19th Asia-Pacific Vibration Conference (APVC2021)

317 / 2021-04-29 12:07:42

Separation and conversion of mono speech and noise

Speech feature parameter; Computational auditory scene analysis; Cycle generative adversarial network,Signal processing;

As a carrier, sound is important to the representation and transmission of information. A silent world is unimaginable. Speech and noise are two forms of sound. Speech is the bridge of human communication. Although noise is disturbing, it contains a lot of natural information. Usually, the two are mixed and need to be separated before accurate characterization and transformation. The conversion between different sounds is also of great significance. In this paper, the separation and conversion of speech and noise are deeply explored and studied. A wider range of mono channel speech and noise separation problems, a multi feature hybrid separation model is proposed. Through time-frequency decomposition, multi speech feature parameter extraction and selection, auditory field calculation scene analysis, speech information flow reorganization, deep neural network to achieve the optimal model selection and adapt to multi scene speech and noise separation. Combined with the above separation model and method to build a new active noise reduction system. Finally, aiming at the problem of voice conversion, a conversion model based on MFCC parameters, self-encoder and Cycle Generative Adversarial Network is proposed to realize the fast conversion between male and female voice with different styles.