Observability

Observability

Vox treats transcription telemetry as a first-class runtime feature.

Dimensions

Each performance sample is tagged with:

• clientId
• route
• modelId

Metrics

Current metrics include:

• fileCheckMs
• modelCheckMs
• modelLoadMs
• audioLoadMs
• audioPrepareMs
• inferenceMs
• totalMs
• audioDurationMs

Additional useful derived values:

• realtimeFactor
• warm vs cold path behavior from modelLoadMs
• effective audio-to-text speed from audioDurationMs / inferenceMs

Storage

The runtime appends JSON lines to:

~/.vox/performance.jsonl

That local store powers the CLI dashboard today and can later be exported to another metrics backend if needed.

Operator Commands

vox transcribe file --metrics /tmp/sample.wav
vox transcribe bench /tmp/sample.wav 5
vox perf dashboard
vox perf dashboard --client vox-cli

Philosophy

Loaded-model inference speed and end-to-end latency are different things.

Vox records both:

• inferenceMs tells you how fast the hot model is.
• totalMs tells you what the user actually experienced.

Example sample

{
  "clientId": "raycast",
  "route": "transcribe.file",
  "modelId": "parakeet:v3",
  "audioDurationMs": 5110,
  "inferenceMs": 151,
  "totalMs": 165
}

How to read the dashboard

• Compare clients against each other only when the audio shape is similar
• Use inferenceMs to judge loaded-model speed
• Use totalMs to judge end-user experience
• Treat large modelLoadMs spikes as warm-up lifecycle events, not steady-state inference regressions