AlephTTS-data / README.md
alejandro-hp's picture
Initial upload: AlephTTS voice dataset (CC0-1.0)
059c109 verified
|
Raw History Blame Contribute Delete
4.55 kB
metadata
license: cc0-1.0
language:
  - en
pretty_name: AlephTTS voice dataset
task_categories:
  - text-to-speech
tags:
  - audio
  - speech
  - single-speaker
  - tts
  - voice
size_categories:
  - n<1K

AlephTTS voice dataset

~47 minutes (854 clips) of studio-quality English speech by a single speaker, Aleph One-Half, recorded specifically to train a text-to-speech voice. This is the exact data used to fine-tune alejandro-hp/AlephTTS.

The recordings are dedicated to the public domain under CC0 1.0 by the speaker (see License for the one exception: the CMU ARCTIC prompt text).

Quick start

from datasets import load_dataset

ds = load_dataset("alejandro-hp/AlephTTS-data")
ex = ds["train"][0]
ex["audio"]["array"], ex["audio"]["sampling_rate"], ex["text"]

Contents

split clips duration
train 842 ~46.5 min
validation 12 ~0.7 min
source clips duration what
cmu_arctic 542 27.7 min phonetically balanced read sentences (CMU ARCTIC prompts)
conversational 173 11.4 min casual, talk-to-chat sentences
stream_lines 90 4.7 min stream-style lines with a delivery style (hype, greeting, reaction, calm, laugh, scared, ...)
custom 49 3.3 min spoken math explanations and lines containing the name "Aleph One-Half"

Fields

field description
audio FLAC, 48 kHz, 24-bit, mono
id clip id; prefix shows the source
text transcript of what was actually spoken (use this for training)
prompt the prompt the speaker read; differs from text for 5 clips where the reading differed
style delivery hint shown while recording (neutral, convo, explain, hype, chat, ...)
source cmu_arctic, conversational, stream_lines or custom
duration seconds

Numbers, math and symbols are written out as spoken words ("x squared plus y squared equals one").

How it was made

  • Recording: Shure MV7+ (USB), 48 kHz / 24-bit, quiet room (noise floor about −85 dBFS), consistent mic placement across 8 sessions. Each line was read from a prompt; the recorder auto-stopped after speech to avoid keyboard noise.
  • Quality control: every take was transcribed with Whisper large-v3 and compared to its prompt. Misreads were re-recorded, or the transcript was corrected to match the audio (5 clips, see prompt). Ambiguous takes were excluded. Takes with mid-phrase hesitations or unusually slow pace were re-recorded to favor fluent, natural delivery.
  • Processing: trimmed to speech with Silero VAD (about 0.12 s of padding before, 0.2 s after), with 10 ms fades, keyboard clicks in the padding muted, and one global gain for the whole set (no per-clip normalization, so loud and quiet styles keep their natural difference). Peaks at or below −1 dBFS; no clipping.
  • The validation split is the 12 clips held out from training for evaluation (8 neutral, 4 stream-style).

Intended use

Training and evaluating speech models (TTS, voice conversion, speech research), and reproducing AlephTTS.

A request from the speaker: CC0 places no legal restrictions on this data, but please do not use it to impersonate the speaker or anyone else, to deceive listeners about who is speaking, or for fraud or harassment, and please label synthetic speech made with it as AI-generated. Uses must also follow the Hugging Face content policy.

Consent: all audio was recorded by the speaker for this purpose and is published by the speaker.

Limitations

Single speaker, American English, under an hour. Mostly read speech; the expressive styles (hype, scared, laughing, ...) have only a few clips each. Many CMU ARCTIC sentences come from early-1900s fiction, with archaic vocabulary and names.

License

  • Audio (all clips), and the text of the conversational, stream_lines and custom transcripts: CC0 1.0 Universal, public domain dedication (see LICENSE).
  • Transcript text of the cmu_arctic clips: sentences from the CMU ARCTIC prompt list © Carnegie Mellon University, redistributed under CMU's permissive license, which requires keeping its notice (see NOTICE_CMU_ARCTIC.txt). Modified transcripts are marked by text ≠ prompt.