Atlas record: https://www.aitimeline.dev/en/events/evt-0772/
Research · March 29, 2017
Tacotron: speech straight from letters
On 29 March 2017 Yuxuan Wang and colleagues at Google described Tacotron, a model that synthesises speech directly from text characters and is trained from scratch on text-audio pairs. Its mean opinion score for naturalness was 3.82 of 5, against 3.69 for a production parametric system.
Why it matters
Speech synthesis, built from a text-analysis frontend, an acoustic model and a vocoder, became one network that can be trained without phonetic expertise. It opened the way to neural voices learned from recordings alone.
Training used about 24.6 hours of internal North American English; the waveform is recovered with the Griffin-Lim algorithm. A concatenative system scored 4.09 in the same test, so Tacotron did not reach it. Generating by frame makes the model substantially faster than methods that generate sample by sample.
Event record
- Event date
- March 29, 2017
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 25, 2026
- Lines
- ID
- evt-0772
The day the first version of the preprint was submitted; the figures were read in it. The second version of 6 April 2017 renamed it "Tacotron: Towards End-to-End Speech Synthesis".
Related events
Earlier
- Contrasts WaveNet 2016
The authors set themselves against WaveNet: slow because it generates sample by sample, and not end to end because it needs linguistic features from a separate frontend; Tacotron generates by frame and learns from text-audio pairs.
- Builds on Attention for translation 2014
Tacotron's backbone is a seq2seq model with attention after Bahdanau et al.: text characters are translated into spectrogram frames.
- Related A voice from a statistical model: HTS 1999links to this record
Tacotron compares itself with a parametric system in production at Google: 3.82 against 3.69.
- Related Unit selection: a voice from pieces of recording 1996links to this record
Tacotron is compared with a unit selection system in production at Google: 4.09 against its own 3.82. It does not cite Hunt and Black.
Later