Text-to-speech. TTS, is the last layer in the voice-agent pipeline. The model produces text, TTS turns it into audio, and the caller hears your agent's voice. The quality determines whether the customer feels they're talking to a person or a phone tree.
Modern TTS is generative, the voice is synthesized from a learned model, not stitched from recorded samples. The best generation in 2026 is indistinguishable from a human voice in a thirty-second sample. In a five-minute call, with prosody and pacing variations, the gap is still detectable but narrowing.
The practical choices for a buyer are the voice library (which voices ship and whether you can custom-record one), the latency to first audible word (most TTS streams audio so the first chunk lands in 120-240 ms), and the SSML / control surface for tone and emphasis.
Vorel ships per-customer voice selection from ElevenLabs + a custom voice option for enterprise plans. Latency-to-first-audio-chunk is part of the contract SLA.