Skip to content

Feature · Streaming TTS

Streaming TTS that starts talking in about 53 ms.

PyAI Speak sends the first audio chunk while later words are still being synthesized, so agents skip dead air.

Get a sandbox key$50 credit. No credit card required.
~53 ms warm to first byte~98 ms on a cold startChunked HTTP body by default$0.04/min, no platform fee

What PyAI gives you for streaming TTS

  • Streaming TTS on PyAI Speak returns the first audio chunk in about 53 ms warm in region at $0.04 a minute, so a voice app starts talking before the rest of the reply has finished rendering.
  • Cold requests reached first byte in about 98 ms in the same 24 June 2026 test, while cutting the model's token stream at each full stop removes the silence of waiting for the full answer.
  • Inside an Omni voice agent the sentence splitting is built in, and users hear a reply about 390 ms after a turn ends, measured in the US region.
  • Streaming costs the same $0.04 a minute as a buffered file with no streaming surcharge, and a cancelled stream bills only the seconds already delivered.

When does the first sound play in each kind of request?

Choose a request type and press play. Timings are an illustration built from PyAI benchmark figures, not a capture.

Pick a moment, then press play

Your appPyAI SpeakClock
Ready00:00

How does PyAI streaming TTS work, step by step?

A single Speak call passes one access check at the PyAI Edge, then synthesis and delivery overlap so the first piece leaves early.

Your code posts a sentence

The request carries the text, a voice id, an audio format and streaming switched on, sent to the PyAI Speak API.

The Edge checks access once

Key, speech permission and credit balance are verified a single time, with no rerouting afterwards.

The voice id picks the lane

Stock, designed or cloned, the voice id settles both the voice and the language of the audio.

Chunks leave as they render

Each piece passes a quality guard and a loudness pass, gets converted to your format, then joins the chunked body.

The exact text was spoken before

A clip cache can answer an identical request at once. Billing matches a fresh render.

Your listener disconnects mid-sentence

Metering finalizes on the audio seconds actually delivered, so the unplayed tail costs nothing.

Where does streaming text to speech matter most?

Streaming matters wherever a person is waiting for a voice to reply.

AI agents

Voice agents with no dead air

Silence after a user finishes sounds like a dropped connection in an app, a car or a phone line. Omni speaks the first sentence of a reply while the model is still writing the second.

Chat assistants

Model answers read sentence by sentence

A support assistant cuts the token stream at every full stop and posts each sentence to PyAI Speak, so speech starts after sentence one rather than after the full answer.

Learning apps

Tap to listen on long text

A lesson or article starts reading about 53 ms warm after the tap, instead of making the learner wait for a long file to download.

Games and kiosks

Spoken lines generated on the fly

Dialogue written at runtime can be voiced as it is produced, which keeps a character or a kiosk responsive when no line was recorded ahead.

Streaming TTS questions people ask

How do I cut latency when a language model feeds the TTS?

Stop waiting for the full model answer. Cut the token stream at each full stop and post that sentence to the speech API immediately, so audio starts after sentence one. PyAI Omni agents run this splitting inside one session and reply about 390 ms median after a user stops talking.

Is first byte the same as the first word a listener hears?

No. First byte marks when audio data leaves the server, while the first audible word can come a little later if a chunk opens with near silence. PyAI publishes about 53 ms warm to the first byte, so teams should also time audible speech on their own playback device.

Does PyAI charge extra for streaming TTS?

No. A streamed PyAI Speak request costs the same $0.04/min of delivered audio as a buffered file, with no platform fee. A stream your app cancels early bills only the seconds already sent, and silence or malformed audio is caught before billing. Starter credit is $50.

Is PyAI Speak a TTS tool for Twitch streamers?

No. Twitch and YouTube streamers who want chat or donation messages read aloud need a streamer chat reader with overlays and alerts. PyAI Speak is an API for developers building assistants, apps and voice agents, where audio must start in milliseconds after a sentence is ready.

PyAI is voice infrastructure. It hands you the audio and transcript, and does not replace your own models or review tools.

Post one sentence and time the opening chunk yourself.

Start building with $50 creditSandbox keys never bill, and you can start without a credit card.

One stack: Omni, Hear, Speak, Recap and Trace.