District names speech corpus for Pakistani Languages

Abstract

This paper presents a speech corpus that is developed for Urdu automatic speech recognition (ASR) system. The corpus comprises of single word utterances fixed vocabulary consisting of district names of Pakistan. The data is recorded over a telephone channel from all over Pakistan to cover six major accents; Punjabi, Urdu, Saraiki, Pashto, Sindhi, and Balochi. The data was collected in challenging acoustic environments; the major issues were silence, background noise and alternate pronunciations, which can affect the performance of the system. In order to address these issues, comprehensive data verification and cleaning guidelines are presented. The proposed process serves as a data preprocessing step for the development of ASR, which is successfully integrated in an Urdu dialog system to provide weather information of Pakistan.