This paper introduces a new frequency domain technique for modal identification of output-only systems, where modal parameters are estimated without knowing the input. The technique is based on the decomposition of the spectral density function matrix, allowing the response spectra to be separated into single degree of freedom (SDOF) systems, each corresponding to an individual mode. This method is user-friendly and can accurately identify close modes even in the presence of strong noise. It also clearly indicates harmonic components in the response signals. The technique is an extension of the classical frequency domain approach, known as the Basic Frequency Domain (BFD) technique or the Peak Picking technique. However, it improves upon the classical approach by using Singular Value Decomposition (SVD) to decompose the spectral matrix into auto-spectral density functions. This decomposition is exact when the loading is white noise, the structure is lightly damped, and the mode shapes of close modes are geometrically orthogonal. If these assumptions are not satisfied, the decomposition is approximate but still more accurate than the classical approach. The theoretical background of the Frequency Domain Decomposition (FDD) technique is based on the relationship between the unknown inputs and the measured responses. The FRF can be expressed in partial fraction form, and the output PSD can be reduced to a pole/residue form. The residue matrix is an m x m Hermitian matrix and is proportional to the mode shape vector in the case of light damping. The identification algorithm involves estimating the power spectral density matrix and decomposing it using SVD. The first singular vector is an estimate of the mode shape, and the corresponding singular value is the auto power spectral density function of the SDOF system. The natural frequency and damping can be estimated from the crossing times and the logarithmic decrement of the SDOF auto correlation function. The technique was tested on a simulation of a two-storey building with close modes and noise. The results showed that the technique accurately identified the natural frequencies and damping ratios, even in the presence of noise. The mode shape estimates were very close to the exact results, especially for the nearly repeated modes. The technique clearly indicates the presence of harmonics in the response signal, allowing users to separate harmonic peaks from structural response peaks. The technique has been successfully applied to several civil engineering cases and mechanical engineering applications where the structure was loaded by rotating machinery.This paper introduces a new frequency domain technique for modal identification of output-only systems, where modal parameters are estimated without knowing the input. The technique is based on the decomposition of the spectral density function matrix, allowing the response spectra to be separated into single degree of freedom (SDOF) systems, each corresponding to an individual mode. This method is user-friendly and can accurately identify close modes even in the presence of strong noise. It also clearly indicates harmonic components in the response signals. The technique is an extension of the classical frequency domain approach, known as the Basic Frequency Domain (BFD) technique or the Peak Picking technique. However, it improves upon the classical approach by using Singular Value Decomposition (SVD) to decompose the spectral matrix into auto-spectral density functions. This decomposition is exact when the loading is white noise, the structure is lightly damped, and the mode shapes of close modes are geometrically orthogonal. If these assumptions are not satisfied, the decomposition is approximate but still more accurate than the classical approach. The theoretical background of the Frequency Domain Decomposition (FDD) technique is based on the relationship between the unknown inputs and the measured responses. The FRF can be expressed in partial fraction form, and the output PSD can be reduced to a pole/residue form. The residue matrix is an m x m Hermitian matrix and is proportional to the mode shape vector in the case of light damping. The identification algorithm involves estimating the power spectral density matrix and decomposing it using SVD. The first singular vector is an estimate of the mode shape, and the corresponding singular value is the auto power spectral density function of the SDOF system. The natural frequency and damping can be estimated from the crossing times and the logarithmic decrement of the SDOF auto correlation function. The technique was tested on a simulation of a two-storey building with close modes and noise. The results showed that the technique accurately identified the natural frequencies and damping ratios, even in the presence of noise. The mode shape estimates were very close to the exact results, especially for the nearly repeated modes. The technique clearly indicates the presence of harmonics in the response signal, allowing users to separate harmonic peaks from structural response peaks. The technique has been successfully applied to several civil engineering cases and mechanical engineering applications where the structure was loaded by rotating machinery.