anime-voices

Speaker-labelled speech for zero-shot TTS and voice cloning. Every example is a single-speaker utterance with its transcript, plus a reference prompt: a different clip of the same speaker.

utterances	1011
hours	2.335
speakers	86
sources	45
sample rate	24000 Hz

series	hours	utterances	speakers	sources
blackclover	1.07	464	48	20
rentagf	0.53	228	12	12
smoking	0.52	235	13	10
naruto	0.22	84	13	3

Audio is stored in Parquet shards under data/ (one per source) with audio and ref_audio columns. Speaker IDs are stable across updates (speaker_registry.json), speakers.csv summarises the speakers, metadata.jsonl holds per-utterance metadata and quality metrics, and manifest.json records the update history.

from datasets import load_dataset

ds = load_dataset("backpropSukuna/anime-voices", split="train")
row = ds[0]
row["audio"]       # {"array": ..., "sampling_rate": 24000}: the utterance
row["ref_audio"]   # a different clip of the same speaker (voice prompt)
row["text"], row["speaker_id"], row["ref_text"]