 is required for the paper-faithful spectrogram and is shipped as an optional extra:

pip install 'deepSTRF[le]'

Filtering

Each stim_meta dict carries name, motif, critical_interval, variant, syntax ("natural" / "scrambled"), experiment, sample_rate_hz, duration_s, ci_onset_s, ci_offset_s.

Each nrn_meta dict carries cell_id, animal_id, animal_uuid, cohort, experiment, area, hemisphere, familiar_motifs, sex, age_days, pprox_file, plus signal_power / noise_power / snr when compute_reliability=True (default).

Combined with the base-class selection API and the variant-aware helpers:

ds.select_variant('C')                       # only unoccluded motifs (8 stims)
ds.select_motif('nat8mk0')                   # all 12 variants of one motif
ds.select_critical_interval(1)               # CI=1 + the CI-independent C / CM
ds.select_pop_by_nrn_attr('cohort', 1)       # familiarity sub-experiment
ds.select_pop_by_nrn_attr('area', 'NCM/L3')  # one auditory subdivision

# The paper's core restoration analysis runs on a (motif, CI) × {C, CB, GB, GM} quartet:
ds.select_restoration_quartet('nat8mk0', 1)

Filter on the precomputed reliability stats to drop weakly responsive cells:

good = [n for n, meta in enumerate(ds.nrn_meta)
        if meta.get('snr', 0) > 0.1]
ds.select_population(good)

Cohort / sub-experiment layout

The handoff between experiment and on-disk directories:

zebf-auditory-restoration-1/
├── metadata/
│   ├── recordings.csv           area / hemisphere by site (cohorts 2 & 3)
│   ├── song-birds.csv           motif-name mapping (nat8a ↔ nat8b) + CI timings in ms
│   └── ephys-birds.csv          cohort / sex / age / familiarity-group by bird
├── nat8a-alpha-responses/       cohort 1 (familiarity, 407 units, legacy pprox)
├── nat8a-beta-responses/        cohort 2 (unfamiliar, 387 units, stimtrial pprox)
├── nat8a-stimuli/               shared by alpha + beta; bird-named motifs
├── nat8b-responses/             cohort 3 (445 units, natural syntax)
├── nat8b-stimuli/               same motifs as nat8a, renamed nat8mk0..7
├── synth8b-responses/           cohort 3 (476 units, scrambled syntax)
└── synth8b-stimuli/             synth8mk0..7

experiment='nat8a' unifies the alpha and beta response dirs (they share the same stimulus set); the per-unit cohort field distinguishes them. Cross-experiment concatenation works via the standard concat_neural_datasets:

nat = Le2025Dataset('...', experiment='nat8a')
syn = Le2025Dataset('...', experiment='synth8b')
both = nat + syn   # block-diagonal coverage matrix; bidirectional rule applies