Dynamic out-of-vocabulary retrieval for automatic speech recognition

Speaker: Amélie Greiner

Date: September 14, 2017


To perform a transcription, a speech recognition system relies on a vocabulary that contains all the words that can be transcribed. In practice, it is impossible to include all the existing words in this vocabulary, which therefore contains only the most common words of the language. Out-of-vocabulary words can never be transcribed. This problem concerns in particular proper names because new proper names appear constantly, and often contains essential information to understand the audio document. During my internship, I focused on the problem of out-of-vocabulary proper names. To address this problem, one possible approach is to build a list of out-of-vocabulary proper names that are relevant to the audio document that we want to transcribe, using an external source of information, in order to add these proper names to the vocabulary. I studied different methods to constitute such a list from the online encyclopedia Wikipedia. To do so, I studied different ways of representing textual documents in a space and exploiting these representations.