Voice changer
Voice changer latency, explained
Latency is the delay between you speaking and the changed voice coming out. Some voices add almost none of it, some add a third of a second, and the difference is not a settings problem. This page gives you the honest numbers and the arithmetic behind them.
If your voice changer feels laggy, the first thing to know is that there are two very different kinds of voice under the hood, and they pay very different latency bills. This page separates them, puts real numbers on each, and shows which parts of the delay you can shrink and which parts are physics. If you want the broader comparison of how the two kinds of voice actually work, read DSP vs AI voice changing first and come back.
The two honest numbers
Latency here means one thing: the time from sound leaving your mouth to the changed voice arriving wherever it is going, your game chat, your stream, your recording. Every voice changer adds some. The honest question is how much, and the answer splits cleanly in two.
- Instant DSP voices: almost nothing added. These voices are built from standard signal processing, pitch shifts, filters, modulation. The math runs as fast as the audio arrives, so the voice adds essentially no delay of its own on top of the audio path you already had.
- AI voices: about 350 ms added. An AI voice runs a trained model on every frame of your audio, and that takes real compute. In Persona that costs about 350 ms, plus a few seconds of one-time warm-up the first time you load a voice. After that first load, switching between loaded voices is near instant, with a clickless fade so the swap never pops.
Whether 350 ms matters depends entirely on what you are doing. In a tight two-way conversation, a third of a second is the gap between a comeback and an awkward pause; your teammates will feel it. For streaming and recorded content it barely registers, because your voice and your video leave together: the audience hears the delayed voice against equally delayed gameplay, perfectly in sync. Same number, completely different verdict.
The engine line is identical on both receipts. The difference is the voice you ordered: arithmetic is free, a new identity takes about a third of a second.
The full path, and what you can actually shrink
The voice is only one line on the bill. Your audio also has to get through the engine: in from the mic, through the effects chain, out to every destination. In patchd that trip costs about 10.7 ms through the bus at a 512 buffer, and smaller ASIO buffers take it lower. So the total delay you experience is always two numbers stacked, the engine trip plus the voice's own cost.
Keep those two numbers separate. It is tempting to describe a voice changer with one blended figure, but an average of 11 ms and 361 ms describes neither voice: it makes the instant voice sound sluggish and the AI voice sound faster than it can be. Two kinds of voice, two bills, no averaging.
Here is what you can actually control:
- The buffer, in your interface's own control panel. patchd has no buffer selector; the engine locks onto the fastest path your device offers automatically and reports the result live in the master clock readout. Buffer size is set in your audio interface's own control panel, and dropping from 512 to 256 or 128 samples cuts the engine leg of the trip roughly in half each step. See what buffer size is for the trade-off.
- A lean chain. Standard voice DSP, the EQ, gate, and compressor kind, adds effectively no buffering of its own, so a sensible cleanup chain is close to free. Just do not stack nodes you are not using.
- The voice itself. This is the big lever. If you are in a fast conversation, pick an instant DSP voice and the whole path stays around the engine figure. Save the AI voices for streams, videos, and anywhere your audio and video travel together.
What you cannot shrink is the model. The roughly 350 ms an AI voice costs is compute, not a misconfiguration, and no slider makes it disappear.
The arithmetic
Everything above is one addition and two definitions. The total is the engine trip plus whatever the active voice costs:
total_ms = engine_ms + voice_msWith an instant voice the second term is approximately zero and the total is basically the engine. With an AI voice it is about 350 ms and dominates everything else on the page. Addition remains the most underused tool in latency marketing.
The engine term is set by buffering. Audio moves in blocks, and each block has to fill before it can move, so the buffer size and sample rate fix the cost:
engine_ms = buffer_samples / sample_rate * 1000512 samples at 48 kHz is 10.7 ms per buffer. The buffer charges the same toll no matter what effects you run through it.
The AI term has a hard floor. A model works on frames of audio, so it must first wait for a frame to fill, then spend real compute on it, and both costs happen on every frame forever:
ai_voice_ms >= frame_ms + compute_msA model cannot transform audio it has not received yet, and it cannot compute in zero time. Any tool promising a dramatic AI voice at zero delay is glossing over physics.
Now the worked example. At a 512 buffer the engine trip is about 10.7 ms. Run an instant DSP voice and the total stays around 11 ms, which most people read as immediate. Run an AI voice and the total is about 10.7 + 350, call it 361 ms. Notice what the arithmetic did not do: it did not blend the numbers. Shrinking the buffer to 128 samples trims the engine leg to about 2.7 ms, which is a real win for the instant voice and a rounding error next to the AI voice's 350. The warm-up is a separate item entirely: a few seconds, once per voice per session, and zero on every frame after that. It belongs on the receipt as a one-time charge, not in the per-word total.
Where this lands in practice
In patchd, Persona sits at the end of the effects chain, after your cleanup nodes, added like everything else through the + Add Node menu. The mixer shows the engine numbers live in the master clock readout, the hardware and bus latency in milliseconds, so the part of the bill you can tune is always visible. The instant DSP voices ride that same low-latency path; the AI voices pay their model time on top, honestly and separately. patchd is in development now. Join the waitlist and we will tell you the moment it is ready to install.