<p>Detecting synthetic speech is a critical challenge in protecting speaker verification systems from spoofing attacks generated by modern text-to-speech and voice conversion technologies. Existing detection systems primarily depend on low-level acoustic features, which are effective but often lack interpretability and robustness against evolving attack methods. This paper introduces a phonetic-driven framework for synthetic speech detection that leverages phoneme posteriorgrams as features to explicitly model phoneme-level information. The proposed approach employs a residual convolutional neural network to learn discriminative representations from phonetic posterior distributions. Additionally, we propose a phoneme-importance analysis that quantifies the variation in Equal Error Rate when frames associated with specific phonemes are excluded. This analysis is used to prune less informative phonemes, resulting in a reduced feature representation. Experiments conducted on the Logical Access scenario of the ASVspoof 2019 dataset demonstrate that pruning more than half of the phonemes from cepstral features significantly reduces the Equal Error Rate. This observation highlights that synthetic speech artifacts are concentrated in specific phonetic units. Furthermore, combining the pruned cepstral-based features with raw waveform representations yields substantial improvements in detection performance, achieving competitive results compared to state-of-the-art systems.</p>