Improve Voice Recognition Model

[pedrooaugusto]

Description

The first step in converting speech into code, is to convert speech into text, for this we use Azure Speech to Text, which has a pretty good accuracy rate. As you can see on this example all you have to do is choose a language say something and Azure STT will do its best to transform what you said in text.

Azure STT gives you an option where you can train and use a custom model to enhance the accuracy for domain-specific language instead of using the default model. If you are using Azure STT to create an application with speech recognition capabilities for a Nuclear Plant you could use the default speech recognition model and it would work fine, however if you train a custom model on the jargon and terminology used by nuclear engineers it would work tremendously well.

The aim of this issue is to train a custom Azure Speech to Text model capable of better recognizing the phrases we used as commands. The phrases used as commands are described by (and can be extracted from) these automata.

A possible issue with training such a model is that the phrases used for training are really just patterns, in the sense that: new variable (any) equals (any) is a valid command and (any) could be almost anything. So if you train a model for recognizing new variable dog equals number 42 it may not recognize new variable cat equals call function add with arguments number 40 and number 2, but it should since they follow the same pattern. Maybe the solution is to train your model to recognize phrases that fall in the pattern described by the automata... (not my area...)