Modelpublic

berrkkaayy/IndexTTS-2.5-Turkish-LoRA

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes

Model Card

IndexTTS-2.5 Turkish LoRA (step 29000)

Fine-tuned GPT weights that add Turkish (<|tr|>, language embedding row 9) to IndexTeam/IndexTTS-2.5 while preserving zero-shot voice cloning and the five base languages (Chinese, English, Japanese, Spanish, Arabic).

Training was done with index-tts-2.5-ft (itts25ft).

Model Details

Field	Value
Base model	IndexTeam/IndexTTS-2.5
Fine-tuning repo	oberkayb/index-tts-2.5-ft
Method	LoRA (rank 32, alpha 64) + trainable `lang_embedding` row 9
Language init	`--lang-init-from es` (Spanish row as seed)
Checkpoint	step 29000 (~12.2 epochs, effective batch 36)
Trainable params	~12.5% of GPT (LoRA adapters + lang row + text/mel heads)
Vocab / BPE	Unchanged — no tokenizer surgery

Files in This Repository

File	Description
`gpt.pth`	Merged LoRA export; drop-in replacement for `checkpoints/gpt.pth`
`export_metadata.json`	Step, epoch, and export provenance

All other weights (s2mel.pth, codec.pth, vocoder, BPE, etc.) come from the base IndexTeam/IndexTTS-2.5 checkpoint.

Quick Start

1. Install upstream IndexTTS-2.5

bash

git clone https://github.com/index-tts/index-tts.git && cd index-tts
pip install -U uv
uv sync --all-extras
hf download IndexTeam/IndexTTS-2.5 --local-dir=checkpoints

2. Replace GPT weights

bash

# back up the stock file first
cp checkpoints/gpt.pth checkpoints/gpt.pth.stock
cp /path/to/this/repo/gpt.pth checkpoints/gpt.pth

3. Synthesize Turkish (recommended path)

Clone the fine-tuning repo for the Turkish text frontend (turkish normalizer + tr_lower casing). Training and inference must use the same frontend.

bash

git clone https://github.com/oberkayb/index-tts-2.5-ft.git
cd index-tts-2.5-ft

uv run --project ../index-tts python scripts/synthesize.py \
  --gpt-checkpoint /path/to/gpt.pth \
  --lang tr --normalizer turkish --case tr_lower \
  --prompt-audio ref.wav \
  --text "Merhaba, bugün hava çok güzel." \
  --output out/tr_test.wav

Or load via IndexTTS2 after replacing checkpoints/gpt.pth, but you must still apply the Turkish frontend before calling infer(..., lang="tr", text_normalization=False).

Cross-lingual check

bash

uv run --project ../index-tts python scripts/synthesize.py \
  --gpt-checkpoint /path/to/gpt.pth \
  --lang tr --normalizer turkish --case tr_lower \
  --prompt-audio ref.wav \
  --cross-lingual-check \
  --output out/check.wav

Training Summary

Item	Value
Utterances	43,292 (~856 speakers)
Train pairs	85,824
Preprocess	`turkish` + `tr_lower`, 1–20 s clips
Learning rate	1e-4 (lang row ×10)
Precision	bf16
Target epochs	12

See export_metadata.json for checkpoint-specific fields.

Limitations

—Turkish requires the itts25ft text frontend; stock webui.py does not apply turkish + tr_lower normalization.
—Long text is segmented; prosody may not carry across segment boundaries (same as base model).
—Occasional word substitutions under high temperature; lower temperature for production use.
—Subject to the bilibili Model Use License Agreement inherited from IndexTTS-2.5.

Citation

If you use this checkpoint, please cite the base IndexTTS-2.5 paper and link this repository:

bibtex

@misc{li2026indextts25technicalreport,
  title={IndexTTS 2.5 Technical Report},
  author={Yunpei Li and others},
  year={2026},
  eprint={2601.03888},
  archivePrefix={arXiv},
}

Fine-tuning pipeline: https://github.com/oberkayb/index-tts-2.5-ft