Observability

Observability

Telemetry is built into the runtime for both transcription and synthesis. Each performance sample includes:

• clientId
• route
• modelId
• voiceId for synthesis routes when a voice is selected
• outcome
• nested metrics

Metrics

Transcription metrics:

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

Synthesis metrics:

• modelCheckMs
• modelLoadMs
• voiceResolveMs
• synthesisMs
• totalMs
• audioDurationMs
• outputBytes
• characterCount

Derived values: realtimeFactor, warm vs cold from modelLoadMs, audio-to-text speed from audioDurationMs / inferenceMs, and text-to-audio speed from audioDurationMs / synthesisMs.

Storage

The runtime appends JSON lines to:

~/.vox/performance.jsonl

The CLI dashboard reads from this file. You can also export it to another metrics backend.

Current emitted performance routes:

• transcribe.file
• transcribe.live
• synthesize.generate
• synthesize.startSession

Cancelled synthesis sessions are recorded under synthesize.startSession with outcome: "cancelled".

Operator Commands

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

Reading the numbers

inferenceMs, synthesisMs, and totalMs measure different things.

• For ASR, inferenceMs is how fast the hot model ran.
• For TTS, synthesisMs is how long audio generation took once the request was inside the model.
• totalMs is what the user experienced end-to-end.

Example sample

{
  "timestamp": "2026-04-25T17:54:26Z",
  "clientId": "menu-bar",
  "route": "synthesize.generate",
  "modelId": "avspeech:system",
  "voiceId": "com.apple.voice.compact.en-US.Samantha",
  "outcome": "ok",
  "textLength": 14,
  "metrics": {
    "audioDurationMs": 1240,
    "synthesisMs": 182,
    "inferenceMs": 182,
    "totalMs": 196
  }
}

Dashboard tips

• Only compare clients when the prompt or audio shape is similar.
• Use inferenceMs for loaded-model ASR speed.
• Use synthesisMs for loaded-model TTS speed.
• Use totalMs for end-user latency.
• Large modelLoadMs spikes are warm-up events, not inference regressions.