Automatic identification of judges' speech in labor court hearings using machine learning: contributions to innovation consulting in the Judiciary and to the development of teaching cases

Abstract

The Brazilian Labor Justice faces significant challenges related to procedural delays, particularly in the analysis of hearing recordings that support the preparation of draft decisions and opinions at the appellate level. In this context, the use of Artificial Intelligence (AI) techniques emerges as an alternative to support document analysis activities and reduce the time spent by court staff in identifying relevant information. The Regional Labor Court of the 8th Region (TRT8) already uses the Degrava tool for automatic transcription of hearings; however, at the time of this research, it did not provide mechanisms capable of automatically identifying the judge as the speaker. Accordingly, this study proposes and evaluates an audio classification model for identifying judges’ speech in labor court hearings. The dataset was built from public recordings made available by TRT8 on YouTube, specifically from the playlist “Sessões da 1ª Turma”. The audio files were extracted, segmented, normalized, and processed using an automated methodology developed in this research, including acoustic feature extraction based on Mel-Frequency Cepstral Coefficients (MFCCs). The final dataset comprised five classes, each corresponding to a different judge. Several Machine Learning algorithms were evaluated, including Support Vector Machines (SVM), Logistic Regression, Random Forest, AdaBoost, and Multi-Layer Perceptron (MLP) neural networks. The results indicated high performance across all models, with accuracy above 94%, highlighting the MLP, which achieved accuracy close to 99% and balanced metrics among classes. The findings demonstrate that traditional models, combined with consistent preprocessing, are effective in noisy acoustic environments typical of labor court hearings. It is concluded that the automatic identification of judges’ speech is technically feasible, compatible with the computational infrastructure available at TRT8, and suitable for integration into the Degrava ecosystem, providing gains in efficiency, legal certainty, and potential expansion to the identification of other interlocutors. Beyond its technical contribution, the study presents practical utility for business and organizational consultancies focused on innovation in the public sector, especially in initiatives for digital transformation, workflow automation, and improvement of institutional efficiency. The findings can also support the development of teaching cases on the application of artificial intelligence in the legal context, allowing its use by teachers and facilitators in problem-based learning activities and case analysis.