Ivan Sorokin

These variables comes from the classical forward/backward algorithm. `alphas` and `betas` must be equal with the small measurement errors. For some reason the values looks very small. Please check that...

I can't remember from my practice that these values was so small. Maybe we should add additional check not only for ratio, but also for abs value as well. Fell...

If I remember correctly, U includes "empty" output, very similar to the first element in the scoring matrix when you align two sequences, for example like this https://en.wikipedia.org/wiki/Smith–Waterman_algorithm

Wow, looks cool! Thank you! I will check it in next few days. As far as I remember, there was a research paper about the compact layout. Maybe we should...

Exactly, Section `3.1. Efficient encoder and prediction output combination` is it similar or am I misunderstood?

I did review and I can't understand how it will be used in practice. @maxwellzh could you shed some light on how to prepare `log_probs` and `labels` arrays in practice?...

Thanks for clarification! Now it make sense, you have the additional functions. Could you add these into MR? Without these functions, it's not clear how to use the compact version...

Thank you! I will check it on the weekend.

You shouldn't average over frames. If I remember correctly, theoretically it doesn't make sense. The loss is calculated for the entire utterance. There is no a direct link between the...

Yes, the loss function implemented only for float values. I have to generalize the implementation for other types. Currently, you can convert logits to float explicitly.