Glossary · Models & serving
What is time to first byte (TTFB)?
Time to first byte (TTS)
Time to first byte, or TTFB, is how long a text-to-speech service takes to send back the first piece of audio after you ask for it. On a live call it decides how soon the caller hears the agent start speaking.
On a call
The agent's model has written "जी, आपका order कल पहुँच जाएगा." The TTS request goes out. TTFB is the wait until the first slice of that sentence's audio arrives, which is the gap the caller hears before "जी".
AI agents from ₹1 a minute.¹
Explained
Here is how it works, in plain words.
The term comes from the web, where it means how long a server takes to start answering a page request. For speech, the idea carries over: the clock starts when you send the text, and stops when the first bytes of audio reach you.
It matters because a streaming voice starts playing as soon as that first slice arrives. The rest of the sentence keeps coming while the caller listens. So on a live call, first-byte time is the part of synthesis the caller actually waits through. Total synthesis time matters much less.
Be careful what you compare. A provider’s quoted TTFB is usually measured close to its own servers. Your number also includes the trip from your server to theirs and back, plus any new connection you open. A fresh HTTPS connection needs a handshake before the first byte can even start, so reuse connections, or keep one socket open for the whole call.
TTFB also changes with load. One request on a quiet server is the best case. Ask for the figure at the number of calls you’ll run at once, and at a high percentile, since a few slow sentences are what callers remember.
And keep it in proportion. On a voice agent, TTFB is one slice of the wait. The caller also waits for the system to notice they’ve finished, for the model’s first sentence, and for the network and playback buffers. A fast voice behind a slow turn detector still feels slow.
A last trap: the first chunk can be tiny. If playback starts on a sliver of audio and then pauses, the caller hears a clipped word. Holding a short buffer before you start playing fixes that.
In Frequency
This is what Frequency does about it.
Every request in your Fourier TTS log shows first_byte_ms, so you can tell our share of a wait from your network's. In our load test with many streams on one GPU, nearly every request got its first audio within 275 ms. That's measured next to the engine, so your own network time comes on top.