What is text-to-speech?
Text-to-speech (TTS) is the generation of spoken audio from written text by software. Older systems stitched together recorded fragments and sounded mechanical. Current neural TTS models generate the waveform directly and produce speech with natural rhythm, emphasis, and pauses, in many languages and voices. Some can clone a specific voice from a short sample. TTS is the output side of a voice interface; speech recognition is the input side, and conversational AI connects the two.
Text-to-speech in practice
A regional bank uses TTS in its phone system to read balances, transaction details, and branch hours in a consistent voice, including names and amounts that cannot be pre-recorded. A logistics company generates spoken dock and routing instructions for drivers from dispatch data so they do not read a screen while driving. A hospital produces audio versions of discharge instructions in the patient’s language. In each case the text comes from a system of record and the voice adds only the delivery.
What to watch for
- Check pronunciation of names, drug names, and product codes, and maintain a pronunciation dictionary for the ones the model gets wrong.
- Decide whether to disclose that the voice is synthetic; callers increasingly expect it, and the EU AI Act requires it in some settings.
- Control voice cloning tightly, because the same capability that produces a branded voice also produces a deepfake of an executive.
- Test latency for live use, since a voice assistant that pauses for seconds before speaking feels broken.