Show HN: I made a neural text to speech website as my pandemic side project

vo.codes

30 pointsechelon5 comments

This was a long project that required a lot of work. The quality of sample data hugely impacts the results, and so I built a (mostly automated) pipeline to clean, annotate, and improve the data. I have no doubt that the companies that build ML training sets will hold an edge in the future.

I'll be happy to answer any questions if there are any. I learned a lot from this. :)

This looks really impressive, do you have any writeup on how you coded all of it?

Not yet, but that's something I can put together!

There are a lot of short cuts I wish I'd known at the outset, mostly in terms of curating training data and monitoring the learning, but I also learned a few new systems-y things from the overall engineering project.

I'd love to read that if you ever put it together!

How many hours of audio did it need to train one of the voices?

It varied a lot. Some speakers (that sound great) only had about 45 minutes of audio. Other speakers had up to five hours.

The key to training is that all of the models were transfer learned from the Linda Johnson speech dataset (LJS).

Reply on news.ycombinator.com