Hence, automatically disambiguating these cases is a challenging task because of different ambiguity issues existing in both un-normal texts and other forms of texts. In the running example of Figure 1, the first occurrence of ‘year ’ is ambiguous with the digit ‘five, ’ or the person name ‘Sáu ’ is ambiguous with the number ‘six ’ if their surrounding contexts are not fully considered. General speaking, detecting such entities requires quite a bit of linguistic sophistication and native speaker intuition (Schutze ( 1997)). Such proper nouns and formatting are called entities in this paper.
Three main types of proper nouns which are person names, organization names, and location names and three typical context-specific formattings such as dates, time, and numbers are considered in this research. We, hence, concentrate on the latter task which aims to automatically transcribe proper nouns and typical context-specific formatting. In this study, we assume that the former task is solved by using a simple technique based on the information of long silence between speeches, to identify sentence boundary. Punctuator detection which mainly focuses on periods (sentence boundaries).Īutomatically recognize and convert the spoken form of texts into their written expressions adhering to a single canonical rule. Normalizing transcribed texts, therefore, plays an important role in STT systems. Restoring the norm-texts greatly improves the readability of transcripts and increases the effectiveness of subsequent processing, like machine translation, summarization, question answering, sentiment analysis, syntactic parsing, and information extraction, etc.
This type of automatic speech recognition systems generally produces un-normalized text (as indicated in Figure 5 Figure 1) which is difficult to read for humans and degrades the performance of many downstream machine processing tasks. As the name would indicate, Speech-to-Text is a system that gets speech input and instantly generates texts as it is recognized from streaming audio or as the user is speaking.