This study investigates the challenge of rhythm recognition in piano compositions, aiming to accurately identify and compare rhythmic patterns within musical pieces, which is crucial for applications in music analysis, transcription, and education. This challenge arises from the complexity of musical rhythm, involving intricate temporal structures and subtle nuances in timing and note sequences. We use deep learning models to analyze rhythmic patterns and address the complexities of sound processing.  The focus is on Long Short-Term Memory (LSTM) networks and Transformer models, examining their effectiveness in rhythm recognition and comparison. Our approach explores various preprocessing techniques, including the use of Mel spectrograms, to enhance model performance. The study employs adapted variants of the BLEU score to measure the accuracy of different models in recognizing rhythmic sequences, assessing their ability to identify the correct notes, maintain their order, and preserve the sequence length. Our findings indicate that both LSTM and bi-directional LSTM models struggle with longer sequences, suggesting the need for models capable of capturing long-term dependencies. Transformer models (Linear and Conv2D) showed superior accuracy, particularly in identifying rhythm and sequence length, but still exhibited limitations in error minimization.