CSTR / Nagoya Institute of Technology2002
Merlin / HTS
HMM / parametricresearchstreamvendor demo
The pre-neural baseline every 2017–2019 paper had to beat.
02 · 1990s–2016
HMM / statistical parametric
Instead of gluing recordings, predict a compact description of speech and run it through a vocoder. The muffling is the method.
Playable examples
The FastSpeech demo page put Merlin next to Tacotron 2 on the same sentence. That pair is the plot twist of 2019.
Hidden Markov Model TTS, and the DNN systems that replaced the HMMs, treat speech as a sequence of statistical features: spectral envelopes, pitch, aperiodic energy. HTS and Merlin are the names on the papers. Google Cloud’s Standard voices are a production cousin of the same idea.
A vocoder (STRAIGHT, WORLD, later neural vocoders) turns those features back into a waveform. Early vocoders subtracted too much: the famous ‘buzzy’ colour. That colour is not a bad recording. It is missing phase and missing residual.
Parametric systems won on flexibility — you could scale pitch, average speakers, speak a language you had little audio for — and lost on presence. WaveNet, in 2016, was in part an argument with the vocoder.
On this plate
The FastSpeech demo page put Merlin next to Tacotron 2 on the same sentence. That pair is the plot twist of 2019.
Google Cloud2014
Google Standard
HMM / parametriccommercialstreamvendor demo
Cheap, intelligible, obviously synthetic.