What is text-to-speech?
It is why a voice agent with a beautiful voice can still sound robotic on a call. The first half-second of silence gives it away before the voice ever does.
Text-to-speech (TTS)
speech synthesis
Text-to-speech is the technology that converts written text into spoken audio. In a voice agent, it turns the model's reply into the words a caller hears. How quickly that audio starts is what a caller actually judges.
Vendors sell TTS on voice quality. They demo a warm, human-sounding accent and call it done. On a real phone call, quality is not what a caller notices first. Timing is.
A natural voice that takes a beat and a half to start still reads as a machine thinking. A slightly flatter voice that starts almost instantly reads as attentive. Callers forgive tone. They do not forgive silence.
The metric vendors don't lead with
Time-to-first-audio is how long a caller waits between the model finishing a thought and the first sound reaching their ear. It is separate from how the voice sounds once it starts. It is the number that decides whether a phone agent feels present or feels like a hold queue.
A voice agent has to do three things before a caller hears anything: transcribe the speech, generate a reply, and speak it back. TTS is the last of those three steps, so any delay it adds stacks on top of the other two. A caller does not know which step was slow. They only know the reply arrived late.
Streaming TTS shrinks this wait. It speaks the start of a sentence while the rest is still generating, instead of waiting for the full reply. On ZhoopZhoop's inbound voice agent, that is the difference between a reply starting mid-thought and dead air while a full sentence renders.
- Model repliesText generates, often token by token.
- First chunk readyEnough text exists to start speaking.
- TTS startsAudio synthesis begins on that chunk.
- First audio outCaller hears sound. This is the number that matters.
- Rest streamsRemaining audio follows while the caller listens.
Step four is time-to-first-audio. Everything before it is invisible to the caller; everything after it just needs to keep up.
Common questions
01Is TTS the same as a voice agent?
No. TTS is one component. A voice agent also needs speech recognition to hear the caller and a model to decide what to say. TTS is the final step that speaks the reply back.
02Does a more natural-sounding voice mean better TTS?
Not by itself. Voice quality and response speed are separate properties. A convincing voice that starts late still feels robotic on a call. Callers react to timing before they react to tone.
03Why does TTS latency matter more on phone calls than in text chat?
A chat reader can see a reply streaming in and knows it is coming. A phone caller only has silence or sound. Any gap reads as the system doing nothing, so the latency budget is stricter on a call than on a screen.
