Big Data & Speech Recognition
The field of Speech Recognition offers countless possibilities for big data analysis. Speech recognition, which deals with the translation of spoken language in written text, is being used for a wide range of applications — from automatic telephone answering services to medical dictation systems. Some of the areas dealt with us in the development process are:
Speech recognition context
Context is a subject. The more a context can be narrowed down, the better the search results and the lower the error rate. While radiologists have a limited vocabulary for their diagnoses, the context for a legal notary is far more extensive. For the editor of a newspaper, who deals with a diversity of themes, a single context is not sufficient. Big data helps not only in defining appropriate areas but also in assigning texts correctly to these areas.
In spoken language punctuation marks are usually not dictated. The software must fill these in automatically, for which there are two methods. If the software guesses where the punctuation marks need to be added, it will miss only a few, but most will be set in the wrong positions. If the software adds fewer punctuation marks, the error rate will be low, but the rate of missing punctuation marks would be high. With the help of big data, the so-called z-value can calculate the optimal compromise between these two strategies.
Speech recognition software generates text from predefined dictionaries. The larger the lexicon, the greater the search field, but also a large number of possible matches for different words. In order to keep the search field small, it helps to create subcategories for each word. This method can be applied to liaisons in French, the 16 different case endings in Finnish, or the long compound nouns in English. Big data provides the strategies for optimal Word categorization.