<p>Securing large-scale audio streams without compromising perceptual fidelity remains a significant challenge, as conventional chaos-based ciphers often suffer from shallow diffusion and static key schedules. This study proposes a two-round dynamic audio encryption framework driven by a Fractional Graph Convolutional Network (FrGCN), which integrates fractional Fourier transform (FrFT)-based amplitude-phase separation with content-adaptive chaotic keystreams. First, the FrFT decomposes audio frames into magnitude and phase spectra in the fractional domain. The magnitudes are perturbed by Poisson noise, while the phases are scrambled via a piecewise linear chaotic map, effectively mitigating structured spectral leakage. Next, a k-nearest-neighbor graph, constructed from frame features, is processed by a two-layer FrGCN. The resulting semantic representations are hashed to initialize refined chaotic maps, generating plaintext-sensitive keystreams. Finally, these keystreams drive two rounds of encryption, including: semantic XOR diffusion, logistic-assisted chaotic scrambling, and DNA-based complementary mutation, collectively achieving high nonlinearity and deep diffusion. The proposed algorithm operates with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\mathcal {O}(N\log N)\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="script">O</mi> <mo stretchy="false">(</mo> <mi>N</mi> <mo>log</mo> <mi>N</mi> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> complexity and millisecond-level latency, ensuring real-time feasibility. Experimental results and analyses demonstrate that the framework provides a robust and efficient solution for real-time, large-scale audio protection.</p>