What kind of speech synthesis is "HMM synthesis"?

Question

The Wikipedia page on synthetic speech claims:

The two primary technologies generating synthetic speech waveforms are concatenative synthesis and formant synthesis.

It then goes on to describe those technologies, as well as a separate section on "HMM-synthesis" almost as an afterthought:

HMM-based synthesis is a synthesis method based on hidden Markov models, also called Statistical Parametric Synthesis. In this system, the frequency spectrum (vocal tract), fundamental frequency (voice source), and duration (prosody) of speech are modeled simultaneously by HMMs. Speech waveforms are generated from HMMs themselves based on the maximum likelihood criterion.[34]

The link at the end takes the reader to a webpage without much description of what hmm-synthesis is. But maybe those of you more familiar with speech synthesis have some insight.

Is hmm-synthesis Concatenative synthesis, Formant synthesis, or something else altogether? I am boggled so any help would be great. Just want to get an idea of the kinds of speech synthesis so I understand better the overall strategies for synthesis speech generation. Thanks!

Answer 1

HMM-based systems and concatenative waveform-based unit-selection systems are similar in that they both involve the use of a large database of recorded human speech, and they both involve concatenating stored units to construct something that results in a whole output utterance.

However, whereas waveform-based unit-selection systems store actual waveforms from the database and concatenate those waveforms, HMM-based systems store HMMs that have been trained on the database and concatenate the HMMs. These HMM "units" are used to construct a single HMM for the desired output utterance. A set of parameter values is generated from that utterance HMM, and a synthesizer synthesizes the output based on those parameter values.

You can find many resources for further reading by Googling speech synthesis based on hidden markov models.