Text-to-speech (TTS)

What it is

Text-to-speech synthesises the agent's reply as audio. In a voice agent it runs in streaming mode: the language model produces text a few words at a time and the synthesiser starts speaking before the sentence is complete. Vendors offer a catalogue of voices, sometimes cloned voices, and controls for speed, pronunciation and pauses.

Two things distinguish telephone-grade TTS from a demo voice. First, time to first audio: how long after the text starts arriving does the caller hear something? Second, correctness on structured content: phone numbers read digit by digit, dates in the caller's convention, currency, postcodes and proper names.

Why it matters when buying

Voice quality is what vendors demo. Latency and pronunciation are what callers notice. A beautiful voice that takes a second to start, or that reads "Dr" as "drive", loses trust on the first turn. Voice cloning of a real employee raises consent questions, and several jurisdictions now expect a synthetic voice to be disclosed (see EU AI Act Article 50).

What to ask

Ask for time to first audio measured on a phone call, not in a browser. Ask the agent to read back your own address, a phone number and three staff names during the demo. Ask whether the synthesiser is the vendor's own or licensed, where it runs, and whether audio from your calls is used to train voices. Ask whether the caller can interrupt the voice mid-sentence and how quickly it stops.