TokAN: Accent Normalization Using Self-Supervised Speech Tokens

📰 ArXiv cs.AI

Learn how TokAN uses self-supervised speech tokens for accent normalization, improving speech quality without requiring parallel L1-L2 data

advanced Published 7 Jul 2026

Action Steps

Extract self-supervised discrete speech tokens from audio data using TokAN
Train a token-based accent normalization model with the extracted tokens
Evaluate the model's performance on a test dataset with diverse accents
Fine-tune the model for specific accent pairs or speaking styles
Integrate the TokAN framework into existing speech recognition systems for improved accuracy

Who Needs to Know This

Speech recognition engineers and researchers can benefit from this technique to improve accent normalization in their models, enhancing overall speech quality and speaker identity preservation

Key Insight

💡 Self-supervised speech tokens can effectively capture accent characteristics, enabling high-quality accent normalization without parallel data

Key Takeaways

Learn how TokAN uses self-supervised speech tokens for accent normalization, improving speech quality without requiring parallel L1-L2 data

Full Article

Title: TokAN: Accent Normalization Using Self-Supervised Speech Tokens

Abstract:
arXiv:2607.03928v1 Announce Type: cross Abstract: Accent normalization (AN) seeks to convert non-native (L2) accented speech into standard (L1) speech while preserving speaker identity. The current techniques either require naturally recorded parallel L1-L2 speech for training, or suffer from quality degradation when supervised by synthesized targets. In this paper, we present TokAN, a token-based accent normalization framework that operates on self-supervised discrete speech tokens extracted fr

Read full paper → ← Back to Reads