Text-to-speech (TTS)

In a phone assistant, TTS is the last step: the language model delivers the text of a reply, and speech synthesis turns it into audio. Modern methods generate the voice with neural networks and sound considerably more natural than earlier voices assembled from recorded sound units.

Three properties matter on a call: how quickly the first syllables can be heard, how the voice pronounces numbers, names and abbreviations, and whether stress and pauses match the content. The best test uses real sentences from the business, such as addresses, times and product names. Voices also differ by language; a voice that is convincing in English can have an accent in German.