Supervised Speech Separation Using The Optimal Ratio Mask

Speech separation technique is aimed at separating interested target speech from a noisy mixture.Speech separation is widely applied in automatic speech recognition?hearing aids design and telephone communication,etc.Although there are already many achievements in the area of speech separation,while further developments are still expected.This dissertation is focused on monaural speech separation.Separation problem can be treated as a supervised learning problem and solved by adopting supervised learning algorithm.Training target is one of the most important part in a supervised speech separation algorithm,which have great impact on the performance of speech separation.The ideal binary mask and the ideal ratio mask are most popular training targets,which need the assumption that the clean speech is dependent with the noise which is hard to obtain in real environment.The complex ratio mask and phase sensitive mask are hard to estimate leads to its practical performance worse than that in theory.This dissertation utilize the optimal ratio mask which gains the maximize signal-to-noise ratio(SNR)in theory.ORM consider the relationship between clean speech and the noise in real environment.We construct a new separation target function by combining ORM with the supervised speech separation system,utilizing the ORM as training target,which is proposed as a new separation algorithm in this dissertation.We carried out experiments under various noise situation and SNRs,analyze and compare the separation results with several popular training targets.The results show that the separation algorithm we proposed outperform the others as a better solution of speech separation problem in real environments.Considering that ORM separates mainly based on the relations between clean speech and noise,we further experiment with the separation of speech from different speakers,our method also shows superiority.Monaural speech dereverberation is another key problem in the area of signal processing.As the advancing of deep learning,researchers adapt it in speech dereverberation and achieve excellent results.In this dissertation,we utilize the algorithm we proposed into the dereverberation problem.The experiment result shows ideal performance.