This study presents a cascaded Bengali-to-English speech-to-speech translation system, addressing the challenges of low-resource languages through a modular architecture. The pipeline integrates three modular components: Speech-to-Text (STT), Neural Machine Translation (NMT), and Text-to-Speech (TTS) modules. Leveraging datasets like SLR37, SLR53, and LJSpeech, the system employs various algorithms at each stage, including transformer-based ASR, LSTM-driven translation, and FastSpeech2 for high-quality speech synthesis. The proposed system demonstrates competitive performance, achieving an 85.49% STT validation accuracy, 82.89% NMT validation accuracy, and a Mean Opinion Score (MOS) of 3.9 for TTS, with the final pipeline MOS score at 3.73. This work highlights the potential of cascaded frameworks for low-resource linguistic environments and provides insights into potential avenues for further refinement and scalability to other language pairs.