normantud / dscorrector

Very simplistic approach to create di- and trigrams of uttered words to select a statistically probable version of a DeepSpeech Utterance

DSCorrector

Allows to create a graph of di- and trigrams of utterances from text files. E.g.

This house is yellow.

This creates:

2:

"this" -> "house" = 1

"house" -> "is" = 1

"is" -> yellow = 1

3:

"this" -> "house" -> "is" = 1

"house" -> "is" -> "yellow" = 1

When you add other sentences, the probabilites will change according to how many utterances of this were recorded.

How to train

Put all of your txt-files into the folder folder. Run perl create_graph.pl.

How to use

You need to be able to pass the JSON-output of DeepSpeech into STDOUT and pipe it into the correct.pl after creating a graph.

Example:

cat dat.json | perl correct.pl graph.db

Caveats

Not every word is checked. This is about to be changed soon hopefully. Right now, only the sentence alternatives as a whole are checked.

The default graph.db consists of Mars by Fritz Zorn and the german Bible.