SE-Bridge-TTS | ICML 2026

SE-Bridge-TTS is a low-resource Thai and Lao speech synthesis project. It studies a practical failure mode in spoken language models: synthetic data improves pronunciation stability, but too much flat synthetic speech erodes prosody and speaker expressivity. The release provides public Thai and Lao CosyVoice2-compatible checkpoints on Hugging Face.

DGSA uses the prosody-timbre separation in flow-matching SLMs to build preference pairs that reward stable, expressive speech without collapsing speaker identity.

TDSC samples candidates across conservative-to-expressive temperatures, filters them with automatic quality checks, and iteratively improves low-resource synthesis when real target-language speech is unavailable.

Paper accuracy is reported as Accuracy = 100 - WER, so higher is better. Higher SIM, NMOS, and SMOS are also better.

Selected demos are available on the project page, including Thai standard TTS, Lao standard TTS, Thai cloning, and Lao cloning.

We also publish a reproducible FLEURS benchmark comparing the released SE-Bridge-TTS checkpoints with three recent open multilingual TTS systems: Higgs Audio v3, OmniVoice, and X-Voice Stage1. The evaluation uses 255 paired Lao/Thai target sentences and reference prompts from Lao, Thai, Chinese, and English. It tests whether a model can synthesize Lao/Thai speech while preserving the prompt speaker.

Although SE-Bridge-TTS was developed before these newer open systems, it still reaches the best overall calibrated accuracy in the Chinese/English prompt -> Lao/Thai target setting, while remaining competitive on speaker similarity.

Accuracy keeps the CER signal but makes it easier to read:

calibrated CER = max(0, generated CER - ground-truth CER)
Accuracy = 1 - calibrated CER

Ground-truth CER is the ASR error measured on original FLEURS target audio, so the metric discounts recognizer baseline errors and focuses on synthesis degradation. Higher is better for Accuracy and speaker similarity.

Best results are bold. Second-best results are underlined.

Each detail cell below is Cal. CER↓ / SIM↑.

This public run covers same-language Lao/Thai prompts plus Chinese and English prompts to Lao/Thai targets. Unsupported Lao directions for X-Voice Stage1 are counted as coverage failures and excluded from quality averages.

The full protocol, machine-readable results, and table renderer are in evaluation/fleurs-lo-th-255pair.

The release checkpoints are hosted at:

https://huggingface.co/isabeth/SE-Bridge-TTS

For inference:

1. Open the Hugging Face model card above.
2. Download thai_tts.pt or lao_tts.pt from the model repository.
3. Follow the CosyVoice2 loading example in the model card.

This GitHub repository is intentionally lightweight: it hosts the project page, audio demos, paper links, and release pointers; the runnable checkpoint package lives on Hugging Face.

@inproceedings{geng2026bridging,
  title = {Bridging the Stability-Expressivity Gap: Synthetic Data Scaling and Preference Alignment for Low-Resource Spoken Language Models},
  author = {Geng, Yizhong and Li, Yanliang and Yang, Jinghan and Jiang, Tianhan and An, Boxun and Li, Ya and Shen, Xiaoyu},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year = {2026}
}

SE-Bridge-TTS thanks the following open-source projects:

• FunAudioLLM/CosyVoice, which provides the CosyVoice2 toolkit and inference stack used by the released checkpoints.
• facebook/mms-tts, a multilingual open-source TTS baseline included in the project demos.
• Higgs Audio v3, OmniVoice, and X-Voice, recent open multilingual speech generation systems referenced in the public FLEURS evaluation.

We thank the maintainers and contributors of these projects for making multilingual and low-resource speech synthesis research easier to reproduce, compare, and extend.