News for hackers. Stuff that matters.
Re:Tacotron 2: Generating Human-Like Speech from Text by mhh__ on Tuesday December 19, 2017 @03:14PM (#15963791)
Apparently not? I imagine it would have to be replicated by someone else. One also assumes that the volume of training data - I don't actually know - is not insignificant!
Re:Tacotron 2: Generating Human-Like Speech from Text by heinrichf on Tuesday December 19, 2017 @05:19PM (#15964951)
From the paper, the training data is surprisingly small:

"We train all models on an internal US English dataset, which contains 24.6 hours of speech from a single professional female speaker."

Re:Tacotron 2: Generating Human-Like Speech from Text by mhh__ on Tuesday December 19, 2017 @06:06PM (#15965303)
Uncompressed, that's still quite a lot of data - perhaps not relative to other ML projects, but still.
Re:Tacotron 2: Generating Human-Like Speech from Text by 21 on Tuesday December 19, 2017 @07:40PM (#15965948)
They say they split the audio in "80-dimensional audio spectrogram with frames computed every 12.5 milliseconds". The picture from the post supports that.

For 24 hours that would be 7 mil frames of 80 values ~= 500 mil data points, or 2 GB of raw data (assuming floats)

Re:Tacotron 2: Generating Human-Like Speech from Text by PeterisP on Wednesday December 20, 2017 @09:20AM (#15969667)
So, something that easily fits into RAM, One might even keep a copy of the whole dataset in GPU memory to avoid copying to and from even if matrix operations are done on smaller batches.

E.g. ImageNet is 50 gb of compressed data, and there are many much larger datasets in practical use.