Provided is a quaternion ESPRIT parameter estimation method of a sparse circular sound vector sensor array. The array receives K far-field narrowband sound signals of different frequencies and independent from one another, according to each of two groups of complex-domain data received by the array, the same sound pressure of sampling data of the array is overlapped with vibration speed data in three directions vertical to one another to form quaternion linear sampling data; characteristic decomposition and related operation are carried out on an autocorrelation matrix of full-array reception data to obtain estimated values of array guiding vectors; the estimated values of the array guiding vectors are used to reconstruct subarray guiding vectors, and the rotation-invariant relation among the subarray guiding vectors is used to obtain a roughly estimated value of the cosine in the signal direction; and the roughly estimated value of the cosine in the signal direction is used to determine a phase period fuzzy number of a spatial domain guiding vector, and a least square method is used to obtain an accurate estimated value of the arrival angle of the signal. According to the invention, orthogonal feature among components of the sound vector sensors is maintained, error robustness of a model is higher, and the problem of phase fuzziness of the sparse vector sensor circular array is solved.