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

Tileward research report R-007. PersonaPlex is NVIDIA's speech model: it listens and talks at the same time, so it has 80 milliseconds to produce each frame of audio. As published it is 16.7 GB and spends about 40 ms on each frame on one laptop. Tileward's TWX4 build is 6.3 GB, 62% smaller, spends 20 ms, and given the original's own conversations chooses the same next word 96% of the time.

Page: https://tileward.com/research/a-smaller-voice-model/

## How it was measured

Every build ran on the same laptop, in the same runtime, on the same conversations, and each is compared with [PersonaPlex as NVIDIA publishes it](https://huggingface.co/nvidia/personaplex-7b-v1).

- **Builds:** the published model; a standard 8-bit conversion (the one the runtime ships with); the same converter taken down to the size of the TWX4 build; and Tileward TWX4.
- **Machine:** one MacBook Pro with an M4 Max. Every build ran in Apple's MLX, each from its own weights.
- **Size:** files on disk.
- **Compute:** the model alone on one conversation, with nothing else running; the audio codec runs beside it and is not part of the figure. The published model and the TWX4 build ran three sessions each; each standard conversion ran one.
- **Same words:** four conversations recorded from the published model (106 seconds, with 375 moments where a word is due), played back through each build. A point either way is within chance.
- **Speaking freely:** four spoken questions, five runs each, transcribed by Whisper medium.

## Results

| Build | Weights | Compute for each 80 ms frame | Same next word as the original |
|---|---|---|---|
| PersonaPlex as published | 16.7 GB | 37 to 42 ms | 99.2% (replayed in this runtime; the ceiling) |
| Standard 8-bit conversion | 8.2 GB | 23 ms | 98.1% |
| Standard conversion at our size | 6.2 GB | 19 ms | 92.8% |
| Tileward TWX4 | 6.3 GB | 20 ms | 96.0% |

- Speed follows size: the standard conversion at our size is as fast as TWX4. What separates them is how closely each follows the original.
- The standard 8-bit conversion comes closest to the original, at 98.1%, for 1.9 GB more and 3 ms more a frame than the TWX4 build.
- Speaking freely, a speech recogniser heard 4.4% of the TWX4 build's words differently from the words it meant to say, against 7.0% for the published model (95% ranges 3.0 to 5.8% and 4.3 to 10.2%). That gap is within the margin of the test, so the builds are level. No build went silent or repeated itself.

## Deployment meaning

A build that uses a quarter of each frame leaves room for the application, for more than one call, or for a smaller machine. 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.

## Related

- [Voice model eval card](https://tileward.com/evals/#compressed-voice-model): the result and limits in one screen
- [PersonaPlex on Hugging Face](https://huggingface.co/nvidia/personaplex-7b-v1): NVIDIA's published model
- [Tileward Models](https://tileward.com/models/): what we serve and what it costs
- [Talk to an engineer](https://calendar.app.google/hg2DzFTxKJgwFWm19): bring the voice model you run and the machine you run it on
- [Research](https://tileward.com/research/): every report