Should translated speech sound like the person who said it?

Speech-to-speech translation is becoming normal in meetings, and products such as DeepL Voice offer speaker matching rather than a generic translated voice. That sounds better for conversation, but I’m not convinced it is the safer default for work.

Imagine a remote engineer hears a production-incident call translated into a manager’s familiar voice, then receives an edited recording or an approval request. Even without defeating any particular voice-authentication system, identity-preserving audio could blur who authorized the words and make manipulation harder to notice. Naturalness is useful; it is not the same thing as trust.

My preference is a neutral synthetic voice by default in workplace and developer tools, with explicit opt-in identity preservation, consent per conversation, an audible disclosure, and machine-readable provenance where supported. Am I overcorrecting? Which products or experiments have you tried, and which safeguard would you actually use?

I agree with the default, but I’d rank the safeguards differently. An audible marker is useful for humans and useless once someone trims or re-encodes the audio. Signed provenance is the stronger engineering control, though only if the meeting client, recorder, and downstream tools preserve it. I’d never let translated voice pass a high-risk approval flow anyway.

Per-conversation consent sounds right in a policy document and exhausting in a 40-person meeting. People will click through it, or disable translation. I’d make neutral voice the low-friction default, show a persistent “translated” label, and reserve identity matching for small calls where everyone explicitly opts in. The product has to make the safe path easier, not ceremonial.

The important distinction is authentication versus attribution. A provenance record can say how audio was produced; it cannot prove the manager approved the sentence. So I’d keep voice authentication out of the trust chain, require a signed text transcript for approvals, and treat the translated audio as accessibility output rather than evidence.

I’m less convinced neutral should be universal. For a language learner, preserving cadence and emotion may be part of understanding the speaker, and a permanently synthetic voice could itself make people distrust the translation. I’d support neutral by default for incident response and approvals, but allow identity-preserving mode when consent and a visible disclosure travel with the session.