How we measure TTS latency.

Every latency figure on this site comes from one harness with one set of rules, described here so you can hold us to it — or rerun it against us on a call.

5 rules

behind every figure on this site

The rules

Production API only — the endpoint customers ship on, not a staging box or a lab build.
Server time to first audio: request accepted to first playable byte, includes gateway, queue, and synthesis. Model-only time is never quoted.
Call-shaped load: short conversational utterances with cloned voices, not paragraph batches with stock ones.
Percentile pairs, always: a p50 is published only next to its p95.
Conditions stated with every figure: stream count, hardware share, failure count.

What the clock includes

The single-stream figure this harness produces is 107 ms p50, 108 ms p95. Here is exactly where that clock starts and stops:

Fig. — The measurement window behind 107 / 108 ms

request sent →

API gateway

queue

synthesis

→ first playable byte

server time to first audio — the published number

Your network sits outside this clock on both ends — measure it separately, because we cannot measure it for you. Model-only time would be a subset of the third segment alone.

What we do not publish

Numbers we cannot reproduce on demand. Under-load figures live on their own page, with the harness pointed at one GPU and the conditions stated. The way to check any figure is to bring your script to a 30-minute call and watch the readout while your own prompts synthesize — the bench runs on the production API, and the recording leaves with you.

Related sheets.

benchmark

813 ms

How the API behaves under load

What one GPU does under 16 concurrent callers — 430 ms median, 813 ms p95, zero failed requests — and how the API refuses traffic past absolute capacity.

glossary

107 / 108

Time to first audio (TTFA)

TTFA is the delay from sending text to the first playable audio byte. How it differs from model latency and TTFB, and how to measure it honestly.

glossary

< 1

Real-time factor (RTF)

RTF is synthesis time divided by audio duration. Why RTF below 1 is table stakes, and why time to first audio matters more for conversations.

The harness reruns on demand, on the production API, with your script in it.

Watch us rerun this on a call