Teaching our Text-to-Speech to sound natural, every time | HappyRobot

Weixiao Xia

www.happyrobot.ai 5 pages Read the paper

The short version

The problem: A synthetic voice may sound natural in one take and awkward in the next, even when reading the same sentence. This makes it hard for teams using voice assistants to deliver a consistently fitting, clear performance.
The idea: Instead of only copying recordings, the voice tries several ways of saying a sentence and gets feedback on how natural each sounds and whether it said the right words. It then learns to use the better versions more often—like coaching a singer who already has the ability, but needs to perform well consistently. The training also limits how far the voice can drift and uses real listeners to choose the version to keep.
What it got: The paper reports that listeners preferred the trained voice, and an early live test showed fewer callers hanging up quickly, more positive sentiment, and more calls completing their task. It gives no numerical effect size for these improvements.

Modules from this paper

Preference Learning for Natural Speech

A speech model can say the words but deliver them awkwardly; generate several takes, score naturalness and word accuracy, and use group-relative feedback to make better deliveries more likely.

text-to-speechreinforcement-learningGRPOspeech-quality

Guardrails for Keeping Reinforcement-Trained Speech Trustworthy

Reinforcement training can push a speech model toward judge-pleasing quirks; a KL leash, word-error safeguard, and blind human checkpoint selection constrain that drift.

speechreinforcement-learningevaluationalignment