Researchers have introduced GRAFT, a new mechanism for fine-grained pronunciation in zero-shot text-to-speech systems. Developed by Antonis Asonitis and six co-authors, the paper was submitted on July 2, 2026, and aims to enhance the intelligibility of rare words and proper nouns in synthesized speech.
Understanding GRAFT's Mechanism
The GRAFT system provides a per-word pronunciation conditioning mechanism that allows text-to-speech neural models to utilize short spoken samples for accurate word pronunciation. Traditional systems often struggle with mispronouncing uncommon terms, but GRAFT addresses this issue effectively.
By encoding audio samples with the model's speech tokenizer and linking them to specific words in the text prompt, GRAFT ensures that the pronunciation is accurate. This method allows for voice conversion during training, which separates the hint speaker from the target speaker, enabling the use of diverse voices while maintaining output quality.
Performance and Benchmarking Results
In a recent blind English listening study, GRAFT was rated highest by human evaluators for its accuracy in pronouncing difficult words compared to reference recordings. The results showed that GRAFT reduced target-word phoneme error rates by 22-39% when benchmarked against traditional text-only systems.
Moreover, GRAFT outperformed several competitive open-source zero-shot systems, both phoneme- and text-conditioned, in terms of target-word pronunciation while preserving naturalness and speaker similarity.
Implications for Text-to-Speech Technology
The introduction of GRAFT could significantly impact various applications that rely on accurate speech synthesis, including virtual assistants, audiobooks, and language learning tools. As text-to-speech technology continues to evolve, GRAFT represents a step forward in addressing the challenges of pronunciation accuracy.
- Authors: Antonis Asonitis, Francesco Verdini, Aref Farhadipour, Vijeta Avijeet, Pierre-Edouard Honnet, Marzieh Razavi, Juan Pablo Zuluaga Gomez
- Submission Date: July 2, 2026
- Error Rate Reduction: 22-39%
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv Machine Learning. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
