Benchmark

The Benchmark section explains how OpenTalking records end-to-end experience metrics and how it references inference baselines from external model services. OpenTalking is the orchestration layer, so the documentation separates two data classes:

Type	Directly owned by OpenTalking	Examples
End-to-end experience metrics	Yes	First frame latency, TTS first packet, events, WebRTC playback, A/V sync.
Model inference baselines	No, owned by the selected backend	OmniRT FlashTalk, Wav2Lip, QuickTalk local adapter throughput.

Reading Order

1. Metrics — shared field names and measurement boundaries.
2. Runbook — how to collect QuickTalk and end-to-end data.
3. Results and Baselines — current reference numbers and the result template.

Recording Rules

• Every result must include hardware, model, backend, resolution, input audio duration, and startup state.
• Cold start, hot state, and steady chunk measurements must be kept separate.
• OmniRT or other external service results must be labeled as backend data.
• A mock run proves orchestration health, not real talking-head performance.