The problem you state seems very interesting. If I understand your statement correctly, I'd guess whether it is possible to do at all even in theory with a decent accuracy... For sure its challenging.

There is numerous amount of research on language identification, where the classifier (neural net, vector representations, etc) is learned in a supervised way (given training utterances with known languages to predict language)

However, when you want to cluster utterances in unsupervised way, the problem starts looking more like speaker clustering, but much harder, cause using the acoustic observations you will likely come up with channel and speaker clustering.

I imagine transcribing the data with pseudo-phonetic units and making clustering on distributions of their sequences, but I should check if smth like that was done.

Quick question: In your problem, is the number of languages fixed at least?

I am really interested in any answers on this topic, too, and in all your findings on this.

Arseniy thanks for your interest. I'm aware of research in Language Recognition and Identification task. I'm currently investigating language clustering in i-vector space but I'm interested in any other research on that topic and especially in possible applications. I found this interesting but I'm not sure that there are any real world application of solution to such problem. In article I mentioned there is not much about that. 

In my research (similary to cited article) I cluster the data without knowledge about classes, but for evaluation of usability of my clustering I use labeled data and assign labels to clusters based on that.

I managed to do some experiments on that topic with NIST i-vector Language Recognition 2015 data and my paper is currently under review. My paper was already rejected once and one objection was it poses a unrealistic problem.

They are wrong, the problem is beautiful.

I'll try to spend some time on literature review. Something may be in related area, but what you could generalize to clustering task.

As for applications, we have a lot of unstructured audio data everywhere. Clustering languages (if working) can help in finding garbage, code switching, mixed dialects. There are also bunch of low-resource languages, for which we cannot train a LID classifier. Some even do not have written forms. It may be that we just did not start thinking about these problems yet

If the reviewers are not convinced, there is always theoretical aspects and practical research curiosity :)

Hi Stanisław. Sorry for not replying for a long while

I had a look on a couple of articles. You may be already familiar with these. Nevertheless, the initial intuition was correct - your problem's practical application can be related to code switching and dialect detection. Another intuition is that pure i-vector clustering is not likely to be enough as we should more focus on detecting phone-like units and clustering based on their distribution. Well, if works, i-vectors are also interesting, of course.

Have a look at these, for example:

1) Hierarchical Language Identification based on Automatic Language Clustering

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.437.1164&rep=rep1&type=pdf

2) LANGUAGE DIARIZATION FOR CODE-SWITCH CONVERSATIONAL SPEECH

http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6639083&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D6639083

I hope some ideas could be useful. Wish you luck with this challegning task.

To make your work motivation more strong, maybe it would be worth clustering some close languages or even dialects that can be frequently found in one audio stream: like French+Arabic/Creole, Russian+Kazakh/Tajik(Or any country where people tend to mix native language with non-native one learned in the past). The data may be difficult to find though...

Yes, you are right, my problem practical application could be related to code-switching. The second article looks interesting to me. Problem is similar to one I stated, except authors know number of languages (2) and the method is supervised. But on the other hand, languages are mixed (this is problematic for i-vector extraction) and I was considering only mono-lingual utterances (samples). I will investigate on topic of code switching in speech. Thank you.