Spectral proper orthogonal decomposition (SPOD) is applied to direct numerical simulation (DNS) datasets of a lean and a stoichiometric methane/air turbulent premixed jet flame. SPOD is used to extract the coherent structures that correlate with the radiated sound by using an inner product based on a linearized disturbance energy. Two types of structures are prominent in the data. The first type arises in the jet's shear layer and is linked to the Kelvin-Helmholtz (K-H) instability, which is an important mechanism of sound generation in non-reacting jets. These structures produce sound through deformation of the flame front in the shear layer. They contain most of the acoustic energy and are dominant at Strouhal numbers (defined based on the jet's diameter and the inlet mean velocity) less than unity. The second type of structures is found near the jet centreline, where large fluctuations of the flame surface are observed. The structures are linked to small non-linear flame dynamics and to the Orr mechanism. They travel at a speed close to the inlet mean velocity and are important at higher Strouhal numbers. Regardless of their energy content, both types of structures have important contributions to the broadband nature of combustion noise.