Response times
These figures are measured on real API traffic, not on a benchmark. Each row gives the median (half of all calls are faster), the quantity it corresponds to, and the number of calls it is computed from — so that you can judge its solidity for yourself.
| Operation | Median | For | Calls measured |
|---|---|---|---|
| Translation | 1.8 s | ≈ 200 characters | 498 |
| Text to speech | 5.5 s | ≈ 200 characters | 447 |
Measurement window: 60 days, successful calls only. These values are indicative and are not a contractual commitment.
What to expect from the slowest calls (p95)
The median above describes the common case. To set a timeout on the client side, the median is not enough: you need to know how long almost all calls take, not just half of them. Over a recent one-week sample, here is the time within which 95% of successful calls finish (p95), for each operation.
| Operation | p95 (95% of calls) |
|---|---|
| Translation | ≈ 95 s |
| Text to speech | ≈ 34 s |
p95 measurement window: 7 days (a more recent, narrower sample than the median above; the two tables do not cover exactly the same period).
In case of high demand, a call can occasionally take noticeably longer than the median, without exceeding the p95 window above in the vast majority of cases. Recommendations:
- Client timeout: at least 120 seconds for translation, 60 seconds for text to speech.
- Retry with back-off on
503: respect theRetry-After, see the code example in Errors. - For a large volume, prefer several short calls (or several subtitling/dubbing jobs) over one giant call.
Long jobs are not synchronous
Subtitling and dubbing do not return their result in the response: they create a job, and you fetch the result afterwards. Their duration depends on the length of the video, not on an API response time. Use a webhook rather than a polling loop.
The API responds in request/response: each call returns its full result, there is no streaming mode (SSE/WebSocket) yet to receive audio or text incrementally. Well suited to pre-generated messages and announcements; for real-time conversational use (voice chatbot, IVR), this is a current limitation to plan around.