Classic speech synthesis is interesting, in that relatively simple approaches, produce useful results. Formant synthesis takes relatively simple sounds, and modifies them according to the various distinctions the human speech tract can make. The basic vowel quality can be modelled as two sine waves that change over time. (Nothing more complex than what's needed to generate touch tone dialing tones, basically.) Add a few types of buzzing or clicking noises before or after that for consonants, and you're halfway there. The technique predates computers; it's basically the same technique used by the original voder [1] just under computer control.
Join that with algorithms which can translate English into phonetic tokens with relatively high accuracy, and you have speech synthesis. Make the dictionary big enough, add enough finesse, and a few hundred rules about transitioning from phoneme to phoneme, and it's produces relatively understandable speech.
Part of me feels that we are losing something, moving away from these classic approaches to AI. It used to be that, to teach a machine how to speak, or translate, the designer of the system had to understand how language worked. Sometimes these models percolated back into broader thinking about language. Formant synthesis ended up being an inspiration to some ideas for how the brain recognizes phonemes. (Or maybe that worked in both directions.) It was thought, further advances would come from better theories about language, better abstractions. Deep learning has produced far better systems than the classic approach, but they also offer little in terms of understanding or simplifying.
[1] https://en.wikipedia.org/wiki/Voder
Join that with algorithms which can translate English into phonetic tokens with relatively high accuracy, and you have speech synthesis. Make the dictionary big enough, add enough finesse, and a few hundred rules about transitioning from phoneme to phoneme, and it's produces relatively understandable speech.
Part of me feels that we are losing something, moving away from these classic approaches to AI. It used to be that, to teach a machine how to speak, or translate, the designer of the system had to understand how language worked. Sometimes these models percolated back into broader thinking about language. Formant synthesis ended up being an inspiration to some ideas for how the brain recognizes phonemes. (Or maybe that worked in both directions.) It was thought, further advances would come from better theories about language, better abstractions. Deep learning has produced far better systems than the classic approach, but they also offer little in terms of understanding or simplifying.
[1] https://en.wikipedia.org/wiki/Voder