CosyAccent

Official implementation of the paper: "CosyAccent: Duration-Controllable Accent Normalization Using Source-Synthesis Training Data" (Accepted to ICASSP 2026)

Paper: https://arxiv.org/abs/2602.19166v1

Dataset

The training data (L2-LibriTTSR) is hosted on Hugging Face: https://huggingface.co/datasets/Piping/L2-LibriTTSR

Inference Quick Start

Install inference dependencies:

pip install -r requirements.txt

Run wav-to-wav inference:

python infer_wav.py \
  --source_wav /path/to/source.wav \
  --output_wav outputs/result.wav

Model selection is done by --model_tag:

• default: emilia_pretrained
• paper setting: paper

Example with paper setting:

python infer_wav.py \
  --source_wav /path/to/source.wav \
  --output_wav outputs/result.wav \
  --model_tag paper

• --reference_wav (default: None -> source_wav; used for timbre conditioning)
• --n_timesteps (default: 32; flow-matching steps)
• --full_cfg (default: 1.0)
• --cond_cfg (default: 1.0)
• --spk_cfg (default: 0.0)
• --preserve_total_duration (flag; disabled by default)
• --speech_len_ratio (default: 1.0; used only when --preserve_total_duration is set)
• --device (auto/cpu/cuda, default: auto)
• --mel_path (optional output path to save predicted mel tensor)
• --hf_repo_id, --hf_subfolder, --model_filename, --config_filename, --hift_filename, --hf_revision, --hf_token (optional advanced Hugging Face download settings)
• --model_checkpoint, --model_config, --hift_checkpoint (optional local paths; when --model_checkpoint is set, --model_config is required)

infer_wav.py auto-downloads required checkpoints from Hugging Face if local checkpoint paths are not provided.

Acknowledgements

This repository builds on several excellent open-source projects:

• CosyVoice: the HiFT checkpoint is leveraged from CosyVoice2.
• ESPnet: the Conformer/Transformer encoder components are adapted from ESPnet (via CosyVoice and direct modifications in this repo).
• OpenAI Whisper: Whisper encoder frontend is used for speech feature extraction.

Citation

@inproceedings{bai2026cosyaccent,
  title={CosyAccent: Duration-Controllable Accent Normalization Using Source-Synthesis Training Data},
  author={Bai, Qibing and Shi, Shuhao and Wang, Shuai and Ju, Yukai and Wang, Yannan and Li, Haizhou},
  booktitle={ICASSP 2026},
  year={2026}
}