Concatenative & formant

Formant synthesis (eSpeak, older DECtalk) does not play a recording. It computes a source tone and a set of resonant filters that mimic a throat. The result is small, fast, and unmistakably synthetic — the sound people mean when they say ‘computer voice.’ eSpeak NG still ships in phones, and it is the phonemizer sitting under Piper and Kokoro.

Concatenative TTS records a speaker, slices the recording into diphones or longer units, and concatenates the slices to match new text. Festival’s kal voice and Flite’s slt voice on this bench are diphone systems. You can hear the splice points: a click, a pitch jump, a vowel that belongs to a different word.

Unit-selection systems (later commercial concatenative engines) searched a huge recorded corpus for longer matches, which hid some joins and created a new failure: a perfectly natural syllable next to a dead one. Neural models did not make concatenation obsolete so much as they made the joins disappear.