You probably already tried DPGMM from skleran. If not, you could start with it. From my experience I'd say its not fast at all, but its easy to start with.

There was also an interesting project here https://github.com/amir1981/hdphmm_lib

It was designed for speech and does inference for both observations and states (so it is DP HMM in fact). The idea and the paper are rely interesting, but I did not find time to play with the tool. If you try and it works for you, let us know.

Arseniy thank you for your answer, I did not  play with HDPHMM yet, but I found another implementation I'm satisfied with (Matlab interface and use of multiple cores).

publication -> http://people.csail.mit.edu/jchang7/pubs/publications/chang13_NIPS.pdf

code -> http://people.csail.mit.edu/jchang7/code.php

Usability in unsupervised acoustic modeling was shown in http://lxie.nwpu-aslp.org/papers/2015Interspeech_CHJ.pdf