TTS (Text-to-Speech)

Whatever the language model decides to say, the caller only hears it through text-to-speech. TTS is the last step in the reply path of a voice agent and the part callers judge most harshly: a flat or robotic voice makes people hang up, and a voice that mispronounces their street name makes them doubt everything else it says.

How it works. The language model emits the reply as a stream of tokens. The TTS engine begins synthesizing audio from the first clause while later clauses are still being generated, so time to first audio is measured in hundreds of milliseconds instead of waiting for the full sentence. The engine takes text, a voice identity and optional controls for speed and emphasis, and returns audio in the codec the telephony leg expects. When the caller interrupts, playback stops and the remaining audio buffer is discarded.

Why it matters for voice agents. Numbers, times, names and addresses are where synthesis breaks. "10:30" read as "ten thirty" is right; read as "ten colon three zero" it is wrong. A long pause before the first word reads as the agent not understanding. Voice choice also carries brand: a clinic and a debt-collection line should not sound the same. Because audio is the whole interface, a TTS defect is a customer-facing defect on every call.

Example. A dental practice uses a voice agent for rescheduling. A patient asks to move a cleaning. The agent replies: "I have Thursday the 16th at 2:15 pm or Friday at 9 am. Which works for you?" Good synthesis reads the ordinal, the time and the day naturally and pauses at the question. The patient picks Friday, and the confirmation is read back with the practice's address spoken street by street rather than as one run-on string. The booking lands in the calendar the agent already checked.

TTS vs voice cloning. Standard TTS uses a library voice supplied by the model vendor. Voice cloning builds a voice from recordings of a specific person and requires that person's documented consent. Most business lines use a library voice chosen to fit the brand; cloning is a separate decision with its own legal review.

What to measure.

• Time to first audio, measured on a real phone call rather than in a browser.
• Pronunciation accuracy on a test set of your own names, addresses, prices and times.
• Naturalness rating from a small panel, repeated when the voice or model changes.
• Consistency across long calls: the voice should not drift in pace or pitch.
• Barge-in stop time, from the caller starting to speak to playback stopping.

Voice language coverage on Botcadence depends on the voice model selected for the agent, while chat agents can reply across 50+ languages. You can hear the production voice on the US demo line at +1 (469) 454-2702 before going live; see voice agents and multilingual support.