A Multi-criteria Text Selection Approach for Building a Speech Corpus

摘要

Speech corpus is an important and primary requirement for several speech tasks. Building a speech corpora is a lengthy, time consuming and expensive process, it typically involves collection of a large set of textual utterances and then selective distribution of these text utterances among a set of speakers, called speaker sheets. These speaker sheets are articulated by speakers to generate the speech corpora. Depending on the task at hand the speech corpora needs to satisfy certain criteria; For example, a phonetically balanced speech corpora is essential for building an automatic speech recognition (ASR) engine, while for a text dependent speaker recognition engine there is a need for several spoken repetition of the same text by several speakers. In this paper, we formulate a method that enables creation of speaker sheets from a predetermined set of text utterances such that the speech corpora satisfies the desired requirement.