What makes a voice expressive?

A synthetic voice can become easier to transcribe while sounding less natural. That tension helps explain the question explored in Julia Turc's voice-evaluation video: how to assess expressive delivery as well as correct pronunciation. The video's description, which is all that was reviewed here, points to the study below.

That study tested speech on 50 samples drawn from Korean call-center conversations. Reinforcement training that rewarded accurate transcription improved transcription but reduced pitch variation, and listeners perceived the result as unnatural. The researchers then trained on human preferences between alternative readings to target natural delivery more directly. This specific experiment shows why a pronunciation score alone can miss changes in rhythm and intonation. Study methods and results.