The growing number of audio deepfakes calls for strong detection tools to protect digital security and voice authentication. Presenting a sophisticated deepfake detection system that uses Constant Q Cepstral Coefficients (CQCC) and a multi-scale neural network with hierarchical attention, we get state-of-the-art performance on the ASVspoof 2019 benchmark. Our model achieves a record 4.91% Equal Error Rate (EER) by exposing minor synthetic artefacts via phase-aware feature extraction and adaptive attention techniques, hence improving baseline systems by 48.7%. Among the main innovations are: (1) Multi-scale phase analysis capturing vocoder distortions, (2) Dynamic frequency band weighting via cross-attention, and (3) Dilated residual blocks retaining temporal accuracy. While visual analysis shows the capacity of our model to detect discontinuous harmonics and anomalous energy spikes in synthetic speech, ablation studies show phase characteristics account for 33% of detection accuracy. Maintaining computational efficiency, the system far outperforms current methods (6.5% lower EER than prior best). With significant impact size (Cohen’s d = 1.24), statistical validation verifies p < 0.01 significance. These developments set fresh benchmarks for actual deployment in voice authentication, forensic analysis, and media verification. Generalisation to unknown spoofing tactics and real-time detection situations will be the subject of future development. With direct relevance in financial security, legal forensics, and disinformation avoidance, this study offers vital methods for fighting manipulation of audio produced by artificial intelligence. Our open-source solution allows quick integration into current speech verification systems.