Text to speech latency
Latency is the wait before speech starts. Here is what drives it and how to cut it.
- First-byte latency is the key metric
- Streaming reaches sub-300ms first byte
- File synthesis waits for the full render
- Matters most for real-time apps
Definition
Text-to-speech latency is the delay between sending text and hearing speech. For anything live, the number that counts is first-byte latency: the time until the first chunk of audio arrives. A low first-byte time is what makes a voice interaction feel immediate instead of laggy.
First byte VS total
There are two clocks. First-byte latency is how long until playback can start. Total render time is how long until the whole clip is ready. For a real-time app, first-byte is what the user feels. For pre-rendered content, total time is what matters. Streaming optimizes the first; file synthesis optimizes for a finished file.
How to cut it
Use streaming. It returns audio in chunks as the model generates them, so the client plays the opening while the remainder is still being produced. That overlap collapses the perceived wait to the sub-300ms first-byte time rather than the full render.
Where it matters
Voice agents, live read-aloud, and interactive alerts all depend on low first-byte latency; a pause before speech reads as a fault. An audiobook pipeline cares less, because nobody is waiting live. Match the mode to the use case. See the streaming guide.
Frequently asked questions
What is text-to-speech latency?
How do I reduce TTS latency?
When does latency matter most?
Start building
Latency is the wait before speech starts. Here is what drives it and how to cut it.