Vendor / paper demo
Tacotron 2 paper demo · Tacotron 2 official demo, gan_or_vae.wav.
Generative adversarial network or variational auto-encoder.
Google · 2017 · Tacotron + WaveNet
Tacotron 2
Tacotron’s spectrogram predictor plus a WaveNet vocoder. MOS 4.53, next to 4.58 for professional recordings, on the paper’s test. This is the ‘AR seq2seq + neural vocoder’ stack that FastSpeech later made non-autoregressive.
Known for
The canonical two-stage neural TTS recipe.
Vendor / paper demo
LJSpeech (FastSpeech comparison) · Tacotron 2 baseline from the FastSpeech demo page, sample 1.
I will quote an extract from the reverend gentleman's own journal.
- Approach
- AR seq2seq
- Year / era
- 2017
- License class
- research
- Cloning
- No
- Streaming
- No
- Emotion controls
- No
- Languages (tags)
- en
- Official
- Source