# CosyAccent: Duration-Controllable Accent Normalization Using Source-Synthesis Training Data

arXiv: https://arxiv.org/abs/2602.19166 · HTML: https://mlsanity.com/paper/2602.19166/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 22 of its 37 references that are in the map.
- First posted on arXiv 22 Feb 2026.

## Builds on (most cited first, 22 of 22)

- [Scalable Diffusion Models with Transformers](https://mlsanity.com/paper/2212.09748/) (2022)
- [Classifier-Free Diffusion Guidance](https://mlsanity.com/paper/2207.12598/) (2022)
- [Flow Matching for Generative Modeling](https://mlsanity.com/paper/2210.02747/) (2022)
- [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://mlsanity.com/paper/2104.09864/) (2021)
- [Robust Speech Recognition via Large-Scale Weak Supervision](https://mlsanity.com/paper/2212.04356/) (2022)
- [CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models](https://mlsanity.com/paper/2412.10117/) (2024)
- [LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpus](https://mlsanity.com/paper/2305.18802/) (2023)
- [Autoregressive Diffusion Transformer for Text-to-Speech Synthesis](https://mlsanity.com/paper/2406.05551/) (2024)
- [Textless Speech Emotion Conversion using Discrete and Decomposed Representations](https://mlsanity.com/paper/2111.07402/) (2021)
- [Any-to-One Sequence-to-Sequence Voice Conversion using Self-Supervised Discrete Speech Representations](https://mlsanity.com/paper/2010.12231/) (2020)
- [Voice-preserving Zero-shot Multiple Accent Conversion](https://mlsanity.com/paper/2211.13282/) (2022)
- [HiFTNet: A Fast High-Quality Neural Vocoder with Harmonic-plus-Noise Filter and Inverse Short Time Fourier Transform](https://mlsanity.com/paper/2309.09493/) (2023)
- [Convert and Speak: Zero-shot Accent Conversion with Minimum Supervision](https://mlsanity.com/paper/2408.10096/) (2024)
- [TTS-Guided Training for Accent Conversion Without Parallel Data](https://mlsanity.com/paper/2212.10204/) (2022)
- [High-Fidelity Neural Phonetic Posteriorgrams](https://mlsanity.com/paper/2402.17735/) (2024)
- [DurFlex-EVC: Duration-Flexible Emotional Voice Conversion Leveraging Discrete Representations without Text Alignment](https://mlsanity.com/paper/2401.08095/) (2024)
- [MacST: Multi-Accent Speech Synthesis via Text Transliteration for Accent Conversion](https://mlsanity.com/paper/2409.09352/) (2024)
- [Improving Accent Conversion with Reference Encoder and End-To-End Text-To-Speech](https://mlsanity.com/paper/2005.09271/) (2020)
- [Transfer the linguistic representations from TTS to accent conversion with non-parallel data](https://mlsanity.com/paper/2401.03538/) (2024)
- [Accent conversion using discrete units with parallel data synthesized from controllable accented TTS](https://mlsanity.com/paper/2410.03734/) (2024)
- [Improving Pronunciation and Accent Conversion through Knowledge Distillation And Synthetic Ground-Truth from Native TTS](https://mlsanity.com/paper/2410.14997/) (2024)
- [Accent Normalization Using Self-Supervised Discrete Tokens with Non-Parallel Data](https://mlsanity.com/paper/2507.17735/) (2025)

## Built on by (most cited first, 2 of 2)

- [Controllable Accent Normalization via Discrete Diffusion](https://mlsanity.com/paper/2603.14275/) (2026)
- [TokAN: Accent Normalization Using Self-Supervised Speech Tokens](https://mlsanity.com/paper/2607.03928/) (2026)

## Abstract (arXiv)

Accent normalization (AN) systems often struggle with unnatural outputs and undesired content distortion, stemming from both suboptimal training data and rigid duration modeling. In this paper, we propose a "source-synthesis" methodology for training data construction. By generating source L2 speech and using authentic native speech as the training target, our approach avoids learning from TTS artifacts and, crucially, requires no real L2 data in training. Alongside this data strategy, we introduce CosyAccent, a non-autoregressive model that resolves the trade-off between prosodic naturalness and duration control. CosyAccent implicitly models rhythm for flexibility yet offers explicit control over total output duration. Experiments show that, despite being trained without any real L2 speech, CosyAccent achieves significantly improved content preservation and superior naturalness compared to strong baselines trained on real-world data.