the-pickler/tts-synthetic
Dataset Card for 3-voice TTS (HiFiTTS-light style) Same column style as MikhailT/hifi-tts-light, built from three single-voice source sets. Original wav names were insignificant, so files were canonically reindexed per speaker. Speakers / shards speaker source rows hours audio handling tortoise tortise-merged (union of large-1 + large-2, orphans dropped) 4313 5.40 24 kHz float, copied as-is aincrad aincrad-synthetic clean metadata.csv (618 dirty-only… See the full description on the dataset page: https://huggingface.co/datasets/the-pickler/tts-synthetic.
Dataset Card for 3-voice TTS (HiFiTTS-light style)
Same column style as MikhailT/hifi-tts-light, built from three single-voice source sets. Original wav names were insignificant, so files were canonically reindexed per speaker.
Speakers / shards
Total: 14024 rows, ~18.9 h. Single split: clean/train (aka all/train.clean).
Columns
speaker(string):tortoise,aincrad, orsteffanfile(string): canonical path, e.g.audio/steffan_clean/0/000123.wavduration(float32): secondstext(string): lowercased, punctuation stripped except apostrophestext_no_preprocessing(string): original transcripttext_normalized(string): same as original (no digits/abbreviations in sources)audio(audio, 24000 Hz): embedded WAV bytes,audio.pathis the basename
Load
from datasets import load_dataset
ds = load_dataset("./hifitts-3voice", "clean", split="train")
# single voice: ds.filter(lambda r: r["speaker"] == "aincrad")