Why Kaldi with Tensorflow?

[Jon Nichols]

Hi, I’m still new to machine learning, and recently someone pointed me to Tensorflow. I’m sure almost everyone reading this knows it’s Google’s API for deep neural networks in general, while Kaldi is made specifically for speech. So, I tried to search for “Kaldi vs Tensorflow” to try to figure out the differences between the two and instead of finding results for the differences the first few results were all projects combining the 2 together, Kaldi using Tensorflow made models. I even saw a recipe was being made from the main Kaldi branch, https://github.com/kaldi-asr/kaldi/pull/1256, but devolvement has since stopped.

 

But I still wasn’t able to find the simple answer of why use Tensorflow with Kaldi? I have a couple guesses from what I’ve read but nothing definite.  My main guesses were its because Tensorflow looks easier to use and to see if its models were  better/faster. But also, wasn’t sure if it can do things to the models that kaldi currently can’t.

thanks again for all the help.

[Daniel Povey]

Tensorflow is a lower-level, more general-purpose tool.

Unless you're the kind of person or team that knows deeply about speech recognition and has months or years to work on it, you wouldn't be able to use Tensorflor for ASR.

People have tried using TensorFlow with Kaldi and generally speaking they have not gotten as good results as Kaldi's own inbuilt neural net tools.  Which is not to say it's not possible, but it's a lot of work and it doesn't make your life simpler than using Kaldi's own tools.

Dan

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.