Kinder Voice · Safety principles

Repair delivery without inventing a different person.

A live speech system can create harm if it changes meaning, conceals urgency, or speaks with authority the user never intended. These are product boundaries, not optional polish.

Preserve the claim

The system must not convert disagreement into agreement, certainty into uncertainty, a refusal into consent, or a warning into reassurance. When preservation is doubtful, it should return the original signal or stop.

Do not impersonate

Kinder Voice must not fabricate emotion, humor, apology, affection, authority, or intent. The output should remain recognizably attributable to the speaker’s actual message.

Do not soften emergency meaning

Threats, calls for help, medical urgency, safety warnings, and crisis language require special handling. The system must never make an emergency sound routine merely to reduce harshness.

Bypass must be immediate

The user needs a clear way to hear or send the original signal without delay. A transformation layer that cannot be instantly disabled is not acceptable for live communication.

Fail openly

Model crashes, high latency, uncertain interpretation, device changes, and network loss should produce visible failure states. Silent degradation is more dangerous than an honest stop.

Release boundary: The live product should remain experimental until these behaviors can be tested under stress, not merely demonstrated under ideal conditions.