Fact · Voice and AI
20times
In the VoiceAgentBench measurements, the two-stage pipeline's 90th percentile time to first token was about 20 times that of the end-to-end speech model.
Checked against its source on October 17, 2026
- What it measures
- The ratio of two 90th percentile times to first token, both measured on the same H100 GPU in the VoiceAgentBench appendix.
- How it is calculated
- 800 ms / 40 ms = 20
- What it does not show
- The ratio compares two different systems, Whisper with Qwen3 8B and Qwen2.5-Omni 7B, so it shows the size of the design gap, not a like-for-like model comparison.
Calculated from
- A pipeline that runs Whisper large v3 first and then the Qwen3 8B language model had a 90th percentile time to first token of about 800 milliseconds on the same hardware.
- Qwen2.5-Omni 7B, an end-to-end speech language model, had a 90th percentile time to first token of about 40 milliseconds on one H100 GPU.