Python ASR Server - Websocket Changing Acuracy

Hi, thanks for the great project.

I believe I found some odd behavior on the Python ASR Server - Websocket and it is 100% reproducible:

Trying to detect the word "sim", with the following phrase list "phrase_list" : ["assim", "sim","não","ok", "sim obrigado", "não obrigado", "quero","[unk]"]

The client is the microphone example in python. the behavior is similar, changing sample rate, changing microphone brand, changing chunk size in the cliente size. Also, the behavior ramains the same playing in the server side with mfcc.conf and model.conf.

If I speak "sim" (pt-br equivalent to yes) several times, it does not detect in the PartialResult and almost all the time it does not detect the Result either. Here is the log of several "sim" spoken:

{ "partial": "" } { "partial": ""}

[sniped several lines of empty partial to save space]

But, if I speak "não" (thas pt-br equivalent to no), it does detect PartialResult and, for my surprise, it now detects the "sim" easily and produces also Result for "sim", its acuracy improves a lot:

{ "partial" : "não"} { "partial" : "não"} { "partial" : ""}

[snip]

Now I am able to detect "sim" easily:

{ "partial" : ""} { "partial" : "sim" }
[snip]

{ "partial" : "sim" }

From now on, for the same connction, all works ok. For a new connection, the behavior is the same.

If you need more logs, please let me know.

Thanks you so much.

Eduardo