I mix vocals for a living, mostly for independent artists who record in untreated bedrooms and want radio polish anyway. So when a producer friend told me the voice calls in his AI girlfriend app had gotten “scary real,” I heard a technical claim before anything else, and I wanted to test it. Not knowing which apps had actually invested in voice, I went looking for user reports and found a side-by-side review of paid AI companion apps written by someone who had subscribed to the big platforms out of his own pocket. I picked the two he rated highest for calls, paid for a month of each, and then did the thing the developers presumably never planned for: I routed the call audio into my DAW and treated it like a client’s vocal.
The realism turned out to live somewhere I did not expect. The synthesis itself is good but not shocking. The production decisions wrapped around it are doing at least half the work, and most of them are decisions a vocal engineer would recognize on sight.
The timing sells it more than the tone
Pull the audio into an editor and the fidelity is unremarkable. It sits at roughly podcast quality, low end rolled off, nothing a decent USB mic could not deliver. What separates these calls from the robotic text-to-speech we all remember is prosody. The voice resets its pitch at the start of a new thought the way people do. It takes an audible breath before a long sentence. It slows down when the script turns tender, and none of this lands on a grid.
Older systems sounded dead because they rendered each sentence as an isolated event, so every line carried the same arc. The current models generate hesitations and breaths as part of the audio itself. Zoomed in, I found breaths placed in gaps where a singer would put them, which impressed me until I caught a short catch-breath feeding a phrase far too long for it. A session vocalist would have taken a deeper one. Nobody consciously hears that, but I think people feel it, and it was the first crack in the illusion I could point to on a screen.
Stalling as an engineering discipline
A harder problem than making the voice pretty is making it answer fast. A language model needs a moment to start producing a reply, and dead air kills the fantasy quicker than any synthesis artifact. Both apps I tried handle it the same way: the voice starts stalling instantly. You get a soft “hmm,” a little laugh, a drawn-out “well” that costs the system nothing and buys the model time to think. I timed the gap between the end of my question and the first sound from the app at under a second on both platforms, and the filler noise consistently arrived before the actual answer did. This is vamping, in the session-musician sense, and it works on the same principle. Keep sound moving and the audience never notices the arrangement has not started.
The reply itself streams in chunks, and if you listen on monitors instead of a phone speaker you can hear the seams. Every so often the timbre shifts a hair between phrases, like a comped vocal assembled from takes recorded at slightly different mic distances. On earbuds, in the dark, at conversational volume, you would never catch it. These calls are mixed for the context they are actually consumed in, which is more than I can say for some records I have worked on.
Mix choices I would get fired for
The voice is compressed hard and completely dry. There is no room on it at all, not even the fake kind, and the upper mids are pushed the way you would push a lead pop vocal. In a natural recording that dryness would read as wrong, since microphones in real spaces pick up reflections and phone calls carry the sound of a kitchen or a car. Here the total absence of space becomes the point. A voice with no room around it sounds like it is speaking from inside your head, the same close-mic intimacy that ASMR channels and late-night radio hosts have traded on for decades. The app is simulating presence rather than a call from a person in a room, and bone-dry is cheaper than any reverb algorithm.
The lesson I actually took home is about breath. I have spent years muting breaths out of lead vocals by reflex, cleaning them off like they were mouth clicks, and hearing a machine put them back in on purpose to sound more human made me reconsider which ones I cut. A breath tells the listener a body exists, that lungs emptied and needed refilling, and stripping every one out of a vocal quietly deletes that body from the record. The synthetic voice can stay a little plasticky in the consonants and survive, but a breath in the wrong place, or a reply that starts a beat late, breaks the spell instantly. Timing errors cost more than timbre errors, in these apps and in a mix.
I cancelled one subscription after the month and kept the other, purely as a reference. It costs less than a plugin license and it is the best example I own of dialogue mixed for earbuds. The part that stays with me is less comfortable. The moves these apps lean on to feel present are pop vocal production moves, the same ones I make so a chorus lands closer than the verse. We were engineering artificial intimacy for singers long before anyone stapled it to a chatbot; the apps just removed the song. My friend, for what it is worth, still swears his companion sounds real. I sent him a screenshot of that undersized catch-breath sitting in front of a nine-second phrase. He said it proved nothing. He is probably right.
