1
$\begingroup$

End-to-end deep learning systems for automatic speech recognition (ASR) have been around for a while now since Deep Speech (2014), but I noticed that DNN-HMM based methods are still performing well and making it to the charts like here.

Does that mean it is still not settled which system is better? Or do they win based on conditions? Who is better when you just have speech training data in the order of hundreds of hours and not tens of thousands? Which system is better in real life and not on super clean data?

$\endgroup$

1 Answer 1

1
$\begingroup$

Does that mean it is still not settled which system is better?

It is more or less clear which system is better in which condition and what are the downsides and advantages.

Or do they win based on conditions?

DNN-HMM is good for 1000-10000 hours of data and when you do not have too many GPUs (just 2-4 cards)

DNN-E2E is really good for huge amounts (10000 hours - 100000 hours) of data and if you really have extreme computing power (100 of GPUs/TPUs).

So unless you have resources like Google you still go with DNN-HMM.

Who is better when you just have speech training data in the order of hundreds of hours and not tens of thousands? Which system is better in real life and not on super clean data?

For hundred hours it is more effective to collect more data than to choose the algorithm. It really helps to get 1000-2000 hours.

$\endgroup$
2
  • $\begingroup$ I know that it's typical that e2e would need more data, but I can't find research papers confirming that for ASR. In fact, I found a paper suggesting otherwise arxiv.org/abs/1707.00722v2. What point am I missing? $\endgroup$ Mar 17, 2019 at 16:35
  • $\begingroup$ This is a common knowledge $\endgroup$ Apr 13, 2019 at 20:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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