Sign language serves as a primary means of communication for the deaf and hard-of-hearing community, utilizing hand gestures, facial expressions, and body movements to convey meaning. Translating this multi-modal system into text or speech presents substantial challenges due to its concurrent, spatially rich signals. We propose a transformer-based architecture that converts sign language videos into gloss sequences and natural language text, leveraging pre-extracted visual features from a convolutional backbone alongside emotional and gestural cues obtained via a perception framework. These modalities are fused through a cross-attention mechanism, and hyperparameters are tuned with Optuna to optimize BLEU (Bilingual Evaluation Understudy), WER (Word Error Rate), CHRF (Character n-gram F-score), and ROUGE-L (Recall-Oriented Understudy for Gisting Evaluation). Our key contributions include a multi-modal transformer with modality-specific encoders, a cross-attention fusion strategy, and a systematic optimization pipeline. Experiments on the RWTH-PHOENIX-Weather-2014T dataset demonstrate the effectiveness of our approach, yielding competitive translation performance on this benchmark.