Advancing the science of human-like AI

GRAFT introduces a new approach that lets users guide the pronunciation of a specific word by providing a short spoken reference, addressing challenging cases such as rare names, loanwords, and technical terminology while preserving the target voice.

Across five languages, GRAFT reduced target-word phoneme error rates by 22–39% compared with the same text-only model, while maintaining speaker similarity and naturalness.

This research explores how the speaking rate of autoregressive text-to-speech models can be controlled through activation steering. Instead of retraining the model or introducing explicit conditioning, the approach modifies internal model representations to influence the pace of generated speech.

The work contributes toward more controllable speech generation, enabling AI voices to adapt how they speak while retaining the capabilities of the underlying model.

Modern text-to-speech systems are becoming increasingly natural, making reliable evaluation more challenging. This research examines the limitations of reference-free speech quality metrics when evaluating modern TTS systems and when using these metrics as reward signals for model optimization.

The work investigates established metrics including UTMOS, DNSMOS, and SCOREQ, identifying important limitations in individual metrics and exploring how calibrated combinations can provide a more robust evaluation signal.