[scott-lowe-vapi]

Linear: VAP15-76 Assess whether turn latency definitions belong in public docs. The matching API reference change is VapiAI/vapi#25247.

Description

• Define every per-turn latency field in one place. Understanding latency gets a "Per-turn latency on a call" section. A table says where transcriberLatency, endpointingLatency, modelLatency, voiceLatency and turnLatency start and end, and a short list covers what trips people up: the fields don't add up to turnLatency, voice latency includes the voice.chunkPlan wait, everything is measured on Vapi's servers, turns the assistant didn't answer have no entry, and the first message is recorded only when the assistant waits for the caller, 0 can mean "not measured" (for example, a static tool message), and per-call averages include those zeros while simulation limits skip them.
• Fix the stage order. The page listed endpointing before transcription. Lasensor measures transcription first (end of speech → words known), then endpointing (words known → turn sent to the model).
• Say the model-intelligence total isn't turnLatency. It's a sum of component medians across calls, and it leaves out endpointing.
• Link the new section from the pages that show latency. Simulation latency limits now say each metric is the same field as on a live call, name the field, and note that model and voice sit inside turn and that the synthetic tester can read faster than real callers. The Traffic tab says why its stages don't add up to the total. Performance metrics, call logs and the glossary link to it too.
• Definitions come from Lasensor in VapiAI/vapi (libs/core/src/util/lasensorUtil/Lasensor.ts) and the simulation evaluator (libs/core/src/functions/simulationFunctions/latencyExpectationsEvaluate.ts).

Testing Steps

• Pinned Fern 5.151.1 check: 0 errors and the same 15 existing warnings.
• git diff --check passes.
• In the preview, open Understanding latency and check the new table, then follow the link from Simulations → Advanced → Metrics and aggregations and from the Traffic tab page to the #per-turn-latency-on-a-call anchor.

🤖 Generated with Claude Code

[scott-lowe-vapi]

• TEST-273 docs: document transcriber and endpointing simulation latency limits and the unmeasured flag #1382
• VAP15-76 docs: define each per-turn latency field and link it from sims, Traffic and logs #1381 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[github-actions]

🌿 Preview your docs: https://vapi-preview-01a18faf-f77a-732b-99cf-2fcd054ac1f5.docs.buildwithfern.com

[github-actions]

🌿 Preview your docs: https://vapi-preview-01a18fb5-830f-713f-9a8a-f33ac0d958a2.docs.buildwithfern.com

[github-actions]

🌿 Preview your docs: https://vapi-preview-01a18fd2-acf1-7299-8906-651bf86f3ce8.docs.buildwithfern.com

[github-actions]

🌿 Preview your docs: https://vapi-preview-01a18fda-7eba-772b-9b72-dddf0a53a2ea.docs.buildwithfern.com

[stephenvapiai]

Edit pass on the changed passages. Small wording and style suggestions only; technical content not reviewed.

understanding-latency.mdx line 4 is outside the diff, so a note here: the page now covers the per-turn fields too, so the description (which feeds llms.txt) should mention them. Suggested:

description: Understand voice-to-voice latency, the per-turn latency fields each call records, Vapi performance metrics, and how model capability, endpointing, and transport affect it.

[scott-lowe-vapi]

Front-matter description updated in f95a238. I kept 'affect response time' at the end instead of 'affect it', which read oddly after a list: 'Understand voice-to-voice latency, the per-turn latency fields each call records, Vapi performance metrics, and how model capability, endpointing, and transport affect response time.'

[github-actions]

🌿 Preview your docs: https://vapi-preview-01a1af67-9899-75f7-b73b-96fd21e640be.docs.buildwithfern.com