ADD22_eval_31 / _build_labels.py
korallll's picture
Labels-only ADD22 Track 3 R1 (audio not redistributed, CC BY-NC-ND 4.0)
33561dd
Raw
History Blame Contribute Delete
2.26 kB
"""Build data/labels.parquet for ADD22_eval_31 (labels-only) from the local protocol.
The audio is NOT redistributed (CC BY-NC-ND 4.0). The arena only ever needs the
labels: reproduce --scoring and validate-dataset --labels-only read this parquet
and never touch audio. This script reads ONLY the protocol text file (no audio
decode, never writes into the source dir).
utterance_id = the audio filename stem, e.g. "ADD_E3_00000000" (matches the
notes.utterance_id the original audio packaging used). label is int8 with the
package convention 0 = bonafide, 1 = spoof (higher score = more bonafide):
genuine -> 0 (bonafide), fake -> 1 (spoof).
"""
from pathlib import Path
import pyarrow as pa
import pyarrow.parquet as pq
SRC = Path("/home/kirill/mnt/users_4tb/datasets/add22track31test")
LABEL_FILE = SRC / "track3_R1_label.txt"
OUT = Path(__file__).resolve().parent / "data" / "labels.parquet"
LABEL_MAP = {"genuine": 0, "fake": 1} # 0 = bonafide, 1 = spoof
EXPECT_TOTAL = 112861
EXPECT_BONAFIDE = 20776
EXPECT_SPOOF = 92085
rows: dict[str, int] = {}
for line in LABEL_FILE.read_text().splitlines():
line = line.strip()
if not line:
continue
fname, raw = line.split()
stem = fname[:-4] if fname.endswith(".wav") else fname
if raw not in LABEL_MAP:
raise ValueError(f"unexpected label {raw!r} on line: {line!r}")
if stem in rows:
raise ValueError(f"duplicate utterance_id: {stem!r}")
rows[stem] = LABEL_MAP[raw]
# Deterministic order (sorted by utterance_id) so the parquet is reproducible.
uids = sorted(rows)
labels = [rows[u] for u in uids]
n_total = len(uids)
n_bonafide = labels.count(0)
n_spoof = labels.count(1)
print(f"total={n_total} bonafide={n_bonafide} spoof={n_spoof}")
assert n_total == EXPECT_TOTAL, (n_total, EXPECT_TOTAL)
assert n_bonafide == EXPECT_BONAFIDE, (n_bonafide, EXPECT_BONAFIDE)
assert n_spoof == EXPECT_SPOOF, (n_spoof, EXPECT_SPOOF)
schema = pa.schema([("utterance_id", pa.string()), ("label", pa.int8())])
table = pa.table(
{"utterance_id": pa.array(uids, pa.string()),
"label": pa.array(labels, pa.int8())},
schema=schema,
)
OUT.parent.mkdir(parents=True, exist_ok=True)
pq.write_table(table, OUT)
print(f"wrote {OUT} ({OUT.stat().st_size/1e6:.2f} MB)")