On Evaluation of Blind Audio Source Separation
In this paper, objective and subjective methods on evaluating the performances of audio source separations are proposed. The proposed subjective listening test method is based on the estimated source decomposition and MUSHRA test. In the objective measures, several well-known hearing characteristics, such as the equal-loudness contours and the masking effect, are considered to recently developed measures which are defined by signal-to-distortion ratio (SDR), signal-to-interferences ratio (SIR), signal-to-noise ratio (SNR), and signal-to-artifacts ratio (SAR), to make these measures more relevant to human auditory system. A statistical analysis between the results of the proposed objective method and those of listening tests is performed for verification of the improvement in comparison to the existing performance measures.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is temporarily free for AES members.