Low latency text to speech for voice agents
Streaming speech with a measured time to first audio, the call, the price and the limits.
Measured October 29, 2026 on our own test servers. Our measurement, not a third-party benchmark.
What a voice agent needs from text to speech
- A low time to first audio, so the reply starts quickly.
- Streaming, so audio plays while the rest is generated.
- A price that does not punish short replies, since agents speak many short lines.
Measured
| Test | First audio (median) |
|---|---|
| Catalog voices (GPU server), session and stream path, from a server in us-east-1 | 40 to 47 ms |
| Amazon Polly Neural, same server, same run | 61 ms |
| Amazon Polly Standard, same server, same run | 49 ms |
The stream, in four lines
s = requests.post("https://speakvora.com/api/v1/session", headers={"x-api-key": KEY}).json()
r = requests.post(s["stream_url"], headers={"authorization": "Bearer " + s["token"]},
json={"text": "Your order is ready.", "voice": "af_heart", "lang": "en", "format": "pcm", "request_id": "r1"}, stream=True)
for chunk in r.iter_content(2048): play(chunk) # 24 kHz, 16-bit, monoWhere it does not fit
- No SSML and no speech marks.
- Fewer voices and languages than the large providers; listen to the samples before choosing.
- The streaming server starts on demand, so plan for a slow first request after a quiet period.
Details and method: text to speech latency. Using a different client? switch from Amazon Polly, OpenAI or ElevenLabs.
Questions
What is a good time to first audio for a voice agent?
People notice a pause once the whole reply takes a few hundred milliseconds. Text to speech is one part of that delay, next to speech recognition, the language model and the phone network, so a lower first-audio time leaves more room for the rest.
How fast is Speakvora streaming text to speech?
In our tests on October 29, 2026, from a server in the same cloud region, the first audio bytes of a short sentence arrived in a median of 40 to 47 ms for catalog voices. That measures the first bytes, not the first audible sound, and it is our own measurement.
How does streaming work?
Call POST /api/v1/session with your API key to get a short-lived token, then POST text to the streaming address with that token. Audio comes back as raw 24 kHz 16-bit PCM over one chunked response, sentence by sentence, so playback can start before the whole reply is generated.
What does it cost?
$3 per million characters for catalog voices, with 100,000 characters free. A typical agent reply is 100 to 200 characters, so 10,000 replies is about 1 to 2 million characters.
What are the limits?
Streaming runs on a server that starts when a paying account needs it, so the first request after a quiet period can be slow. Japanese is not available on the streaming lane. Your own cloned voice takes 1.6 to 2.1 seconds to first audio. There is no uptime guarantee yet.