snowboy | Future versions with model training module

I am building a personal assistant on my 3B+. What I am trying to do now is use Snowboy to detect my hotword (works flawlessly), then after the hotword is detected, use SpeechRecognizer to receive a voice command. Hotword detection works fine, the error happens when sr.Microphone() is called.

Example Code:

I am creating a personal assistant in Python. I am using Snowboy to record audio, and it works very well. Snowboy has a saveMessage() method that creates and writes a wav file to the disk. This wav file is later read and used as an AudioFile object by Speech_Recognition. I find it very inefficient that the program has to write and read wav files to the disk. I would much rather have the wav file be passed around as an object without EVER saving it to the disk.

Here is the snowboy saveMessage() module that I would like to reweite.

I have python script which is using google STT engine and then it loops through dictionary to check where is the biggest match (ratio) using difflib SequenceMatcher

UPDATE (better explanation of what script needs to do):

1. variable izgovoreno contains string (from google STT engine,actually microphone) - let's say it's value is "lights"
2. there is nested dictionary where each hotword (like "lights on", "blinds shut" etc.) contains correspoding command. There is "hotword" and there is "komanda". In this example hotword value is lets say "lights on" and corresponding komanda is "execute some code to turn on the lights"
3. komanda will only be executed if there is some percentage(at least 50% or 0.5 in decimal) match between variable izgovoreno and hotword. So, there is a for loop which goes thru dictionary and using difflib.SequenceMatcher compares variable izgovoreno with all hotwords from dictionary, and then it writes result (percentage) to key "razlika_izgovoreno_hotword" for each corresponding hotword.
4. Now we have dictionary with this data (example):

...