Research On Acoustic Feature Extraction Algorithm Based On Adversarial Contrastive Learning

Neural network models can obtain desired outputs through input data by fitting a large mathematical representation using differentiable and linear operations.Compared to traditional feature models,neural network models perform better,especially in the audio domain.Research on the role of noise in audio and how to generate natural audio signals has become a focus for researchers,and developing a general acoustic model that integrates high-dimensional acoustic feature information is crucial for further advancing AI and achieving social and economic benefits.This article focuses on studying audio feature extraction algorithms in different scenarios and proposes the following main research work:In multiple downstream task scenarios,a sound denoising and enhancement algorithm based on a hybrid convolutional network is proposed.Models that use traditional acoustic features such as spectrograms suffer from severe information loss on the input,while models that use time-domain sound signals ignore frequency and phase information.To address this issue,this article uses a neural network to extract different types of information from sound signals and proposes a universal acoustic feature extraction model that can be used in various downstream tasks.By using unsupervised contrastive learning,the model enhances sound quality and reduces noise when given noisy audio samples,and outputs their acoustic features.In the high-quality sound reconstruction scenario,a sound quantization reconstruction algorithm based on conditional diffusion models is proposed.Mainstream reconstruction algorithms use generative training methods,but noise in different scenarios is difficult to distinguish,resulting in low-quality reconstructed samples.This article proposes a quantization reconstruction algorithm that improves the inverse diffusion process of adversarial training and diffusion models,making it applicable to sound reconstruction tasks.The algorithm further uses a quantization encoder and codebook space to control the inverse diffusion process based on the diffusion model.