Translation should not automatically sound like the person

I’m a remote-work engineer and regularly sit in multilingual meetings. I want speech-to-speech translation to be useful without quietly turning every translation into an impersonation. Translation and voice imitation should be separate controls: a neutral synthetic voice should be the default, while a cloned voice is opt-in for each session.

Vocal likeness can change how authority, emotion, urgency, and trust are perceived. It also creates obvious security and audit questions: who approved the clone, which words were translated, and how do we distinguish the original speaker from generated audio? I’d want an always-visible “translated and synthesized” indicator, plus privacy-minimized metadata linking the output to the original speaker or recording without exposing more identity than necessary.

Is preserving someone’s voice an accessibility feature worth the identity risks? Disagree, or share a meeting example where voice preservation genuinely helped.

The per-session part matters more than a permanent preference, in my view. Consent can be valid for one meeting and wrong for the next, especially when an external guest joins. Metadata helps audits, but it is not a security boundary by itself; labels and access controls still need to survive export and replay.

I agree with the default, but products will be tempted to make the recognizable voice the “delight” feature because users perceive it as more personal. The interface should make neutral voice the path of least resistance, not bury it under an advanced-settings menu.

I’d be careful calling the neutral voice safer in every sense. It may reduce identity confusion, but the research basis here is mostly about how voices shape judgments generally, not translated meetings specifically. That supports a cautious design proposal, not a measured claim about outcomes.

The visible indicator is a good idea, provided it is readable in the actual meeting flow. Tiny badges become decoration. I’d also let people hear the original and synthesized channels separately, because agency includes being able to check what the system did.

Accessibility is the strongest counterargument. Someone who has lost speech, or relies on a familiar communication identity, may reasonably value voice preservation. That argues for explicit, revocable consent and a safe fallback—not for making imitation the invisible default.

From an implementation standpoint, “who approved the clone?” needs an answer that is machine-verifiable, not a checkbox in a meeting app. Otherwise the audit trail records a claim made by the same system producing the audio. Separate signed events for consent, translation, and synthesis would be the cleaner model.

I’m not fully convinced a neutral voice avoids authority effects. People will still infer competence or urgency from timing, wording, and the selected synthetic voice. Maybe the honest default is disclosure plus user choice, rather than implying one voice is socially neutral.

My preference is simpler: translate the words, show the label, and stop trying to make the computer sound like my colleague. If voice preservation becomes necessary for a particular person, let that be a deliberate exception with an expiry date.