TTS (Text-to-Speech)

TTS (text-to-speech) converts the agent's written response into audio. On a voice agent it is the last step in the loop: the caller speaks, speech-to-text transcribes it, the LLM writes a reply, and TTS speaks that reply back.

Why it matters when you are buying or building

Voice quality is the first thing anyone judges, and it is the thing clients have the strongest opinions about. But the quality problem on a phone call is not usually the voice model. Phone audio is narrowband, so a lot of the fidelity a demo showcases disappears the moment the call is real. Always audition voices over an actual call, never through your laptop speakers.

What actually separates good from bad in production:

• Time to first audio. A voice that starts speaking in 200ms feels alive. One that takes a second feels broken. Streaming TTS, which starts speaking before the full sentence is generated, is close to mandatory for phone. See latency.
• Pronunciation of your specific words. Brand names, street names, product SKUs, and clinical terms are where TTS embarrasses you. Most platforms let you override pronunciation. Build a list during testing.
• Handling of numbers and formatting. If the agent writes "$1,200" or a URL, a naive voice reads it literally and awkwardly. The fix lives in the system prompt: instruct the agent to write for the ear, not the page.
• Interruptibility. When a caller talks over the agent, audio has to stop fast. That is barge-in, and it is a platform capability, not a voice one.

What I have seen in production

Across the agents I run, voice choice is more of a brand decision than a technical one. A jewelry retailer's shopping assistant, a UK windows-and-doors company's agent, and a receptionist on my own agency's main line all want different things, and matching the voice to the market matters more than picking the objectively "best" model. The UK build is the clearest case: a US accent on a UK line reads as an outsourced call center, whatever the audio quality.

On cost, TTS is usually billed by characters or by audio duration and is a real line item, though on most small business builds it is not the dominant one. A voice agent's total usage cost lands in cents per minute all-in, and the model and telephony share of that is often larger than the voice share.

One habit worth adopting: have the client listen to two or three candidate voices on a real call and pick. It takes ten minutes and prevents a rebuild after launch, because voice is the one thing everyone at the company will have a reaction to. More in choosing a voice for your agent.