4 ms·
It sounds like they're training a parametric speech synthesis platform on samples in order to learn the parameters. I wonder if there are are approaches at gene
by echelon 10y ago
It sounds like they're training a parametric speech synthesis platform on samples in order to learn the parameters. I wonder if there are are approaches at generating n-phones for concatenative models, or using a hybrid approach.
I built a toy concatenative Donald Trump speech system [1], but I don't have an ML background. I've been taking Andrew Ng's online course in addition to Udacity's deep learning program in an attempt to learn the basics. I'm hoping I can use my dataset to build something backed by ML that sounds better.
Is anyone in the Atlanta area interested in ML? I'd love to chat over coffee or join local ML interest groups.
[1] http://jungle.horse http://jungle.horse
- kastnerkyle 10y agoI tried similar approaches long ago (~2 years now?) with something related to RNN-RBM and it showed some slight glimmer of promise, and still think there might be some clever ways to combine concatenative methods and deep learning to avoid a lot of the noise issues present in parametric models. Then again, maybe it just needs to train longer - it's always hard to tell. I liked jungle.horse, awesome stuff!