[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-doc:docs\u002Fconcepts\u002F5.latency":3},"---\ntitle: Understanding Latency\ndescription: There is no single latency number for a live speech system. What users experience as latency is the result of several different stages.\n---\nThere is no single latency number for a live speech system.\n\nWhat users experience as \"latency\" is the result of several different stages.\n\n## The latency pipeline\n\nA simplified real-time pipeline looks like:\n\n```text\nspeaker\n  ↓\naudio capture\n  ↓\nnetwork transport\n  ↓\nspeech recognition\n  ↓\ncontext processing\n  ↓\noptional translation\n  ↓\napplication rendering\n```\n\nEach stage contributes to the final experience.\n\n## First-result latency\n\nThe time between incoming speech and the first usable recognition result.\n\nLower first-result latency generally makes an interface feel more responsive.\n\nHowever, extremely early results may contain greater uncertainty.\n\n## Stabilization latency\n\nThe time required before evolving speech becomes sufficiently stable for a particular use.\n\nThis is different from first-result latency.\n\nFor example:\n\n```text\n300 ms → first interpretation appears\n900 ms → surrounding context resolves ambiguity\n1.2 s → segment becomes stable\n```\n\nThese numbers are illustrative only.\n\nThe important point is that **responsiveness and stability are different measurements**.\n\n## Translation latency\n\nLive translation introduces another dependency.\n\nTranslation quality improves when more linguistic context is available, while live experiences require output before the full conversation is known.\n\nThis creates another latency-quality tradeoff.\n\nLansonAI therefore treats live translation as a streaming context problem rather than simply translating a completed transcript.\n\n## Measure the user experience\n\nFor live applications, useful latency measurements should reflect what the user actually experiences.\n\nDepending on the application, this may include:\n\n- time to first readable text\n- time to stable text\n- time to translated text\n- correction frequency\n- visible reflow\n- end-to-end interaction latency\n\nOptimizing only one number can make another part of the experience worse.\n",1790059118948]