search

julius-speech / julius

Open-Source Large Vocabulary Continuous Speech Recognition Engine
BSD 3-Clause "New" or "Revised" License
1.84k stars 299 forks source link

Is the creation of DNN models with HTK + inference with julius still a good choice #175

Open welliX opened 3 years ago

welliX commented 3 years ago

Hi, I am wondering, if HTK-DNN training combined with Julius inference is still a good choice (with respect to performance, speed, usability, ...). There is almost nothing (scripts, comparing WA results, etc) to find in the web - is this approach inactive/dead since more than 5 years and not worth thinking about it?

I would like to give it a try and am thus searching for a kind of (step-by_step) recipe, tutorial, script, or similar, that describes how the Julius DNN acoustic models are created (prpbably from/via HTK). many thanks for any help! kr Andi

palles77 commented 3 years ago

It's not dead att all. I am planning to publish the recipe at some point this year.

welliX commented 3 years ago

Hej palles77, thanks for coming back, good to know there is somebody ;-) Do you know maybe also about any WA/speed/delay results compared to 'pure' DNN approaches? I didn't find any publication around that topic. And I hope you can publish your recipe soon (maybe even already a quick pre-release...) thanks and br, Andi