Early adductive reasoning for blind signal separation
MetadataShow full item record
We demonstrate that explicit and systematic incorporation of abductive reasoning capabilities into algorithms for blind signal separation can yield significant performance improvements. Our formulated mechanisms apply to the output data of signal processing modules in order to conjecture the structure of time-frequency interactions between the signal components that are to be separated. The conjectured interactions are used to drive subsequent signal separation processes that are as a result less blind to the interacting signal components and, therefore, more effective. We refer to this type of process as early abductive reasoning (EAR); the “early” refers to the fact that in contrast to classical Artificial Intelligence paradigms, the reasoning process here is utilized before the signal processing transformations are completed. We have used our EAR approach to formulate a practical algorithm that is more effective in realistically noisy conditions than reference algorithms that are representative of the current state of the art in two-speaker pitch tracking. Our algorithm uses the Blackboard architecture from Artificial Intelligence to control EAR and advanced signal processing modules. The algorithm has been implemented in MATLAB and successfully tested on a database of 570 mixture signals representing simultaneous speakers in a variety of real-world, noisy environments. With 0 dB Target-to-Masking Ratio (TMR) and no noise, the Gross Error Rate (GER) for our algorithm is 5% in comparison to the best GER performance of 11% among the reference algorithms. In diffuse noisy environments (such as street or restaurant environments), we find that our algorithm on the average outperforms the best reference algorithm by 9.4%. With directional noise, our algorithm also outperforms the best reference algorithm by 29%. The extracted pitch tracks from our algorithm were also used to carry out comb filtering for separating the harmonics of the two speakers from each other and from the other sound sources in the environment. The separated signals were evaluated subjectively by a set of 20 listeners to be of reasonable quality.