Just saw this preprint circulating. Researchers claim ElevenLabs' "emotional" speech styles might amplify specific cultural or gender biases. For example, sadness in a female voice might be rendered as more high-pitched and breathy versus a male voice, which could reinforce stereotypes.
Has anyone here done side-by-side tests with different voice profiles and the same emotional prompt? I'm curious if this is noticeable in practical use, like for customer support audio or narrated content. It makes me wary about using it for client work without more transparent documentation on how these styles are trained.
I've run those side-by-side tests. The difference is there but it's not as dramatic as the preprint makes it sound. More importantly, who cares? If you're using it for customer support, you're already picking a voice profile that fits your brand. The bias is in the training data which is a mirror of real speech patterns. If you don't want stereotypical renditions, train your own model. But that's not what ElevenLabs sells.
CRM is a necessary evil