Contents
At a glance
- Category
- Voice and messaging
- Abbreviation
- TTS
- Role in a voice agent
- The third stage - speaks the model's reply to the caller
- Metric that matters
- Time to first audio byte, not total generation time
- Requirement for live calls
- Streaming synthesis, so playback starts on the first chunk
- Billing
- Per character or per second of audio - verify current rates before scaling
What text to speech does
Text to speech takes written text and produces audio. In a voice AI agent it is the last of three stages: speech to text transcribes what the caller said, a language model decides the reply, and TTS speaks it. Modern neural TTS is convincingly human on ordinary sentences, which is why phone automation stopped sounding like a robot somewhere around 2023.
The quality question in a live call is not mainly how natural the voice is - it is how fast the first audio arrives. A perfect voice that begins speaking a second and a half after the caller stops talking feels broken; a slightly less polished voice that starts in 300 milliseconds feels like a conversation.
Streaming is the thing that matters
Non-streaming TTS generates the whole clip and then returns it, so the caller waits for the entire sentence to be synthesised. Streaming TTS returns audio as it is produced, so playback starts on the first chunk. For live calls, streaming is not an optimisation, it is a requirement.
The metric to ask any vendor for is time to first byte, not total generation time. Combined with the model's own first-token latency it determines whether your agent feels responsive - and it is the biggest single line in a latency budget. Providers publish flagship low-latency models specifically for this; check the current figures rather than trusting a benchmark from last year.
Cost and control
TTS is billed per character or per second of audio, and it is usually the smaller line in a voice agent next to telephony minutes and model tokens. Prices and model tiers move, so verify current rates with the provider before budgeting a high-volume deployment.
Control matters as much as price. Look for SSML or equivalent markup so you can fix pronunciation of names, product terms and postcodes; number and date formatting, since "2/3" should be read as a date or a fraction depending on context; and word or phrase overrides for anything the model consistently mispronounces. Businesses with unusual trade names spend more time on pronunciation dictionaries than they expect.
What you should know
Voice cloning is offered by most vendors and carries real obligations. Cloning a specific person's voice needs that person's documented consent, and several jurisdictions now regulate synthetic voice in calls. Separately, US rules on AI-generated voices in robocalls have tightened, so anything outbound and automated needs checking against current FCC guidance and your telephony provider's own policy before it dials.
On disclosure, the practical advice is to be upfront regardless of what the law strictly requires. Callers work it out, and an agent that opens by saying it is an automated assistant that can book, reschedule or pass them to a person gets better cooperation than one pretending to be a receptionist called Sarah. Keep a fast route to a human on every call - see IVR and barge-in.
Text to speech FAQ
What is text to speech?
Software that converts written text into spoken audio. In a voice AI agent it is the final stage: the caller's speech is transcribed, a language model composes a reply, and text to speech turns that reply into the voice the caller hears.
Why does streaming TTS matter for phone calls?
Because non-streaming synthesis makes the caller wait for the whole sentence to be generated before anything plays. Streaming starts playback on the first chunk of audio, which is the difference between a conversation and an awkward pause after everything the caller says.
How much does text to speech cost?
It is billed per character or per second of audio and is usually the smaller line in a voice agent, behind telephony minutes and model tokens. Rates and model tiers change regularly, so check the provider's current pricing before budgeting high volume.
How do I fix mispronounced business or product names?
Use SSML or the provider's phoneme and alias controls to specify pronunciation explicitly, and build a small dictionary of your trade names, product terms and local place names. Businesses with unusual names spend meaningfully more time on this than they expect.
Is it legal to use a synthetic voice on calls?
It depends on jurisdiction and on whether the call is inbound or outbound. US rules on AI-generated voices in robocalls have tightened, and several jurisdictions regulate synthetic voice disclosure. Check current regulator guidance and your telephony provider's policy before running outbound campaigns.
Should the agent tell callers it is AI?
Yes, as a matter of practice as well as compliance. Callers work it out anyway, and an agent that opens by saying it is an automated assistant which can book, reschedule or hand over to a person gets more cooperation than one pretending to be a named receptionist.
Primary source
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what text to speech means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
