Audio8-TTS-Persian (v0.2)

A Persian (Farsi) fine-tune of Audio8-TTS-Preview-0.6B, a 0.6B-parameter DualAR text-to-speech model with zero-shot voice cloning. Persian is not among the base model's supported languages; to our knowledge this is the first public Persian adaptation of Audio8.

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

🔊 Samples

Text	Audio
سلام، امروز هوا خیلی خوب است و می‌خواهم کمی در پارک قدم بزنم.	ft_s01.wav
دیروز با دوستام رفتیم سینما، فیلمش واقعاً عالی بود!	ft_s03.wav
Voice clone (reference from a second speaker)	clone_s01.wav

Usage

The model uses Audio8's custom Transformers code (trust_remote_code=True). Normalize Persian text the same way as in training (digits → words, Arabic → Persian letters) with normalize_fa from the GitHub repo.

# pip install "transformers>=4.57,<5" soundfile num2fawords
# git clone https://github.com/ard9/Audio8-Persian-TTS  (for fa_tools.normalize_fa)
import sys, torch, soundfile as sf
from transformers import AutoModel, AutoProcessor
sys.path.insert(0, "Audio8-Persian-TTS/src")
from fa_tools import normalize_fa

model_id = "ArminRahimi/Audio8-TTS-Persian"
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_fa("سلام! حال شما چطور است؟")
inputs = processor(text=[text], return_tensors="pt")
# Voice cloning: add reference_audio=["ref.wav"], reference_text=[normalize_fa("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-Persian --text "سلام" --output out.wav.

Keep each input under ~150 characters; scripts/synthesize.py splits longer text into sentences automatically.

Training details

Base model	Audio8-TTS-Preview-0.6B (repo commit 07e40f5)
Data	33 of 77 Mana-TTS parts + GPTInformal-Persian: 31.6 h of audio after filtering (HIGH alignment quality only, 1–20 s clips)
Reference conditioning	~30% of Mana-TTS rows and 100% of GPTInformal rows trained with a same-speaker reference clip
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
Precision	fp32 master weights + fp16 autocast (T4)
Effective batch size	16 (batch 2 × gradient accumulation 8)
Steps / epochs	7000 / 2
Hardware	1× T4 on Google Colab, several sessions resumed from Hub checkpoints
Exported weights	bf16

Evaluation

Metric	Value
Eval loss (held-out, slow + fast AR)	8.00 at step 7000 (v0.1, trained on 10 parts: 8.80; different eval split)
Listening impression	Intelligible Persian speech; prosody still somewhat robotic
CER, Whisper transcription of generated speech	not measured yet
MOS, native listeners	not measured yet
Speaker similarity (voice cloning)	not measured yet

Limitations

• Mostly single-speaker data (31.6 h after filtering: 33 of 77 Mana-TTS parts from one speaker + GPTInformal). Voice variety and prosody are limited; cloning of unseen voices is weaker than in Audio8's supported languages.
• Limited compute (a single Colab T4); no hyperparameter search.
• No G2P: unwritten short vowels, homographs and the ezafe can be mispronounced.
• Dates, abbreviations and Latin-script words are not specially normalized.
• The base model is a Preview checkpoint without Persian pre-training.
• Prosody still sounds somewhat robotic.

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. The training data's authors prohibit using it to imitate their speakers' voices for malicious purposes.

License and attribution

Apache-2.0, as a derivative of Audio8 TTS (Apache-2.0; architecture inspired by Fish Audio S2 Pro's DualAR). Training data: Mana-TTS and GPTInformal-Persian (CC0-1.0).

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}}
}

@inproceedings{qharabagh-etal-2025-manatts,
  title     = {{M}ana{TTS} {P}ersian: A Recipe for Creating {TTS} Datasets for Lower-Resource Languages},
  author    = {Qharabagh, Mahta Fetrat and Dehghanian, Zahra and Rabiee, Hamid R.},
  booktitle = {Proceedings of NAACL 2025 (Volume 1: Long Papers)},
  year      = {2025},
  pages     = {9177--9206},
  url       = {https://aclanthology.org/2025.naacl-long.464/}
}