Font Size: a A A

Research And Application Of Speech Enhancement Technology

Posted on:2021-06-15Degree:MasterType:Thesis
Country:ChinaCandidate:C GengFull Text:PDF
GTID:2518306338985909Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Speech enhancement is a basic research topic in the field of speech signal processing,and its purpose is to improve the quality of noisy speech.This technology is not only related to signal processing theory,but also to human auditory perception and phonetics.Speech enhancement is a supporting technology that plays a vital role in improving the robustness of other application systems.Such as speech recognition and speaker recognition currently on the market use speech enhancement technology to ensure the stability of its overall performance.In recent years,speech enhancement methods based on deep learning technology have made great progress.Compared with traditional enhancement methods such as Wiener filtering and Kaman filtering,they have shown greater advantages in terms of performance and universality.This paper has carried out research work in the following aspects:1)deep learning based speech enhancement methods;2)application of speech enhancement technology in speaker recognition system;3)source separation technology.(1)Propose an end-to-end speech enhancement system based on zero phase.To solve the problem of phase spectrum estimation in speech enhancement,a zero-phase feature extraction scheme is proposed,and an end-to-end speech enhancement system is designed in combination with Unet.In terms of the objective function of the system,the original wSDR function is modified,which effectively improve the performance of enhanced speech.(2)Design and implement a robust end-to-end speaker recognition system.By integrating above speech enhancement system with the I-vector based speaker recognition system,and proposing the improved scheme of the speaker model and its scoring mechanism.Experiments show that the above improvements improve the robustness of the basic speaker recognition system.(3)Based on the above speech enhancement system,a deep learning architecture is proposed to accomplish the task of source separation.In this architecture,the objective function of the system is optimized by combining the evaluation indicators of source separation.
Keywords/Search Tags:speech enhancement, zero phase, I-vector, end-to-end, source separation
PDF Full Text Request
Related items