from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
import torch

# Load model
config = XttsConfig()
config.load_json("config.json")

model = Xtts.init_from_config(config)
model.load_checkpoint(
    config,
    checkpoint_path="model.pth",
    vocab_path="vocab.json",
    use_deepspeed=False,
)
model.cuda()
model.eval()

# Generate speech
# Text should be in Cyrillic script
text = "Селям алейкум! Бу къырымтатар тилинде сёйлеген модель."

gpt_cond_latent, speaker_embedding = model.get_conditioning_latents(
    audio_path="reference.wav"
)

out = model.inference(
    text=text,
    language="ru",  # Use "ru" for Cyrillic text
    gpt_cond_latent=gpt_cond_latent,
    speaker_embedding=speaker_embedding,
)

# Save audio
import torchaudio
wav = torch.tensor(out["wav"]).unsqueeze(0)
torchaudio.save("output.wav", wav, 24000)

Text Format

The model expects text in Cyrillic script. If your text is in Latin script, convert it first:

Latin	Cyrillic
a	а
ğ	гъ
q	къ
ñ	нъ
j	ж
ş	ш
ç	ч

Training Details

Dataset: Sevil voice recordings (~7 hours)
Epochs: Averaged weights from epochs 4 and 7 (50/50)
Batch Size: 2
Learning Rate: 5e-6
Optimizer: AdamW

Limitations

Best results with sentences under 200 characters
Some phonemes specific to Crimean Tatar may vary in quality
Works best with reference audio from the same speaker

Our another resources

Crimeantatar TTS model based on SpeechT5 Sevil model on SpeechT5
Fixed Sevil dataset

Citation

bibtex

@misc{xtts-crh-sevil-v1,
  author = {Servin Osmanov},
  title = {XTTS Crimean Tatar - Sevil Voice},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/servinosmanov/xtts-crh-sevil-v1}
}

License

Apache 2.0