2
$\begingroup$

I am wondering, what can we ask about speech recognition on this site.

This question indicates that machine learning is very much included in the site. So how about the following questions:

  1. How do I train an acoustic HMM model
  2. Why does my ngram language model get so big
  3. What tool can I use to do Speech Recognition
  4. What are the advantages of using Finate State Machines for decoding/recognition
  5. How do I do discriminative training
  6. HTK (Hidden Markov Model toolkit) gives an error when I try to align my transcriptions. What is wrong?
  7. How do I adapt my Language model to law text

Do we want this kind of question on the site? If complicated, please indicate which questions you think are ok and which not.

| |
$\endgroup$
4
$\begingroup$

I think you should give him a chance to ask them and then decide. It is better to close some questions than to have everyone asking on meta if they can ask a particular question. IMO 1, 2, 4, 5 and 7 are perfectly ok, 6 may be off-topic, for 3 the answer is R ;-) (ok, is also off-topic).

| |
$\endgroup$
  • $\begingroup$ I like this answer :). If 6 is off-topic, is it then also off topic to ask help with getting weka to run k-means? $\endgroup$ – Peter Smit Jul 23 '10 at 15:07
  • $\begingroup$ Ok, I had an impression that it is more specific to the voice recognition area. $\endgroup$ – user88 Jul 23 '10 at 16:11
  • $\begingroup$ Yeah, the aligning of transcriptions is maybe not the thing to ask, but for example asking about the splitting of states and retraining them is maybe more appropriate. $\endgroup$ – Peter Smit Jul 23 '10 at 16:18
3
$\begingroup$

To echo some of the sentiments above:

I think if you separate the more implementation based questions from the more abstract statistical questions you will find cross-validated more useful.

I come from a related signal processing background to speech but find the dsp.stackexchange site far to focused on implementation and so for more methodological questions around the HMM theory, signal processing approach etc cross-validated is far more useful. For questions (such 6. above) on particularly software packages or other issues with implementation such as bugs, dsp.stackexchange or even stackoverflow might be more useful.

| |
$\endgroup$
1
$\begingroup$

I think at least some of this is out of scope. Maybe Q1 is ok, but I don't know if acoustic HMMs are any different from HMMs in other areas. If the question is about training HMMs generally, it would be fine. But if acoustic HMMs are very different from other HMMs, then maybe it would be to application-specific for this site.

In general, questions that require a lot of application-specific knowledge should be out-of-scope in my view.

| |
$\endgroup$
1
$\begingroup$

I agree with Rob. You could always make the question application agnostic by abstracting away some of the application details. Abstracting away application details can make the question difficult to understand but perhaps may also make it 'answerable' by the community.

| |
$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .