search

UFAL-DSG / cloud-asr

Cloud-based Automatic Speech Recognition (ASR) platform and a public ASR webservice.
http://www.cloudasr.com
Apache License 2.0
65 stars 25 forks source link

Is this project still alive? #13

Open schadom opened 6 years ago

schadom commented 6 years ago

cloudasr.com seems to be dead. Thanks

ondrejklejch commented 6 years ago

Hi,

I am sorry, the service www.cloudasr.com is down as we don't have any servers to host it. The code and models should work and we might release new models next year. Are you interested in the service or do you want to run it on your servers?

Ondrej

schadom commented 6 years ago

Hi Ondrej,

thank you for the quick response. For now I just want to play around a bit in combination with my EchoPi and a small GPU cluster in the back :-) I just noticed that there are docker images available and scaling is fairly simple, which is really cool! I also like the idea of a public and freely available ASR webservice for non-commercial and educational use.

Also I've just found the original website and documentation via https://web.archive.org/web/20160804174158/https://www.cloudasr.com/ . Would it make sense to host the website and documentation again? Maybe it could be hosted via https://pages.github.com/

Thanks

oplatek commented 6 years ago

@choko @schadom We may be able to host it with a demo on our company website.

Can you tell me what is your use-case for it?

schadom commented 6 years ago

Hi @oplatek, currently no concrete use-case besides playing around with different ASR technologies.

dpny518 commented 6 years ago

@oplatek i read your thesis, and this was able to get 1000 parallel request? Do you know the most recent updated version, I want to try and put it up and see if I can use it for real time production use

oplatek commented 6 years ago

@bamba518 it is actually @choko thesis, so let him respond.

Note: I used the code for production before.

ondrejklejch commented 6 years ago

@bamba518 In the thesis we used 1000 fake ASR engines, because we didn't have big enough cluster to run 1000 real ASR engines. In theory, the only limit for the number of parallel requests is a size of your cluster, because one worker can handle only one request.

dpny518 commented 6 years ago

@choko does the size of cluster mean that each core can only handle 1 worker? so 1000 parallel request requires 1000 cores?

ondrejklejch commented 6 years ago

Yes, I think this is a good rough estimate - it all depends on your ASR model. Do you have a good ASR model? Is it for closed or open domain?

dpny518 commented 6 years ago

I was just going to us this model http://goofy.zamia.org/zamia-speech/asr-models/kaldi-generic-en-tdnn_f-r20180901.tar.xz

My goal is to handle thousands of parallel connections using http streaming, do you have any recommendations ?

ondrejklejch commented 6 years ago

I am not familiar with this model, but TDNN-F models should be a good choice. Make sure to rebuild the docker images, because nnet3 models are not supported in the docker images available from the docker hub.