Does the synthesizer only work on "middle length" (+/- 20 words) sentences?

[akrokodile]

Beyond 20 words, it seems to talk a blue streak.
Below 10, it produces pauses and inhumanly noises.
I've tried padding the short sentences with words consisting of the single letter "s" and then subtracting int(0.25 * Synthesizer.sample_rate) * padding_size from the b_ends array. It works, but only to an extent (sometimes it cuts too much, sometimes it leaves in a bit of the padding).
Is there any better way to teach the synth to process shorter sentences?

[ghost]

Poor performance on short inputs results from bad training data. You can try the alternative model in #538 and see if it gets better.

Problems with long inputs are caused by a failure of the attention mechanism. Solution is to implement a better one and retrain the model. Much easier said than done.

[akrokodile]

Thank you so much! The alternative model does get the short sentences right (also gets rid of weird pauses).
Re the long ones--I've got a working method of handling them (basically considering them in chunks of 20 words and seeing if I can split on commas).