| --- |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| pipeline_tag: text-to-audio |
| tags: |
| - text-to-speech |
| - speech-refinement |
| - prosody |
| base_model: yhaha/EmoVoice |
| --- |
| |
| # LoopTTS Refiner 1.5B |
|
|
| This repository contains the **Refiner checkpoint only** for [LoopTTS (arXiv:2608.28970)](https://arxiv.org/abs/2608.28970). Given an input utterance, its transcript, and a natural-language instruction, the Refiner generates corrected speech. The Filter and Judge stages are outside this release. |
|
|
| ## Checkpoint |
|
|
| | File | Description | |
| |---|---| |
| | `model.pt` | Refiner 1.5B checkpoint, epoch 31 / step 7190 (6,408,626,894 bytes) | |
| | `config.yaml` | Portable summary of the evaluated inference configuration | |
| | `SHA256SUMS` | Checkpoint checksum | |
|
|
| SHA-256 of `model.pt`: `995185563a1ae25c2fcc57e0e06b7e196252f8a5745b521902166093ab38580f`. |
|
|
| The checkpoint was selected from the project's tested `TTS_refiner_v1_1.5B_from_FTEmoVoice_GTGlobal_useRAWprompt` run. The March 28, 2026 Refiner test script points to this epoch and step, and the adjacent inference log records the same checkpoint being loaded. It is fine-tuned from [EmoVoice 1.5B](https://huggingface.co/yhaha/EmoVoice). |
|
|
| ## Use |
|
|
| The [LoopTTS standalone inference instructions](https://github.com/Pooookeman/LoopTTS/tree/main/inference) provide one command that accepts an input WAV, transcript, and prosody instruction. The script implements the Refiner input packing and greedy decoding directly. It needs no EmoVoice checkout or patch; command-line paths are relative to the LoopTTS repository. |
|
|
| The runtime needs [Qwen2.5-1.5B](https://huggingface.co/Qwen/Qwen2.5-1.5B) and [CosyVoice-300M-SFT](https://huggingface.co/FunAudioLLM/CosyVoice-300M-SFT) in addition to this checkpoint. These are separate dependencies from their original publishers; they are not included in this repository. |
|
|
| Example instruction: `Speak with an angry emotion at a moderate speed and high pitch. Stress the word 'no'.` |
|
|
| ## Intended use and limitations |
|
|
| This model is provided for noncommercial research on controllable speech and prosody correction. It was evaluated with greedy decoding, three CosyVoice code layers, a 22,050 Hz output rate, and `raw_audio_position=end`. The released input preparation supports WAV utterances up to 25 seconds. The model may change words, speaker characteristics, or intended emotion; review outputs before using them in a study. |
|
|
| `model.pt` is a PyTorch checkpoint read with `torch.load`; load it only from this repository or another source you trust. The checkpoint alone is not a self-contained inference package. The standalone script has not yet been run against the released checkpoint, so output equivalence remains unverified. |
|
|
| ## License and attribution |
|
|
| The Refiner checkpoint is released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). The fine-tuning base is [EmoVoice](https://github.com/yanghaha0908/EmoVoice), whose authors state that their pretrained models are noncommercial. The LoopTTS Refiner inference code uses CC BY-NC 4.0; portions adapted from upstream EmoVoice retain their MIT attribution. Qwen and CosyVoice dependencies retain their own licenses. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{song2026loopt, |
| title = {Diagnose, Then Refine: A Closed-Loop TTS System with AudioLLM-Guided Correction}, |
| author = {Song, Zeyang and Liu, Tianchi and Wang, Tianrui and Xu, Chenglin and Guo, Yiwen and Li, Haizhou}, |
| booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing}, |
| year = {2026}, |
| eprint = {2608.28970}, |
| archivePrefix = {arXiv}, |
| url = {https://arxiv.org/abs/2608.28970} |
| } |
| ``` |
|
|
