Voice changer
DSP vs AI voice changing
Every real-time voice changer belongs to one of two families. DSP effects reshape the voice you already have, instantly. AI conversion replaces it with a modeled voice, for a real latency cost. Knowing which is which tells you exactly when to reach for each.
If you have ever wondered why one voice effect feels like part of your own mouth while another sounds like a different person entirely, you have already noticed the split this page is about. The two families are built on different ideas, they cost different amounts of delay, and they are good at different jobs. If you want the broader picture of what a voice changer even is, start with what is a voice changer and come back.
Reshape versus replace
A DSP voice effect reshapes your voice. Pitch shifts, formant moves, robot and radio character, all of it is arithmetic applied to the signal coming out of your mic. Your timing, your phrasing, your laugh, they all survive the trip. The result is recognizably you, processed. And because it is arithmetic, it happens essentially instantly: an instant DSP voice adds almost no delay on top of your normal mic path.
AI voice conversion replaces your voice. A trained model listens to what you say and produces the same words in a different voice, a wholly different character with its own timbre. You still drive the performance, but the sound that comes out is not a filtered version of you; it is a new voice wearing your delivery. That power has a price: the conversion adds about 350 ms of delay, plus a few seconds of one-time warm-up the first time you select an AI voice.
So the honest decision rule is short. If you are having a conversation, a chatty stream, a call, comms in a game, you want the instant family: DSP effects keep the back-and-forth feeling natural. If you are doing a character bit, an alternate persona, a voice that needs to be genuinely somebody else, the AI family is the only one that can do it, and a third of a second is a fair trade for a whole new identity.
How the two families live together in Persona
In Persona, patchd's voice changer, the two families are not two products or two modes. They sit in the same voice list and behave the same way from where you sit. Every voice, DSP or AI, gets its own hotkey bind, and that includes a bind for your real, unprocessed voice, so dropping character mid-sentence is one keypress, not a menu dive.
Switching is designed to be invisible in the audio: when you jump between voices, Persona crossfades the change so there is no click or gap, just your output smoothly becoming the next voice. The one exception to instant switching is the first time you pick an AI voice in a session: the model loads once, which takes a few seconds, and after that warm-up switching to it is near-instant like everything else.
Persona also does not care where your voice comes from or where it goes. It runs on any input, inside the patchd mixer as the last stage of a channel, or standalone if you do not need the rest of the console. On the Free tier you get 6 rotating taster voices to play with; the launch target is about 25 voices, with the full lineup, including the AI voices, on Studio. If you are building a full stream chain around it, the streaming voice changer guide walks the whole setup.
Why the latency actually differs
The delay gap between the families is not an implementation detail someone could optimize away next quarter. It falls straight out of what each family is doing. A DSP effect is a fixed formula evaluated on every sample as it arrives:
y[n] = f(x[n])Every output sample is a formula applied to input samples the effect has already seen. At 48 kHz that formula runs 48,000 times a second and never falls behind. It is math with no opinions and no waiting room.
Because the formula needs nothing beyond the samples already in hand, a DSP voice adds essentially zero delay of its own. What you hear is just the normal engine path: in patchd that is about 10.7 ms through the bus at a 512 buffer, and smaller ASIO buffers take it lower. The voice effect rides along for free.
AI conversion cannot work per sample, because a single sample carries no meaning. The model needs a frame of audio, a chunk long enough to contain something worth converting, and then it needs time to run on that chunk. A trained model runs on every frame of audio, and that takes real compute. So the floor on its latency is structural:
latency >= frame_time + compute_timeThe conversion can never be faster than the frame it waits for plus the compute it spends. For Persona's AI voices those two terms add up to about 350 ms. You cannot convert audio you have not heard yet; physics has never once granted an exception.
Keep the two numbers separate, because they answer different questions. The 10.7 ms figure is the mixer's bus path, and it is what a DSP voice effectively costs you. The roughly 350 ms figure is what an AI voice adds on top of whatever path it sits in. They do not average, and no amount of buffer tuning moves the AI number, because buffers are not where it lives. The few seconds of first-use warm-up is different again: that is the model loading once, not a per-word cost, which is why switching back to a warmed-up AI voice is near-instant. For the deeper arithmetic on all of this, see voice changer latency, explained.
Two families, one seat
The comparison ends somewhere pleasantly boring: you do not have to choose. In patchd, Persona sits at the end of the channel after your cleanup chain, and the DSP voices and AI voices share the same list, the same per-voice hotkeys, and the same clickless switch. Chat on an instant voice, hit one key for the character bit, hit another to be yourself again. patchd is in development now. Join the waitlist and we will tell you the moment it is ready to install.