How it works
You call. You record. When the recording ends, you hear a prompt with four options. Press a number. Hang up. That's it.
The transformation happens after you hang up, not during the call. There's no real-time processing constraint, no buffering penalty, no reason to cut corners. You record in your own voice. The system applies whatever you chose, cleanly, in the background. What goes live is intentional — not distorted by latency.
If you don't press anything, your message is published as recorded. There's no second chance to change it after you hang up.
The four options
No processing. Your voice, exactly as recorded. If the content of what you're saying doesn't put you at risk, this is the right choice. The archive exists to be heard — and recordings in real voices become part of the greeting pool that future callers hear when they call in.
Your recording passed through a 1960s vocoder. Ring-modulated, downsampled to 8kHz, quantized to 8-bit. There is no human voice left — only the shape of what you said, expressed as machine noise. No gender, no register, no vocal fingerprint. Maximum separation from your original voice at the cost of everything that makes it human.
Your message is transcribed, then re-read by an early Macintosh speech synthesizer — the concatenative text-to-speech engine from the 1980s, with its flat cadence and narrow frequency response. Your words survive; your delivery doesn't. The output sounds like a document being read by a machine, which is what it is.
Your performance, in a different human voice. The nervousness, the pauses, the catches — all of it comes through. Only the voice identity changes. Built for situations where you need to sound like a different person — not just a different pitch — while keeping the raw delivery intact. If you're a man calling about something that could identify you by gender, press 4 then 1 for a female voice, or 4 then 2 for a male voice. The swap is deep enough that people who know your voice well will not recognize it.
What masking can and can't hide
Every option here changes how your voice sounds. None of them change what you say.
Your vocabulary is identifying. Your sentence structures are identifying. The way you hedge, the things you choose to emphasize, the specific details only you would know — all of it survives any voice transformation. If you say you're the only nurse on the night shift at a specific hospital in a specific city who voted for the strike, your hospital administrator will know who you are. No voice mask changes that.
Protect yourself with what you choose to say, not just with how your voice sounds. The transformation handles one vector. It doesn't handle all of them. Think before you call — not after.
Why Voice Swap retains your delivery
The other two masking options — Classic Computer and Mac Era — destroy your vocal performance entirely. That's by design. If you want maximum separation from your voice, those options deliver it.
Voice Swap takes a different approach. It uses AI speech-to-speech conversion, which means your original recording is the input, not a transcript. The breathing, the hesitations, the moments where your voice drops or cracks — the system reads all of that and expresses it through the target voice. The result sounds like a real person speaking with real emotion, because it is. The identity has changed. The testimony hasn't.
This matters because testimony is credible partly because of how it's delivered. A flat, synthesized re-reading of a whistleblower's message carries less weight than a recording where you can hear what it cost them to make it.
Why there's no "sound like a specific person" option
Some voice tools let you upload a target voice — a celebrity, a public figure, someone specific — and apply it to your recording. We don't offer this and won't.
The harm is obvious. A service that lets you record yourself sounding exactly like your CEO saying they embezzled $40 million is not something we want to build. Our options are anonymous personas, not identities. Nobody can be impersonated by a voice called "Rachel." That distinction matters and it's not going away.
Questions about the technical implementation? Write to hello@psa.xyz.