Antalia 1 — open Turkish text-to-speech

Development of this model is discontinued. This repository contains the training, inference, and evaluation code for Antalia 1, a ~300M-parameter Turkish TTS model built on one consenting speaker's recordings, published alongside the foundation model, the evaluation suites, and a technical report. It is archived here so the parts that worked are not lost.

• Weights: https://huggingface.co/cloud0day3/antalia-1
• Foundation: https://huggingface.co/cloud0day3/antalia-1-foundation
• Voice corpus (5.008 h, CC-BY-4.0): https://huggingface.co/datasets/cloud0day3/antalia-voice-corpus
• Evaluation data: https://huggingface.co/datasets/cloud0day3/antalia-eval
• Audio samples: https://0daycloud.github.io/antalia/
• Paper: https://github.com/0daycloud/antalia/blob/main/paper/main.pdf
• Contact: sezgin@patientdesk.ai or GitHub issues

What it is

Antalia 1 is a CrossFlow rectified-flow acoustic model over 24 kHz, 100-band log-mel frames, distributed with a self-describing release format (config.json + model.safetensors). It speaks Turkish in one voice: that of a professional Turkish voice actor who consented to open redistribution of the weights. There is no zero-shot cloning; speaker id 0 is the unconditioned foundation path.

The voice is labelled "Voice B" in prose and voicedata-candidate-b in speaker identifiers, run names, and config filenames. Both are internal labels; the person behind the voice is credited anonymously at her request. Her recordings are released: the 5.008 h the model was fine-tuned on are published as antalia-voice-corpus under CC-BY-4.0, under an addendum that covers redistribution of the audio itself.

On our 120-prompt Turkish evaluation suite (Whisper-large-v3, WavLM-x-vector similarity):

Configuration	CER mean	Speaker sim. mean
Single seed, 32 steps	0.0528	0.9331
Best-of-8, timbre-gated	0.0298	0.9445

Long inputs are chunked at ≤120 characters; numbers, foreign names, and normalization-heavy text have the highest error rates; and automated speaker similarity of 0.93-0.94 does not certify identity — a native listener rated 0/12 synthesized-vs-real pairs as the same person. See the paper.

Install

uv sync            # or: pip install -e .

Python 3.12; CUDA GPU recommended (CPU works slowly). The vocoder is not in the package: clone NVIDIA BigVGAN and apply the one-line hub compatibility patch:

git clone https://github.com/NVIDIA/BigVGAN /opt/bigvgan
git -C /opt/bigvgan checkout 7d2b454564a6c7d014227f635b7423881f14bdac
patch -d /opt/bigvgan -p4 < scripts/patches/bigvgan-huggingface-hub-1.patch
export PYTHONPATH=/opt/bigvgan

Synthesize

python scripts/synthesize-crossflow.py \
  --checkpoint cloud0day3/antalia-1 \
  --vocoder nvidia/bigvgan_v2_24khz_100band_256x \
  --speaker voicedata-candidate-b \
  --prosody-presets prosody-presets.json --preset warm_voice_agent \
  --text-guidance 4.0 --sway -0.8 --steps 32 --mel-clamp 5.0 \
  --min-seconds-per-char 0.085 --chunk-chars 120 \
  --text "Merhaba, ben Antalia." --output out.wav

--checkpoint accepts a Hub repo id, a local release directory, or a training .pt file. prosody-presets.json ships in the weights repo next to inference-recipe.json, which records the full champion recipe.

Evaluate

python scripts/synthesize-crossflow.py \
  --checkpoint cloud0day3/antalia-1 \
  --evaluation-suite configs/evaluation/turkish-v2.jsonl \
  --seed-candidates 8 \
  --output-dir runs/analysis
python scripts/select-best-of-n.py \
  runs/analysis/crossflow.candidates.json \
  --envelope-stats envelope-stats.json \
  --timbre-stats timbre-profile.json
antalia baseline evaluate \
  --synthesis-report runs/analysis/crossflow.synthesis.json \
  --reference-manifest /path/to/reference.jsonl \
  --output runs/analysis/quality.json

Train / lineage

The six-stage training lineage is preserved in configs/crossflow/:

foundation-v3-normalized.json          (100,000 updates, all params)
foundation-speaker-v2.json             (3,000 updates, speaker embeddings)
candidate-b-speaker-v1.json            (100 updates, one new speaker embedding)
cfg-foundation-v1.json                 (8,000 updates, CFG calibration)
candidate-consistency-v1.json          (6,000 updates, replay-heavy fine-tune)
candidate-b-timbre-adapter-v2.json     (1,400 updates, released)

Repository layout

• src/turkish_tts/ — model, training, inference, evaluation, normalizer, release format
• scripts/ — synthesis, best-of-N selection, export, dataset/manifest builders, sweep shells
• configs/ — training, inference, evaluation, campaign manifest configurations
• reports/ — CMOS results, failure analysis, found-audio pilot, campaign brief
• docs/lessons.md — what measurably improved the model, with before/after numbers

Limitations

1. Voice similarity gap (above).
2. Long inputs need chunking.
3. Numbers, normalization, foreign names — weakest categories.
4. Single-seed variance is large; headline numbers depend on best-of-8.
5. One fixed voice, Turkish only, 24 kHz, grapheme input.
6. No memorization audit was run; no audio watermark is embedded.

License

Code: Apache-2.0. Model weights: Antalia Open RAIL-M — no impersonation, fraud, deception, political robocalls, or use without AI disclosure, and anything that uses or redistributes the weights must credit "Antalia 1" by Sezgin Saygili, Emre Kaplaner, Oncel Ozgul and Fikri San Koktas (Patientdesk.ai) with a link to the model or code repository. See the HF model cards.

Contributing

Community pull requests are accepted. There is no roadmap and no support SLA.

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://github.com/0daycloud/antalia}
}