What algorithm would you recommend for minimizing errors in speech recognition?

Question

Speech recognition

There is original text and text recognized in real time using the bing speech api. In general, both texts should match, but there may be deviations (that is, a person speaks from the original text, but can start from anywhere or interrupt for a while and start talking off topic, and then continue reading through the text). How can I get rid of speech recognition errors? The api itself provides error handling in words, that is, such words exist, but a person could pronounce a completely different word, and not the one that the framework recognized, how can one understand, based on neighboring recognized words, that this is just a recognition error, and not a person began to read from another place? Need to use some kind of fuzzy search algorithm for words in sentences?

Answer 1

Try to connect the Stumper.ru API. It gives
very accurate results for matching two strings.

Answer 2

see NLTK
bigrams, trigrams and statistics
about "I read from here, but now not from here" it will be more complicated, but in general the same