A new accent normalization model rewrites non-native speech at the token level — without generating fake training audio first.
Researchers posted TokAN to arXiv (arXiv:2607.03928) to address a stubborn bottleneck: existing accent normalization systems either need paired recordings of the same speaker in both accented and standard English — rare and expensive to collect — or lean on synthesized speech as a training target, which degrades output quality. TokAN sidesteps both by working entirely in the token domain, using a jointly trained vector-quantization tokenizer to extract discrete speech tokens and an autoregressive encoder-decoder to rewrite L2-accented sequences into standard-voice tokens. A flow-matching synthesizer then reconstructs audio conditioned on the original speaker's voice embedding, preserving identity. The team added reinforcement learning post-training via Group Relative Policy Optimization (GRPO), using word error rate and accent-classifier confidence as complementary rewards.
According to the paper, TokAN reduced word error rates from 12.40% to 9.89% after supervised fine-tuning and to 9.23% after RL post-training, tested across seven English accents — outperforming frame-based, direct flow-matching, and prompt-based token-conversion baselines on both accent reduction and intelligibility. The system also ships a duration-aware synthesizer aimed squarely at voice dubbing and live casting, where timing constraints rule out approaches that stretch or compress audio unpredictably. Accent normalization sits at the commercial intersection of call-center AI, broadcast post-production, and accessibility tools, so interest from well-funded labs is a reasonable assumption.
The preprint has not been peer-reviewed, and speech AI benchmarks across seven curated accents are a friendlier test environment than the full range of real-world conditions — a gap worth watching if this moves toward deployment.