This paper presents a comparative study on sentiment analysis applied to legal texts, specifically legal opinions, and judgments. This paper uses embeddings such as T5, XLM Roberta and BERT, Roberta, and Legal BERT to convert the textual data into pre-trained language model word representation, to make it easier for ML algorithms to determine the sentiment (positive, negative, or neutral) of legal documents. The dataset was converted into embeddings, and various models, including KNN, ANN, SVM, random forest, XG boost, logistic regression, and decision tree, were employed for sentiment analysis. Each model was subjected to hyperparameter tuning using Grid Search CV, resulting in the highest accuracy of 66.66% with SVM on T5 embeddings. This research provides valuable insights into the emotional degrees within legal texts, demonstrating the effectiveness of our sentiment analysis approach.