babycry v1 — valence (positive vs distress)

A merged, embedding-ready corpus of 7,574 infant/child vocalization clips for a binary valence task:

• valence — positive vs distress: affective polarity of the vocalization (ambiguous clips are included in the data but excluded from the binary task).

Every clip ships with the raw audio (native sample rate) and three precomputed frozen-encoder embeddings (AST, wav2vec2, CLAP), so you can train a last-layer head with zero audio preprocessing, or work from the waveform directly.

ℹ️ An alert-gate target (attend vs ignore) was explored on this corpus and dropped for v1 (near-chance on this data). The revertible local snapshot + label_map.yaml still retain it, but it is not a column in this published dataset.

⚠️ Per-source licensing — no single combined license. Each clip keeps its source's license, recorded per row in the source_license column: ReCANVo CC-BY-4.0, DonateACry ODbL-1.0, ESC-50 CC-BY-NC-3.0. The ESC-50 subset (40 clips) is non-commercial; filter source_license != "CC-BY-NC-3.0" for a commercial-friendly subset (7,534 clips). See Licensing below.

⚠️ Research preview. ReCANVo speakers are ages 6–23, not infants (affect transfer, not infant data); DonateACry is 8 kHz vs 44.1 kHz for the others (sample-rate/domain confound). Read Caveats before drawing conclusions.

Intended use

• Train and compare last-layer heads (logistic regression, RF, small MLP) on frozen embeddings under leave-one-speaker-out evaluation.
• Probe valence (positive vs distress) and cross-dataset generalization.

Not intended for: clinical decisions; commercial deployment without removing the non-commercial subset; treating ReCANVo as infant data.

Sources

source_dataset	clips	original release	license	native SR
recanvo	7,077	ReCANVo, Zenodo record 5786859 · paper	CC-BY-4.0	44.1 kHz
donateacry	457	github.com/gveres/donateacry-corpus (donateacry_corpus_cleaned_and_updated_data/)	ODbL-1.0	8 kHz
esc50	40	github.com/karolpiczak/ESC-50 (crying_baby only)	CC-BY-NC-3.0	44.1 kHz

Considered but excluded (gated / not openly redistributable): BabbleCor (OSF rz4tx — audio + annotation tags require a signed Data Use Agreement), CRIED (ComParE 2018 EULA), Baby Chillanto (written permission). No clips from these are included.

Citation / BibTeX

@article{johnson2023recanvo,
  title   = {ReCANVo: A database of real-world communicative and affective nonverbal vocalizations},
  author  = {Johnson, Kristina T. and others},
  journal = {Scientific Data}, volume = {10}, year = {2023},
  doi     = {10.1038/s41597-023-02405-7}
}
@misc{donateacry,
  title = {DonateACry corpus}, howpublished = {\url{https://github.com/gveres/donateacry-corpus}}, note = {ODbL-1.0}
}
@inproceedings{piczak2015esc50,
  title = {ESC: Dataset for Environmental Sound Classification},
  author = {Piczak, Karol J.}, booktitle = {ACM Multimedia}, year = {2015}
}

Schema (column dictionary)

column	dtype	meaning
clip_id	string	globally unique id, "<source_dataset>:<filename-stem>"
source_dataset	string	recanvo | donateacry | esc50
orig_label	string	verbatim source label (e.g. dysregulated, hungry, crying_baby)
valence	string	positive | distress | ambiguous (see mapping)
speaker_id	string	speaker/group id (recanvo:P01; per-clip-unique for donateacry/esc50)
lobo_group	string	recommended CV grouping key (= speaker_id); group on this for LeaveOneGroupOut / GroupKFold
source_license	string	per-row license (CC-BY-4.0 / ODbL-1.0 / CC-BY-NC-3.0) — filter for commercial use
duration_s	float32	clip duration in seconds
orig_sr	int32	native sample rate of the stored audio (Hz)
emb_ast	list[768]	mean-pooled AST embedding
emb_wav2vec2	list[768]	mean-pooled wav2vec2 embedding
emb_clap	list[512]	mean-pooled CLAP embedding
audio	Audio (decode=True)	raw waveform; decodes to {array, sampling_rate} at native SR

The audio column is stored at native SR (no resampling). Decoding returns the native sampling_rate (== orig_sr). The audio bytes are embedded directly in the parquet shards (self-contained — no external files needed).

Decoding audio needs a backend (embeddings/metadata do not). datasets >= 4 decodes audio via torchcodec, which requires system FFmpeg (libav versions 4–8) installed. datasets < 4 (e.g. 3.6.x) decodes via soundfile and needs no FFmpeg. Every other column — emb_ast / emb_wav2vec2 / emb_clap and all metadata — loads on any datasets version; only ds[i]["audio"] decode needs the backend. If audio decode errors, either pip install "datasets<4" soundfile, or install FFmpeg and pip install torchcodec. To skip decoding entirely: ds = ds.cast_column("audio", datasets.Audio(decode=False)) (yields raw bytes/path).

Label definitions & mapping rationale

The verbatim source labels are mapped to valence via the single source of truth in data/datasets/v1_attend_ignore/label_map.yaml.

valence — positive vs distress (ambiguous is excluded from the binary task):

• positive ← ReCANVo delighted, laughter, laugh, happy, glee, affectionate
• distress ← ReCANVo frustrated, dysregulated, dysregulation-sick, dysregulation-bathroom; all DonateACry classes; ESC-50 crying_baby
• ambiguous ← all other ReCANVo labels (selftalk, social, request, yes, help, more, protest, bathroom, no, tablet, hunger, greeting)

An alert-gate target (attend vs ignore) was explored on this corpus and dropped for v1 (near-chance under leave-one-speaker-out). Its mapping is retained in the local label_map.yaml snapshot for reference but is not published here.

Sampling, coverage & evaluation

Class balance

task	classes (kept)	counts	dropped
valence	positive / distress	1,571 / 2,829 (4,400)	3,174 ambiguous

Per source: ReCANVo positive 1,571 · distress 2,332 · ambiguous 3,174; DonateACry 457 (all distress); ESC-50 40 (all distress).

Recommended protocol. Group by lobo_group (= speaker_id) with sklearn.model_selection.LeaveOneGroupOut or GroupKFold so no speaker is in both train and test. Use class_weight="balanced" and report balanced accuracy / macro-F1 / ROC-AUC, not raw accuracy. ReCANVo is the only multi-speaker source; DonateACry/ESC-50 are per-clip groups (each its own fold) and are single-class (all distress), best used for the cross-dataset distress test, not pooled clean training (see Caveats).

Feature extraction (reproducible)

Each clip is resampled per-encoder, sliced into 2.0 s windows with a 1.0 s hop, each window embedded with a frozen encoder (eval mode, no grad), then windows are mean-pooled to one vector per clip.

column	encoder model id	resample	dim
emb_ast	MIT/ast-finetuned-audioset-10-10-0.4593	16 kHz	768
emb_wav2vec2	facebook/wav2vec2-base	16 kHz	768
emb_clap	laion/clap-htsat-unfused	48 kHz	512

Resampling applies only to embedding extraction; the stored audio keeps its native SR. Reproduce with src/babycry/embed.py in the project repo.

Baselines (what to expect)

Leave-one-speaker-out, frozen embeddings + linear/RF probe (from results/):

• valence (positive vs distress). Pooled GroupKFold/5, CLAP/RF: balanced-acc 0.662, ROC-AUC 0.757 (CLAP/logreg 0.662 / 0.725; AST/RF 0.644 / 0.750). Within-ReCANVo (44.1 kHz, no SR confound) CLAP/RF: 0.673 / 0.740. CLAP zero-shot floor (no training): 0.615 / 0.648.

Quickstart

from datasets import load_dataset
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import LeaveOneGroupOut, cross_val_predict
from sklearn.metrics import balanced_accuracy_score, roc_auc_score

# Load locally (this directory) or from the Hub once published:
ds = load_dataset("parquet", data_files="data/train-*.parquet", split="train")

# --- VALENCE: positive vs distress, leave-one-speaker-out on CLAP embeddings ---
# Restrict to ReCANVo: 8 real speaker groups at 44.1 kHz = the CLEAN number.
# (Pooling DonateACry/ESC-50 inflates results -- they're all-distress and 8 kHz,
#  so LOBO exploits the sample-rate shortcut. See Caveats.)
keep = [v in ("positive", "distress") and d == "recanvo"
        for v, d in zip(ds["valence"], ds["source_dataset"])]
sub  = ds.select([i for i, k in enumerate(keep) if k])
X = np.array(sub["emb_clap"], dtype="float32")
y = np.array([1 if v == "distress" else 0 for v in sub["valence"]])
groups = np.array(sub["lobo_group"])                 # speaker grouping

clf = LogisticRegression(max_iter=2000, class_weight="balanced")
proba = cross_val_predict(clf, X, y, groups=groups,
                          cv=LeaveOneGroupOut(), method="predict_proba")[:, 1]
print("valence  bal-acc", balanced_accuracy_score(y, proba > 0.5),
      " ROC-AUC", roc_auc_score(y, proba))   # ~0.69 / ~0.76 (RF ROC-AUC ~0.74)

# --- AUDIO: decode a waveform at native sample rate ---
a = ds[0]["audio"]            # {"array": np.ndarray, "sampling_rate": int}
print(ds[0]["clip_id"], a["sampling_rate"], a["array"].shape)

# --- COMMERCIAL-FRIENDLY subset (drop CC-BY-NC ESC-50) ---
commercial = ds.filter(lambda r: r["source_license"] != "CC-BY-NC-3.0")  # 7,534 rows

A runnable, fuller example (logreg + RF heads, proper LOBO) is in examples/train_head.py in the project repo.

Provenance & build

Built from data/datasets/v1_attend_ignore/manifest.csv (immutable v1 snapshot) + data/emb/{ast,wav2vec2,clap}.npz via scripts/build_hf_dataset.py. To retarget labels, edit data/datasets/v1_attend_ignore/label_map.yaml and rebuild.

Licensing

This dataset asserts no single combined license — each clip is governed by its original source license, recorded per row in the source_license column:

source	clips	license	links
ReCANVo	7,077	CC-BY-4.0	license · source
DonateACry	457	ODbL-1.0	license · source
ESC-50	40	CC-BY-NC-3.0	license · source

The ESC-50 subset is non-commercial (CC-BY-NC-3.0). There is no blanket license over the whole collection — honor each source's terms individually. For a commercial-friendly subset, filter to source_license != "CC-BY-NC-3.0" (ReCANVo + DonateACry, 7,534 clips). Always cite the original datasets.