Accurate detection of echoes in audio signals is crucial for applications such as telecommunications, architectural acoustics, and audio forensics. Traditional signal processing methods, such as cross-correlation and autocorrelation, often struggle with accuracy under noisy, reverberant, or complex echo conditions. This paper introduces a Convolutional Neural Network (CNN)-based approach that leverages time-frequency representations (Mel-spectrograms) to improve echo delay estimation. We demonstrate significant improvements over traditional techniques in both controlled and real-world scenarios. Our model achieves lower error rates, higher robustness to noise, and strong generalization across diverse environments. Experimental results and extensive evaluations highlight the potential of deep learning to enhance echo detection, paving the way for improved audio quality in telecommunications, better room acoustic characterization, and more reliable acoustic forensics analysis.