We shrank PersonaPlex by 62% to run at 20ms a frame.

A model that talks while it listens cannot fall behind. Every 80 milliseconds it owes the conversation another frame of audio. For a voice model, size and speed are the product, not a cost line. We took PersonaPlex as NVIDIA publishes it, made three smaller builds, and ran all four on the same laptop.

The TWX4 build is 62% smaller than the published model. The published model spends about half of each 80 ms frame computing. The TWX4 build spends a quarter.

Speed follows size: the standard conversion at our size is as fast as ours. What separates those two is how closely each follows the original.

Smaller and faster only counts if the model still says what the original would. We recorded the original holding four conversations. Then we played them back through each build and, every time a word was due, asked which word the build would say next.

The first row is the ceiling. The standard 8-bit conversion comes closest to it, at 98.1%, for 1.9 GB more and 3 ms more a frame than the TWX4 build.

Then we let each build hold 20 conversations of its own, and had a speech recogniser write down what it heard. It heard 4.4% of the TWX4 build’s words differently from the words the build meant to say, against 7.0% for the published model. That gap is within the margin of this test, so the builds are level. No build went silent or repeated itself.

On hardware you own, a voice model shares the machine with everything around it. A build that uses a quarter of each frame leaves room for the application, for more than one call, or for a smaller machine.

We measured the second of those. Four conversations at once took 34 to 39 ms a frame on the TWX4 build and 49 to 61 ms on the published model. Both are inside the 80.

Both tests compare each build with the model as NVIDIA publishes it.

The machine. One MacBook Pro with an M4 Max. Every build ran in the same runtime, Apple’s MLX, from its own weights. The published model ran from the published weights, and each smaller build from its compressed ones.

Size. Files on disk.

Compute. The model alone on one conversation, with nothing else running on the machine. The audio codec runs beside it and is not part of this figure. The published model and the TWX4 build ran three sessions each: 36.9, 39.7 and 42.2 ms, and 19.9, 20.0 and 20.1 ms. Each standard conversion ran one.

Same words. Four conversations recorded from the published model: 106 seconds, with 375 moments where a word is due. A point either way is within chance.

Speaking freely. Four spoken questions, five runs each, about half a minute a conversation. Whisper medium wrote down what each build said. The 95% ranges are 3.0 to 5.8% for the TWX4 build and 4.3 to 10.2% for the published model.