<p>This paper presents HySyllNet, a hybrid architecture for real-time syllable segmentation that processes input incrementally, character by character, without requiring access to full word boundaries. The proposed model integrates causal convolutional feature extraction, a local attention mechanism, and a unidirectional recurrent module with a lightweight symbolic filter grounded in phonotactic rules—such as CV and CVC syllabic patterns. This neural–symbolic design enables strictly causal inference, ensuring that predictions at each time step depend only on past and present characters, a property that is essential for streaming applications. Experiments conducted on two complementary datasets—a curated corpus of English words and a collection of French spontaneous speech transcriptions—demonstrate that HySyllNet achieves an F1-score of 90.1% on the noisy streaming corpus while maintaining an average inference latency of approximately 2&#xa0;ms per character. These results correspond to a 33% reduction in latency relative to bidirectional baselines, with no loss in segmentation accuracy. The compact architecture, constant memory footprint, and integration of linguistic constraints make HySyllNet a practical solution for latency-sensitive applications such as keystroke-based behavioral biometrics, real-time speech interfaces, and embedded natural language processing systems.</p>