Time to first audio (TTFA)
116 ms
p50 first audio, server side warm
The delay between sending text to a TTS system and getting back the first byte you can play. You hear it as the silence before the voice starts.
Where the milliseconds land
The scale below is one second, roughly as long as a listener will wait. Judge a TTFA claim by where it sits on that line, and by how far its p95 drifts from its p50.
fig.
One second of listener patience, to scale
Fig
200 ms
human turn-taking rhythm
1,000 ms
listeners start interrupting
200 ms
human turn-taking rhythm
1,000 ms
listeners start interrupting
How far a claim’s slow end sits from its median tells you how honest it is. Ours: first audio byte in 146 ms over the open internet, 116 ms p50 first audio, server side warm.
TTFA vs model latency vs TTFB
Model latency counts only the forward pass. TTFB counts any byte, including headers. TTFA counts what the caller hears. A vendor quoting model-only latency can be honest and still quote a number you will never see.
How to measure it honestly
- Measure server time on the production endpoint, not a staging box or a co-located benchmark.
- Report both ends of the run: a median with no slow draw beside it hides the worst turn your listeners live in.
- State what the figure includes. Ours is: first audio byte in 146 ms over the open internet, 116 ms p50 first audio, server side warm.
Every term on this page is measurable. Take a key and read the numbers off your own requests.
Get a key and measure it yourself