Tacotron-2: How long did it take you to write this project?

[sunnnnnnnny]

it is amazing

[Rayhane-mamah]

Thank you!

I started working on it 1 week after the first DeepMind Tacotron-2 paper release which is early January 2018. So overall lifetime of the repo is about 11~12 months. :)
Overall the repo covers Tacotron-2 paper contents with few extras that either help train speed or makes the paper easier to reproduce on a relatively affordable machine. Many people helped achieving that and I appreciate it a lot.

There were few times where I stopped working on it due to studies/absence of computation resources/other projects came up.

Considering that, at the time I started the project I only had academic knowledge about deep learning, I would say that for a well experienced person, it only takes about 2 to 3 months to do it from scratch (from no knowledge about tacotron whatsoever to reproducing paper ).

Thanks again for the support :)

[statcom]

Thanks for sharing your work.

From scratch, I was able to train your model with LJSpeech data and my rig (AMD TR 1920 with 2 x GTX 1080). After 15 hours, it's still running with rate of 1000 steps per 1 hour, but the trained Tacotron-2 model already synthesizes human-like speech, a lot better than traditional/commercial TTS just after 10 hours (~10,000 steps) so I could feel how fast deep learning is changing the world.

[cnlinxi]

Great work! I am your fans@Rayhane-mamah