A method for phrasing music data into meaningful musical pieces (e.g., bar and phrase) is an important function to analyze music data. To realize this function, we propose a method for extracting a unit of music data (musical unit) in the compressed domain of TwinVQ audio compression (MPEG-4 audio). Our key idea is to extract a musical unit from a sequence of autocorrelation coefficients computed in the encoding step of TwinVQ audio compression. We call the sequence of the autocorrelation coefficients the "autocorrelation sequence r". We use the k-th autocorrelation sequence rk (k=1, 2, ..., 20) of music data for extracting a musical unit of music data. First, we calculate the jk-th autocorrelation coefficient akjk of the k-th autocorrelation sequence rk (jk=38, 39, ..., 208; k=1, 2, ...,20). Second, for detecting the peak in the sequence (ak38, ak39, ..., ak208), the Laplacian filter is applied to the sequence. We then obtain the order pk for which the maximum differential coefficient is attained. Finally, we compute the musical unit using pk. To evaluate the performance of extracting the musical unit by our method, we collected 64 music data and obtained autocorrelation sequences by applying the TwinVQ encoder to each data. We then applied our extraction algorithm to each autocorrelation sequence. The experimental results reveal a very good performance in the extraction of a musical unit for phrasing music data.