DNN-Based Supervised Spontaneous Court Hearing Transcription for Amharic

Research in the area of Automatic Speech Recognition (ASR) for Ethiopian languages, especially for Amharic has been conducted since 2001. However, the ASR systems have not been used for real-world applications such as court hearing transcription. In this paper, we present our endeavour towards the development of a DNN-based supervised spontaneous court hearing transcription system for Amharic. Speech and text corpora that are required for the development of the ASR system are collected from the Ethiopian Federal Courts. The text data has been cleaned through a series of pre-processing tasks while the speech data has been manually segmented and aligned with the corresponding transcription. A Deep Neural Network (DNN) based ASR system has been developed using 90% of the pre-processed data and the remaining 10% has been used as an evaluation set. The back-end of the system interface has been developed using Laravel PHP while the front-end has been developed using Java script, HTML5, and CSS. The database is developed with MySql and for the API python programming language has been used. The performance of the court transcription system has been evaluated in terms of word error rate (WER) of the ASR system on the evaluation set. When the ASR system is manually evaluated on 6000 segments (from the evaluation set) by a transcriber who listened to each of the audio segments, it achieved a WER of 7.6%. Based on the result, we can conclude that this transcription system can be used by the Federal courts of Ethiopia with minor editing of the transcription output.