Google2017
Tacotron
AR seq2seqresearchvendor demo
Ending the era of hand-built TTS frontends.
03 · 2017–
Autoregressive seq2seq
Tacotron reads characters the way a translator reads a sentence: one spectrogram frame at a time, each frame looking at the last.
Playable examples
Tacotron and Tacotron 2 read the same out-of-domain sentence about GANs and VAEs. The second clip is the vocoder growing up.
Tacotron (2017) trained a sequence-to-sequence model from characters to spectrograms, with no linguistic frontend. Tacotron 2 kept the predictor and attached a WaveNet vocoder. The MOS numbers in that paper — 4.53 against 4.58 for professional recordings — are why every lab spent two years on this stack.
WaveNet itself (DeepMind, 2016) modelled raw audio sample by sample. It was too slow to be a full TTS system, and exactly fast enough to be a vocoder once Tacotron gave it a mel spectrogram. Google Cloud WaveNet voices are the production echo.
The cost of autoregression is left-to-right decoding and the occasional failure to stop, skip, or repeat. FastSpeech was a reply to that cost. Instructable APIs such as gpt-4o-mini-tts still live in this family: a neural acoustic model, a vocoder, a text encoder, and (now) a language model telling the voice how to say it.
On this plate
Tacotron and Tacotron 2 read the same out-of-domain sentence about GANs and VAEs. The second clip is the vocoder growing up.
Google2017
Tacotron 2
AR seq2seqresearchvendor demo
The canonical two-stage neural TTS recipe.
Google Cloud / DeepMind2016
Google WaveNet
AR seq2seqcommercialvendor demo
The moment neural vocoders entered production.
OpenAI2023
OpenAI TTS
AR seq2seqcommercialemotionvendor demo
Steerability through instructions; one-vendor convenience.