You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access is restricted and inherited from the upstream Speech Accessibility Project data use agreement. Recordings and participant metadata may only be used for research. Do not redistribute.

Log in or Sign Up to review the conditions and access this dataset content.

SAPC1

Speech Accessibility Project corpus, 2025-03-31 release, repackaged for training. Audio is 16 kHz mono PCM_16. Speakers have a range of etiologies (Parkinson's disease, ALS, cerebral palsy, Down syndrome, stroke).

Repacked from yangwang825/sa2025, which ships each split as multi-part 21 GB archives. Here the same audio is one tar per contributor, so a download resumes cheaply and a single speaker can be fetched on its own.

Access is gated and covered by the upstream Speech Accessibility Project data use agreement. The metadata includes participant diagnoses and consent statements. Do not redistribute.

Layout

Train/<contributor-id>.tar         580 archives, 218,900 utterances
Dev/<contributor-id>.tar            83 archives,  31,114 utterances
metadata/Train/<contributor-id>.json
metadata/Dev/<contributor-id>.json
sapc1_train.jsonl
sapc1_dev.jsonl

Each archive expands to <contributor-id>/ holding that speaker's WAV files. Unlike SAPC2, the JSON metadata is not inside the audio archives — it lives under metadata/, mirroring how the upstream release ships it and matching the separate metadata_dir / audio_dir arguments of prepare_sap.py.

Download and extract

pip install "huggingface_hub[hf_transfer]"
export HF_HUB_ENABLE_HF_TRANSFER=1

CORPUS=/path/to/corpus/sapc1
hf download dys-asr/sapc1 --repo-type dataset --local-dir "$CORPUS"

for split in Train Dev; do
  find "$CORPUS/$split" -maxdepth 1 -name '*.tar' -print0 |
    xargs -0 -P 16 -I {} bash -c 'tar -xf "$1" -C "$(dirname "$1")" && rm -f "$1"' _ {}
done

One speaker only:

hf download dys-asr/sapc1 Dev/<contributor-id>.tar \
  --repo-type dataset --local-dir "$CORPUS"

Rebuilding the manifests

The audio here is already 16 kHz mono, so no resampling is needed:

python prepare_sap.py "$CORPUS/metadata/Train" "$CORPUS/Train" \
  sapc1_train.jsonl --workers 24 --include-metadata

Manifests

Each JSONL line is one utterance:

{"audio": "/abs/path/sapc1/Train/<id>/<id>_100_346.wav", "text": "...", "duration": 3.61, "metadata": {...}}

audio holds absolute paths from the machine that generated the manifest. Rewrite the prefix to your own corpus root before training:

OLD=/lus/lfs1aip2/scratch/u6sn/yangw.u6sn/corpus/audio/sapc1
python - "$OLD" "$CORPUS" <<'PY'
import json, sys
old, new = sys.argv[1], sys.argv[2]
for name in ("sapc1_train.jsonl", "sapc1_dev.jsonl"):
    with open(name) as src:
        rows = [json.loads(line) for line in src]
    for row in rows:
        row["audio"] = row["audio"].replace(old, new, 1)
    with open(name, "w") as dst:
        for row in rows:
            dst.write(json.dumps(row) + "\n")
PY

Then confirm every file resolved before starting a long job:

python -m ataxia.cli validate-manifest sapc1_train.jsonl sapc1_dev.jsonl \
  --check-audio --audio-check-workers 96

Related

Downloads last month
-