tutorial yes/no version of the model? #2020
Replies: 3 comments 6 replies
-
|
This is a great request... there actually isn't one (yet!). There are similar, existing tutorials here: https://github.com/coqui-ai/STT/tree/main/notebooks when we've got a yes/no tutorial I'll post here:) |
Beta Was this translation helpful? Give feedback.
-
|
Hi @subzero882 -- here's a quick tutorial on training STT models for yes/no... let me know what you think! https://colab.research.google.com/github/JRMeyer/STT/blob/jupyter/notebooks/yesno_tutorial.ipynb |
Beta Was this translation helpful? Give feedback.
-
|
@JRMeyer First, I want to say how amazed i am at the speed of your response! Second, I didn't realize that the notebooks folder had tutorials in it, i was just going by the ReadTheDoc page. So that is very helpful. Third, after looking at the notebook I think the tutorial that helps the most is this one, https://colab.research.google.com/github/coqui-ai/STT/blob/main/notebooks/easy_transfer_learning.ipynb but going over the Yes/No one was helpful as well. Finally, I am using the docker image for my training which i think i have all setup now. Based on the tutorials you posted (not well versed in colab though I think I should learn it), and after looking over the tutorials above I think I have a direction way better than I did before so thank you very much for that. I think it would also help to see some tutorials similar to the ones you have cover doing this with the docker image. With what I am trying to do (command words), does that sound right to use the transfer learning for a checkpoint in the huge-vocab, and then just have the command words as data for that? Also, since i would be using transfer learning, does it sound reasonable to only have each command word recorded only a couple of time for a few people, or do you think I'll need hours of data for that? Thank you very much!!!! |
Beta Was this translation helpful? Give feedback.
-
the best approach for you and your use case will depend on a few things, but the fastest and easiest one I would first try out is to just take the You should come chat with us in our community STT channel:) that's where you will get even faster feedback! |
Beta Was this translation helpful? Give feedback.
-
|
@JRMeyer Thank you for link to the chat. So previously when we were using deepspeech just to test this all out, we used hotwords but that didn't help completely as we are using some uncommon words and word pairs for commands that probably aren't in the current training data. Example words, 2D map, 3D map, magnify, Heads-up display, etc. I will however, try hotwords with the new model as well and see what results I get before I re-train. |
Beta Was this translation helpful? Give feedback.
-
|
@subzero882 the model was trained without digits like [0,1,2... ], but rather words [zero, one, two... ], so, you should try spelling out the hotwords like "two d map" and "three d map", and there's no dash, so "heads up display" |
Beta Was this translation helpful? Give feedback.
-
|
Sorry I wasn't clear there, yes I do have the numbers spelled out as well as no dash in the hotwords. |
Beta Was this translation helpful? Give feedback.
-
|
If hotwords isn't sufficient, and you have a command-and-control application (which sounds like you might have), then training a custom language model will be your next step to improve performance. |
Beta Was this translation helpful? Give feedback.
-
|
Ok ya that's what i was thinking, so with that in mind do you think Fine Tuning would be sufficient? |
Beta Was this translation helpful? Give feedback.
-
Is there a tutorial on how to retrain coqui into the yes/no only model?
I'm building some software looking to use command words, and i'm having some issues in understanding how to take STT as is and have it reduce to a subset of words. like how the yes/no model would work
Beta Was this translation helpful? Give feedback.
All reactions