Automatic Detection and Correction of Errors in Video Tutorial Transcripts

Speech-to-text technologies continues to become more accessible and more accurate. But while automated personal assistants and transcription software is improving, it still falters in many cases that involve domain-specific terminology, like medicine and technology. If spoken clearly and concisely, some terms can be correctly transcribed, but there are many confounding variables like accents that can make it unlikely to transcribe the correct word or phrase. To fix these errors in domain-specific terminology, I suggest a post-processing step to find and correct these errors. To find the mistakes, I trained a language model is trained on relevant data—in this test case, Stack Overflow posts that were tagged with Java-related keywords. The model is used to find words in automated transcripts that are likely mis-transcribed. Using n-gram and a long short-term memory neural network trained on 80% of the manual audio transcriptions, suggestions for replacements of the mis-transcribed words are offered. The accuracy of these suggestions are found by comparing corrections to the test case manual transcription corresponding with the automated transcription. These results are a mixed bag that reveal an issue with the language model but that the neural network can successfully suggest word corrections for technical words with almost a 70% accuracy.