Text to speech latency by language
Measured time to first audio for each voice, with the method and the gaps.
Measured October 29, 2026 on our own servers; 30 requests per voice. Our measurement of the first audio bytes, not a third-party benchmark.
Time to first audio by voice
| Language | Voice | Median | 95th percentile |
|---|---|---|---|
| English | af_heart | 41 ms | 42 ms |
| English | am_michael | 42 ms | 42 ms |
| Spanish | ef_dora | 40 ms | 53 ms |
| Spanish | em_alex | 40 ms | 53 ms |
| French | ff_siwis | 41 ms | 51 ms |
| Portuguese | pf_dora | 42 ms | 56 ms |
| Portuguese | pm_alex | 41 ms | 55 ms |
| Hindi | hf_alpha | 44 ms | 58 ms |
| Hindi | hm_omega | 45 ms | 58 ms |
| Chinese | zf_001 | 47 ms | 56 ms |
| Chinese | zm_009 | 46 ms | 55 ms |
| Italian | if_sara | 41 ms | 54 ms |
| Italian | im_nicola | 41 ms | 56 ms |
What this does not show
- Japanese is not on the streaming lane.
- Voice quality and pronunciation: the test sentences were not reviewed by native speakers.
- The first request in a language after the server wakes is slower.
- Our own cloned voice is English only and takes 1.6 to 2.1 seconds to first audio.
English comparison with Amazon Polly (Neural 61 ms, Standard 49 ms in the same run): text to speech latency. Prices: pricing.
Questions
How fast is Speakvora text to speech in languages other than English?
In our tests on October 29, 2026, the first audio bytes of a short sentence arrived in a median of 40 to 47 ms in English, Spanish, French, Portuguese, Hindi, Chinese and Italian, from a server in the same cloud region as the API.
Is Japanese included?
No. Japanese voices are not served on the streaming lane: they failed in this test and were removed from it. Japanese is still available as stored audio through the standard API.
How was each language tested?
Three short customer-service sentences per language (about 30 to 110 characters), 10 requests each plus a warm-up, through the session and stream path with an API key, over one kept-alive connection. The sentences were written by us and not reviewed by native speakers, and this test says nothing about how natural a voice sounds.
Why is Chinese a little slower?
Chinese measured 46 to 47 ms against 40 to 45 ms for the others. We have not investigated the cause.
What about the first request in a language?
The first request in each language after the server starts compiles the model for that language and can take from a few seconds to over two minutes. Plan a warm-up request after the streaming server wakes.