simba328's picture
Add files using upload-large-folder tool
4ccc648 verified
Raw
History Blame Contribute Delete
2.46 kB
"""
Example usage for the Designed Vocalizations Dataset.
pip install "datasets<4"
python example.py
Two configs — `raw` (source recordings) and `designed` (effect-processed clips) —
each with a `train` and `test` split.
TRAIN non-parallel: the raw and designed train pools are not aligned.
TEST parallel: evaluate the (source -> reference) pairs listed in
metadata/test_pairs.csv, each with a seen/unseen condition.
"""
import csv
from itertools import islice
from datasets import load_dataset
from huggingface_hub import hf_hub_download
REPO = "NCSOFT/Designed-Vocalizations-Dataset"
# ------------------------------------------------------------------- TRAIN
# Non-parallel: two independent pools, no pairing between them.
raw_train = load_dataset(REPO, "raw", split="train") # 5,654 sources
print(f"train: {len(raw_train)} raw sources")
# streamed just to keep this example light (drop streaming=True to load fully)
designed_train = load_dataset(REPO, "designed", split="train", streaming=True)
clip = next(iter(designed_train))
wav, sr = clip["audio"]["array"], clip["audio"]["sampling_rate"] # numpy waveform, 44100 Hz
print(f" designed (streamed): {clip['file_path']} | preset {clip['preset']}")
# -------------------------------------------------------------------- TEST
# Parallel: pair each source with its reference. Stream so only the test shards
# are fetched (test shares a config with the large designed/train split).
sources = load_dataset(REPO, "raw", split="test", streaming=True) # 120 inputs
references = load_dataset(REPO, "designed", split="test", streaming=True) # 5,640 refs
# the 120 sources are small — collect them for lookup by file_path
source_by_path = {clip["file_path"]: clip for clip in sources}
# seen/unseen condition per reference (from metadata/test_pairs.csv)
pairs_csv = hf_hub_download(REPO, "metadata/test_pairs.csv", repo_type="dataset")
seen_split = {row["reference"]: row["seen_split"]
for row in csv.DictReader(open(pairs_csv, encoding="utf-8"))}
for reference in islice(references, 3):
source = source_by_path[reference["source"]] # `source` column = paired input
src_wav = source["audio"]["array"]
ref_wav = reference["audio"]["array"]
print(f" [{seen_split[reference['file_path']]}] "
f"{reference['source']} -> {reference['file_path']}")
# ... compute your voice-conversion metric between src_wav and ref_wav ...