Automatic speech recognition system for court proceedings in Ethiopia

Abstract Automatic speech recognition (ASR) technologies have revolutionized specialized domains like healthcare and broadcasting, yet their application within legal and judicial proceedings remains unused for under-resourced languages. While foundational ASR research exists for Ethiopian language, a specialized system tailored to the unique linguistic and structural demands of the judiciary has been notably absent. This study addresses this gap by proposing a speaker-independent Tigrigna ASR system specifically designed for court proceedings in Ethiopia. Utilizing a Hidden Markov Model (HMM) framework, we developed a dedicated Tigrigna speech corpus featuring recordings from twenty-four native speakers, partitioned into training (90%) and testing (10%) subsets. The conventional ASR architecture integrates acoustic, language, and dictionary models. Both context-independent and context-dependent phone-level acoustic models were engineered using the CMU Sphinx toolkit, with speech waveforms parameterized into acoustic feature vectors via Mel-Frequency Cepstral Coefficients (MFCC). Furthermore, a phonetic dictionary was developed using a custom Java application, and a trigram language model was constructed via the SRI Language Modeling Toolkit. Experimental results indicate that the best recognition performance is obtained using a phoneme-based, context-dependent model configured with 11 Gaussian mixture components. Furthermore, comparative analysis shows that the context-dependent acoustic model substantially outperforms the context-independent variant, reaching an optimal baseline word accuracy of 73.84%. Accordingly, this study establishes a solid baseline for the digitalization of judicial workflows in Ethiopia.