Remove defunct demo

[simoncorstonoliver]

I have spent ~4 hours working through the demo trying to get it working. It has problems so severe that I don’t think it’s worth trying to salvage it. This example might make a good blog entry for people who are thoroughly versed in Kaldi, an Open Source automatic speech recognition (ASR) project, and who are interested in an MXNet wrapper for it but does not stand alone as an introductory example to speech recognition suitable for inclusion in the MXNet repo. Alternatively, an ASR expert could perhaps rework this example and resubmit it if it illustrated general MXNet concepts.

Specific criticisms that lead me to propose removing the demo altogether:

• Depends on Kaldi which is a bear to install and compile (but I got there in the end).
• The example includes a Python wrapper around Kaldi which does not compile. Editing the makefile to remove the link dependency on ../thread/kaldi-thread.a allowed it to compile but I have no idea if the result will work as intended.
• The discussion of preparing data gives no hint of where you would download compatible data or transcripts. It tells you to run Kaldi scripts that barf and ask where your Fisher transcripts are.
• The example has vague instructions to go read the Kaldi docs but doesn't even give a title or topic that you should go looking for.
• The tutorial is full of acronyms that appear to have ASR-specific meanings that are never explained and that a user of MXNet would not reasonably be expected to know e.g. AMI does not mean "Amazon Machine Image". Similarly TIMIT is a speech corpus.

[piiswrong]

@Soonhwan-Kwon

[Soonhwan-Kwon]

@piiswrong @simoncorstonoliver
Our team had looked the example for 2 weeks about an year ago, and we skipped the few steps to make example works.
As mentioned in pull request there are some problems.
First kaldi preprocessing on dataset took days.
And second we've made many detours to fill the gap of vague instructions.

But we referenced the network architecture of it and also method for reading configuration,
and it was quite helpful.

[simoncorstonoliver]

@Soonhwan-Kwon @piiswrong
Glad to hear the tutorial was helpful for your project but the question really is whether it belongs in the repo. We're trying to improve the quality of examples and would like to move to a world where the examples can be automatically validated for each release. If preprocessing the data using the external ASR tool Kaldi takes days then I propose that the tutorial should not be included in the repo.

if someone with expertise in ASR were to prepare some sample data that could be used with the reference network architecture then perhaps a new example tutorial could be created in the future, but for now let's remove it so it doesn't cause confusion on the web site for people who are new to MXNet.

[piiswrong]

Doesn't the speech_recognition example work? @simoncorstonoliver

[simoncorstonoliver]

I was not able to verify that it works. @Soonhwan-Kwon says that people with more expertise than me were eventually able to get it working after filling in the missing/vague information and training for several days so I would say no, as it stands it doesn't work.

[Soonhwan-Kwon]

@piiswrong @simoncorstonoliver I'm afraid that you've confused the author who wrote the speech-demo example(maybe pluskid?). The example I've made is speech_recognition.