Adaptive hosted text to speech processing

Abstract

Techniques are provided performing text-to-speech translation in situations in which the input texts may contain unanticipated content. According to one aspect of the invention, text-to-speech services are provided by splitting a text into segments that include anticipated-content segments and unanticipated-content segments. Speech for the anticipated-content segments is generated based on pre-recorded sound recordings that correspond to the anticipated-content segments. Speech for the unanticipated-content segments is generated using speech synthesis. Usage statistics are recorded. The usage statistics identify which segments are contained in texts that are translated using the text-to-speech services. In one embodiment, the usage statistics indicate frequency of use of unanticipated-content segments and, based on the usage statistics, a set of unanticipated-content segments for which to make recordings is selected. In another embodiment, the usage statistics indicate frequency of use of anticipated-content segments, and a set of anticipated-content segments is selected based on the usage statistics. The recordings associated with the selected anticipated-content segments are then removed.