cloud0day3/antalia-voice-corpus

Antalia Turkish single-speaker scripted speech

5.008 hours / 1,073 clips of studio-quality Turkish read speech from one professional voice actor, script-aligned, segmented, and quality-gated. This is the corpus the Antalia 1 voice was fine-tuned on.

Clean, consented, single-speaker Turkish speech at this quality is scarce — which is the main reason this release exists. Development of the model is discontinued; the data is published as-is so it stays useful.

• —Model: https://huggingface.co/cloud0day3/antalia-1 · Foundation: https://huggingface.co/cloud0day3/antalia-1-foundation
• —Evaluation suites and filter manifests: https://huggingface.co/datasets/cloud0day3/antalia-eval
• —Code: https://github.com/0daycloud/antalia · Samples: https://0daycloud.github.io/antalia/
• —Report: https://github.com/0daycloud/antalia/blob/main/paper/main.pdf
• —Contact: sezgin@patientdesk.ai

Consent and who the speaker is

The speaker is a professional Turkish voice actor who recorded this material for the project under a commercial agreement, and who then signed an addendum permitting public redistribution of these recordings. She is credited anonymously at her request; the corpus carries a single pseudonymous speaker id (vd-speaker-e4440f86f3f6808056e3) and no personal identifiers of any kind.

Please treat the licence's attribution requirement and the ethical note below as binding in spirit as well as in law: a real person agreed to this, on the understanding that her voice would not be used to impersonate her or to pass synthetic speech off as a genuine recording.

Composition

All clips: 24 kHz, mono, 16-bit PCM WAV. Recorded at 48 kHz and downmixed to mono, resampled to 24 kHz, peak-limited to −1.0 dBFS, no gain applied. Six recording batches.

Signal quality: median estimated SNR 34.9 dB (minimum 28.5 dB), zero clipping. Every clip in this release passed the full quality gate — no segment carries a QA flag.

Fields

metadata.jsonl, one row per clip, file_name relative to the repository root.

from datasets import load_dataset
ds = load_dataset("cloud0day3/antalia-voice-corpus", split="train")
print(ds[0]["transcript"], ds[0]["duration_seconds"])

datasets v4 decodes audio through torchcodec, which needs FFmpeg shared libraries present; if that is awkward in your environment, read the files directly instead — the corpus is plain WAV plus one JSONL, and needs no special tooling:

import json, soundfile as sf
from huggingface_hub import hf_hub_download

repo = "cloud0day3/antalia-voice-corpus"
rows = [json.loads(l) for l in open(hf_hub_download(repo, "metadata.jsonl", repo_type="dataset"))]
audio, sr = sf.read(hf_hub_download(repo, rows[0]["file_name"], repo_type="dataset"))

How it was built

Scripts were authored per category, recorded batch by batch with a pilot-review gate before the remaining batches were unlocked. Each parent recording was force-aligned against its script with an independent ASR pass and split at word boundaries into 3–20 s segments (24 s absolute maximum, 0.12 s padding). A segment was accepted only if it met the script-alignment thresholds (≥0.7 exact-word ratio, ≤0.3 CER against the script, ≥0.45 mean word probability, ≥0.7 language probability) and the acoustic gates (RMS, clipping, SNR, active-audio ratio). Segments that failed were dropped, not repaired.

Scope, and what is not here

• —Only the training split is published (1,073 clips). A parent-disjoint held-out test split exists and is deliberately withheld, so the speaker-similarity figures reported for Antalia 1 are not independently reproducible from this release.
• —Nothing from the foundation corpus is here. Common Voice and FLEURS filter manifests live in antalia-eval; their audio is not re-hosted.
• —Numbers, reference codes, addresses and names in the scripts are fictional, authored for the campaign — the call-centre scenarios in readback_confirmation and numeric_normalization contain invented telephone numbers and reference codes by design, so that the model would learn to read them. Any resemblance to a real number or person is coincidental, and no name is attached to any of them.
• —Single speaker, single language, read speech. No spontaneous conversation, no other speakers.

Licence and acceptable use

Audio, transcripts and metadata: CC-BY-4.0. Attribute to "Antalia (Patientdesk.ai)" and link back to this repository.

The licence permits commercial use. The speaker's consent was given on the understanding that the following are out of bounds, and we ask you to honour it:

• —Do not use this audio, or anything trained on it, to impersonate the speaker or any real person, or to imply someone said something they did not.
• —Do not present synthetic speech derived from this corpus as a genuine human recording; disclose it where a listener could reasonably be misled.
• —Do not use it for fraud, scams, social engineering, or political robocalls.

The Antalia 1 model weights are licensed separately, under Open RAIL-M, which makes these restrictions binding rather than requested.

Citation

@misc{antalia1_2026,
  title  = {Antalia 1: An Open Turkish Text-to-Speech Model from a Rights-Clean Pipeline},
  author = {Saygili, Sezgin and Kaplaner, Emre and Ozgul, Oncel and Koktas, Fikri San},
  year   = {2026},
  note   = {Technical report},
  url    = {https://huggingface.co/datasets/cloud0day3/antalia-voice-corpus}
}

Türkçe veri kartı: README.tr.md