Oliver Adams

Yes, this is a standard approach used in ASR. Typically there will be a group of non-language symbols that represent different noises and the model learns to transcribe them, then...

Examples of such symbols in a Kaldi recipe I was recently working with: ``` (hesitations) (Unknown word) ```

using hyperopt or skopt.

What are the pros and cons of this as compared to what I'm doing with num_feats at the moment? That is: ``` @property def num_feats(self): """ The number of features...

Okay, let's just stick with the standard property decorator for now.

Using it in Travis testing sounds like a good idea!

This is a good idea! There'd have to be some experimentation to determine thresholds for the automatic checking mode. Ideally this should be done across a diverse array of datasets....

Create a roadmap for XNMT compatibility. Tentative: - [ ] Try using XNMT's serializer.py for handling YAML. - [ ] Do the XNMT tutorial to elicit ideas

Hi Niko, thanks for the kind words. > Does it just take first nth training samples? So even if the corpus is larger, it takes only this portion? [...] >...

@nikopartanen Yeah, you got it. The reason this happens is is because of the way the CTC loss function works. It sums over all the possible ways your input-frame ->...