Audio8-TTS-Kurdish-Sorani

A Central Kurdish (Sorani) fine-tune of Audio8 TTS, a 0.6B-parameter DualAR text-to-speech model with zero-shot voice cloning. Neither Central Kurdish nor Persian is among the base model's supported languages. Training started from the Persian fine-tune ArminRahimi/Audio8-TTS-Persian, whose script Sorani largely shares.

Training code, Colab notebook and documentation: https://github.com/ard9/Audio8-Persian-TTS

Non-commercial use only. The training data, central-kurdish-tts4all, is licensed CC BY-NC-ND 4.0, so this model is released under CC BY-NC 4.0: free for research, education and non-commercial demos; not for paid products or services.

🔊 Samples

Text	Audio
سڵاو، ئەمڕۆ کەشوهەوا زۆر خۆشە و دەمەوێت کەمێک لە پارک پیاسە بکەم.	ft_s01.wav
دوێنێ لەگەڵ هاوڕێکانم چووین بۆ سینەما، فیلمەکە بەڕاستی زۆر جوان بوو!	ft_s03.wav
Voice clone (reference from another speaker of the dataset)	clone_s01.wav

Usage

The model uses Audio8's custom Transformers code (trust_remote_code=True). Normalize the text with the same Sorani rules as in training (normalize_ckb in the GitHub repo: Arabic letter variants unified, numbers spelled out in Kurdish).

# pip install "transformers>=4.57,<5" soundfile
# git clone https://github.com/ard9/Audio8-Persian-TTS  (for fa_tools.normalize_ckb)
import sys, torch, soundfile as sf
from transformers import AutoModel, AutoProcessor
sys.path.insert(0, "Audio8-Persian-TTS/src")
from fa_tools import normalize_ckb

model_id = "ArminRahimi/Audio8-TTS-Kurdish-Sorani"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(model_id, trust_remote_code=True, dtype=dtype).eval().to(device)

text = normalize_ckb("سڵاو! چۆنی؟")
inputs = processor(text=[text], return_tensors="pt")
# Voice cloning: add reference_audio=["ref.wav"], reference_text=[normalize_ckb("exact transcript")]
inputs = {k: v.to(device) for k, v in inputs.items()}
with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=1024, temperature=0.8, top_p=0.95, top_k=50,
                         do_sample=True, return_dict_in_generate=True)
    wav, lengths = model.decode_audio(out.codes)
sf.write("output.wav", wav[0, : int(lengths[0])].float().cpu().numpy(), model.config.codec_sample_rate)

Or from the command line: python scripts/synthesize.py --model ArminRahimi/Audio8-TTS-Kurdish-Sorani --lang ckb --text "سڵاو" --output out.wav.

Training details

Starting model	Audio8-TTS-Persian (itself fine-tuned from Audio8-TTS-Preview-0.6B, repo commit 07e40f5)
Data	central-kurdish-tts4all: ≈ 35 h of read Central Kurdish, 3 speakers (male and female audiobook narrators, male studio voice), 24 kHz, clips of 1–20 s
Default voice	The speaker with the most audio is trained without a reference clip; the other speakers always with a same-speaker reference, which teaches voice cloning
Text normalization	Sorani rules: Arabic yeh/kaf → Kurdish forms, ھ → ه, heh + ZWNJ → ە, diacritics and ZWNJ removed, numbers spelled out in Kurdish
Trainable parameters	Both DualAR branches (slow + fast AR), all 601M parameters
Optimizer	8-bit AdamW, lr 3e-5, cosine schedule, 3% warmup, grad-norm clip 1.0
Steps / epochs	TODO
Hardware	TODO GPU on Google Colab
Exported weights	bf16

Evaluation

Metric	Value
Eval loss (held-out, slow + fast AR)	TODO
CER, ASR transcription of generated speech	not measured yet
MOS, native listeners	not measured yet

Limitations

• Three speakers, read speech. Voice variety and prosody are limited; conversational speech and unseen voices are weaker.
• Possible Persian accent when starting from the Persian model, especially in early checkpoints.
• Spelling follows the dataset: the model learns pronunciation from how the transcripts are written; non-standard spellings and Latin-script words may be mispronounced.
• The base model is a Preview checkpoint without Kurdish or Persian pre-training.

Responsible use

This model can imitate voices. Clone a voice only with the speaker's explicit consent and disclose synthetic audio. Do not use it for impersonation, fraud or misinformation, or to imitate the dataset's speakers.

License and attribution

CC BY-NC 4.0 (non-commercial). Model code and architecture: Audio8 TTS (Apache-2.0). Training data: central-kurdish-tts4all by Aran Emini et al. (CC BY-NC-ND 4.0), including audiobook content used with the permission of Malaka Mustafa Soltani, read by Shahen Shahlai and Fatih.

Citation

@misc{audio8_persian_tts_2026,
  title        = {Audio8-Persian-TTS: Adapting a Compact Multilingual TTS Model to Persian},
  author       = {Rahimi Darehbagh, Armin},
  year         = {2026},
  howpublished = {\url{https://github.com/ard9/Audio8-Persian-TTS}}
}

Please also cite the central-kurdish-tts4all dataset as described on its dataset page.