At Addis Ababa University, I research Automatic Speech Recognition for low-resource Ethiopian languages, including Tigrinya, Somali, Wolaytta, and Afar. I adapt pretrained multilingual speech models through fine-tuning and transfer learning.
My ASR work explores tokenization strategies, including Byte Pair Encoding, character-based and other subword approaches, and custom tokenizers such as SentencePiece. I prepare speech and text datasets and evaluate models using Word Error Rate and Character Error Rate.
At Information Network Security Administration (INSA), I contributed to a machine learning-based risk management system for the Ministry of Revenue and to an NLP system covering tasks such as machine translation and question answering. I also researched civilian fatality prediction in conflict zones in Ethiopia and implemented a machine learning algorithm for predicting civilian deaths.
At INSA, I developed a demo application that generates Amharic videos from English videos using speech-to-text and text-to-speech. My research interests include multilingual speech technologies, adapting pretrained speech models, and tokenization for low-resource ASR.
