# We compressed VoiceChat by 74% and made it run in real time on one card

Tileward research report R-008. VoiceChat is NVIDIA's speech model: it listens and talks at the same time, and can call tools while it does. Tileward compressed it from a 44.4 GB file to 11.5 GB, and it still picks the right tool as often as the original. Separately, Tileward's serving changes make it run in real time on one card: each 80 ms frame of sound takes 72 ms, against 114 ms for NVIDIA's open code.

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

## How it was measured

Every figure sets the TWX4 build beside [VoiceChat as NVIDIA publishes it](https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B), on the same inputs.

- **Size:** files on disk. NVIDIA's file stores every number in 32 bits; nine tenths of them fit in 16 bits exactly, which takes the file to 24.2 GB without changing a value. The TWX4 build is 74% smaller than the published file and 53% smaller than that 16-bit copy.
- **Tool calls:** 100 spoken requests in six synthetic voices on a clean line: forty against five operations tools, forty against five everyday tools, and twenty that call for no tool. Both builds ran NVIDIA's own system prompt and tool template; the score is the first call each one made.
- **Same words:** the original's 100 recorded conversations, played back through each build and compared at 1,661 words and 1,848 tool-call tokens.
- **Speed:** one NVIDIA 96 GB card running NVIDIA's open pipeline, one conversation at a time. Times are the average frame.

## Results

| Build | Right tool | Right tool and arguments | Quiet when no tool applies |
|---|---|---|---|
| VoiceChat as published | 52 of 80 | 46 of 80 | 20 of 20 |
| Tileward TWX4 | 55 of 80 | 50 of 80 | 20 of 20 |

- A gap of three or four calls in 80 is within chance, so the two are level. They made the same first call on 89 of the 100 requests.
- Given the original's conversations, TWX4 chooses the same next word 95.4% of the time and the same next tool-call token 99.5% of the time. The ceiling, the published model run in 16 bits against its own recording, is 99.3% and 100.0%.
- Speed, published model on a five-tool prompt: NVIDIA's open code as it ships takes 114 ms per frame (24.7 GiB on the card) and falls further behind as the call goes on; with Tileward's serving changes it takes 72 ms (24.4 GiB).
- With Tileward's serving on a plain conversation, both builds take 71 ms per frame and 24.2 GiB on the card, because the server runs every build from unpacked weights. The saving is in the file you download, store and move: 11.5 GB.
- For both builds, a request's slowest frame took 86 to 89 ms, and 3 of 19 requests had one near 350 ms. A host needs an audio buffer for those.
- Across 100 answers offline and 20 through the streaming path, an answer repeated itself twice for the original and once for the TWX4 build.

## Licence

VoiceChat is NVIDIA's model, released under the [OpenMDW License Agreement, version 1.1](https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B/blob/main/LICENSE) (OpenMDW-1.1), Copyright © 2026 NVIDIA Corporation & Affiliates. The TWX4 build is a modified version of it under the same agreement, and NVIDIA's notice and a copy of the agreement stay with every copy of it.

## Related

- [Voice model eval card](https://tileward.com/evals/#compressed-voice-agent): the result and limits in one screen
- [VoiceChat on Hugging Face](https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B): 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 card you run it on
- [Research](https://tileward.com/research/): every report