Text to speech latency
Measured time to first audio next to Amazon Polly, with the method and the limits.
Measured October 29, 2026 on our own test servers in the same cloud region as the API. Not a guarantee; your results depend on location, text length and traffic.
Time to first audio, short sentence
| Voice and server | First audio (median) |
|---|---|
| Catalog voices (GPU server), through the live session and stream path, from a us-east-1 server | 40 to 47 ms (95th percentile 42 to 58 ms) |
| Same, from a laptop over home internet | 53 to 60 ms |
| Phonon Delia and Easton (English), CPU test server, from a us-east-1 server | 25 to 27 ms (95th percentile 30 to 33 ms) |
| Amazon Polly Neural, same server, same run | 61 ms (95th percentile 76 ms) |
| Amazon Polly Standard, same server, same run | 49 ms (95th percentile 79 ms) |
| Your own cloned voice | 1.6 to 2.1 s |
What this does not show
- Phonon Delia and Easton were measured on a test server, over plain HTTP, and are not in the public voice list yet.
- Japanese voices are not included: they failed in this test and we are fixing that.
- The first request after the streaming server wakes can take much longer while the voice model loads.
- These are our own measurements, not a third-party benchmark.
- Audio quality is not part of this test. Listen to the samples.
- There is no uptime or latency guarantee.
Try it
The same API can be called the way you already call Amazon Polly, OpenAI or ElevenLabs: switch from Amazon Polly. Full measurements: speed page.
Questions
How fast is Speakvora text to speech?
In tests we ran on October 29, 2026, the first audio bytes of a short sentence arrived in a median of 40 to 47 ms for catalog voices (English, Spanish, French, Portuguese, Hindi, Chinese and Italian), measured through the real session and stream path with a customer-style API key from a server in the same cloud region (us-east-1). Amazon Polly Neural measured 61 ms and Polly Standard 49 ms from that same server in the same run. These are test numbers, not a guarantee.
Is Speakvora faster than Amazon Polly?
For time to first audio on short sentences in our test, it was faster than Polly Neural (40 to 47 ms against 61 ms) and about level with or a little faster than Polly Standard (49 ms). The differences are tens of milliseconds. Speed is only one factor: Polly has far more voices, languages and regions, and we make no quality claim.
What is not as fast?
Your own cloned voice took 1.6 to 2.1 seconds to first audio in the same tests. Fast streaming also needs a running server: the streaming lane starts when a paying account needs it, and a first request after a quiet period may use the standard path while it wakes.
How many users can one server handle?
One GPU server produced about 900 characters of speech per second in total. With people pausing six seconds between requests it held 15 to 30 users at about 60 ms median first audio (measured earlier on the same hardware); beyond that requests queue or are refused. A CPU server held about two simultaneous streams at this speed.
How do you measure first audio?
The time from sending the request to receiving the first audio bytes over one kept-alive connection, median of repeated runs, with the network round trip included. The headline numbers were measured from a server in the same region as the API, which is the lowest network time you can get; from a laptop expect about 15 ms more. Method and data are in the speed page.