1
$\begingroup$

I'm trying to understand and implement CTC-loss for speech recognition (here on SO). I'll like to have more information about the use cases of this technique.

From what i understood, it is more dedicated to understand sentences (e.g. "Please close the door.") than word commands (e.g. "close") as it's efficient to learn the sentence's structure (spaces, end of sentence) and fitted to recognize unknown words from the training set.

But considering a "some commands versus the rest" problem, could ctc-loss technique be more efficient than a classical deep learning classifier?

Example:

  • Classical model has 5 classes, 4 are commands ("open", "close", "enter", "exit) and the fifth is a garbage which takes whatever that is not a command.

  • CTC-loss model "translate" letter by letter every words and then we check if it's a command.

What is the best approach?

$\endgroup$
1
$\begingroup$

For short commands you can use more simple model, not necessary end-to-end CTC loss. There is extensive research about that. You need to google for "Keyword spotting". For references you can check

https://arxiv.org/abs/1904.03814

For something simple check

https://github.com/ARM-software/ML-KWS-for-MCU

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.