Runbook¶
QuickTalk Local Adapter¶
The repository includes apps/cli/quicktalk_bench.py for measuring QuickTalk local adapter
load time, first frame, render throughput, and mux time.
Terminal
source .venv/bin/activate
python apps/cli/quicktalk_bench.py \
--asset-root /path/to/quicktalk/assets \
--template-video /path/to/template.mp4 \
--audio /path/to/input.wav \
--output outputs/benchmarks/quicktalk-output.mp4 \
--device cuda:0
The JSON output includes:
init_secondsaudio_feature_secondsfirst_frame_secondsrender_secondsrender_fpsmux_seconds
OpenTalking End-to-End Path¶
End-to-end testing should pin the model, TTS provider, and input audio before collecting browser, API, and Worker logs.
Record:
- OpenTalking commit, config file, and non-secret
.envsettings. - Hardware and driver versions.
- Selected
avatar_id,model, andbackend. - Input audio duration, sample rate, and prompt text.
- First token, TTS first packet, avatar first frame, browser first frame, and A/V sync.
External Model Services¶
OmniRT, FlashHead direct WebSocket, or other model services should be benchmarked with their own tools. OpenTalking documentation should reference those results and record only the OpenTalking-side call, queue, and playback behavior.