ElgekAI/SLED-TTS-Uyghur-UAS-0.2B

AopCode SLED-TTS Uyghur UAS 0.2B

An Arabic-script Uyghur (UAS) text-to-speech and zero-shot voice-cloning model released by AopCode. It is a 0.2B-parameter derivative of ICTNLP/SLED-TTS-Libriheavy.

This release contains the final offline SLED checkpoint, the extended UAS tokenizer, reproducible fine-tuning scripts, and a Gradio deployment app supporting CUDA, Apple Silicon MPS, and CPU.

Highlights

• Native Arabic-script Uyghur input; no UAS-to-Latin conversion at inference time.
• 24 kHz output using facebook/encodec_24khz.
• Reference-audio voice cloning with a matching reference transcript.
• CFG inference; 2.0 is a useful starting point.
• 208.4M total parameters, approximately 194.2M trainable SLED parameters.
• FP32 checkpoint size: approximately 777 MB. It can be loaded as FP16 on Apple Silicon/CUDA to reduce model-weight memory to approximately 390 MB.

Important responsible-use notice

Do not use this model to clone a person's voice without authorization, impersonate another person, deceive, defraud, harass, infringe rights, evade consent, or violate applicable law. Generated audio should be disclosed as synthetic where a listener could reasonably mistake it for a real person.

Read RESPONSIBLE_USE.md before use. The MIT license applies to the released code and weights; it does not grant rights to any third-party voice, recording, text, identity, trademark, publicity right, or training source.

Data statement

The fine-tuning corpus was assembled by AopCode from publicly available datasets and material accessible on the public internet, then filtered and normalized. The training audio and manifests are not included in this repository. Public accessibility does not necessarily imply permission for every downstream use. Users are responsible for confirming that their use, reference audio, target voice, generated content, and jurisdiction are lawful and appropriately authorized.

Model architecture

• Base: SLED-TTS LibriHeavy, MIT licensed.
• Text representation: RoBERTa byte-BPE vocabulary extended without changing existing token IDs.
• Added vocabulary: Arabic-script Uyghur characters; new rows were randomly initialized to avoid inheriting misleading English letter pronunciation priors.
• Acoustic representation: sum of the first eight Encodec 24 kHz embedding vectors, following SLED-TTS.
• Context length: 2048 positions.
• Training utterance limit in this release: 20 seconds.

Quick start

1. Install

Python 3.10–3.12 is recommended.

git clone https://github.com/ictnlp/SLED-TTS.git
cd SLED-TTS
pip install -e .
pip install gradio librosa soundfile huggingface_hub

Download this repository, then run the included app from the model directory:

huggingface-cli download sayiwen/AopCode-SLED-TTS-Uyghur-UAS-0.2B \
  --local-dir ./AopCode-SLED-TTS-Uyghur-UAS-0.2B

cp ./AopCode-SLED-TTS-Uyghur-UAS-0.2B/training/sled.py ./sled/sled.py

python ./AopCode-SLED-TTS-Uyghur-UAS-0.2B/app.py \
  --checkpoint ./AopCode-SLED-TTS-Uyghur-UAS-0.2B \
  --port 6006

Open http://127.0.0.1:6006.

Apple Silicon Mac

For M1/M2/M3/M4 Macs, the app automatically selects MPS and loads the SLED weights as FP16. Enable CPU fallback for any Metal operation that is not implemented:

export PYTORCH_ENABLE_MPS_FALLBACK=1
python app.py --checkpoint . --port 6006

A 16 GB Apple Silicon Mac is sufficient for inference. Intel Macs fall back to CPU and will be slower.

Input format

Provide all three fields:

1. A clean reference recording, preferably 3–9 seconds.
2. The exact Arabic-script Uyghur transcript of that recording.
3. The Arabic-script Uyghur text to synthesize.

Reference-text accuracy materially affects cloning quality. For long-form text, split at sentence boundaries and synthesize segments separately. This checkpoint was trained with utterances up to 20 seconds; very long single requests can become fast, unstable, or truncated.

Training summary

The final checkpoint was produced with FP32 master weights and BF16 compute on an NVIDIA RTX 5090 D:

1. Extend the tokenizer while preserving all original IDs.
2. Randomly initialize the added UAS rows and train only those rows for 500 steps at 5e-4.
3. Full-model UAS adaptation on approximately 225 hours for one epoch at 1e-5.
4. Low-learning-rate refinement on 219,508 filtered samples / 454.66 hours for three epochs at 5e-6, effective batch size 8.

Monitoring used a fixed 100-example held-out subset. Final monitoring loss was 10.9040; this is a training diagnostic, not a standardized perceptual benchmark. See TRAINING.md and training/ for the reproducible workflow.

Limitations

• Pronunciation, rhythm, and intelligibility are not guaranteed.
• Long single-pass text can be spoken too quickly; sentence-level segmentation is recommended.
• Reference-audio noise may be reproduced or amplified.
• The model may hallucinate, omit, repeat, or mispronounce content.
• This release has not been validated for safety-critical, forensic, biometric, medical, legal, or authentication uses.
• It must not be used as evidence that a person said something.

Attribution

Built by AopCode from the MIT-licensed SLED-TTS base model. Please cite the original SLED-TTS work:

@misc{ma2025efficientspeechlanguagemodeling,
  title={Efficient Speech Language Modeling via Energy Distance in Continuous Latent Space},
  author={Zhengrui Ma and Yang Feng and Chenze Shao and Fandong Meng and Jie Zhou and Min Zhang},
  year={2025},
  eprint={2505.13181},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2505.13181}
}