Snooow1029/itts-controllability

CommonVoice ITTS Controllability Validation Set

Ground-truth-labelled English speech clips drawn from fixie-ai/common_voice_17_0 (Common Voice 17.0, CC0), used to validate audio-LM judges for instructable-TTS (ITTS) controllability scoring on age, gender, and native accent.

Each clip carries a human-provided attribute label from Common Voice, so a judge's perceived-attribute accuracy can be measured against real ground truth (blind perceived-accuracy protocol).

Contents

• data/*.mp3 — 198 source clips (~108 MB).
• metadata.jsonl — 271 rows, one per (clip, attribute).

Cells (30 per value unless noted):

attribute	values
age	young_adult, middle_aged, elderly
gender	female, male, non_binary (1)
native_accent	american, australian, british, indian

Age bands: young_adult 13–39 / middle_aged 40–59 / elderly 60+.

Fields

• file_name — relative path to the audio (AudioFolder convention).
• id — <clip>-<attribute>.
• attribute — one of age / gender / native_accent.
• requested_label — ground-truth Common Voice label for that attribute.
• text — reference transcript.
• language — en.
• source — common_voice_17_0.

A clip may appear under more than one attribute (same speaker labelled for age, gender, and accent).

Usage

from datasets import load_dataset
ds = load_dataset("audiofolder", data_dir="<local_dir>", split="train")

License

Audio is Common Voice 17.0 (CC0-1.0); labels derive from Common Voice contributor metadata. See the upstream dataset for details.