Intelligent training method and system for audio noise reduction model

The invention discloses an intelligent training method and system for an audio noise reduction model, and relates to the technical field of audio noise reduction, a loss assessment module analyzes a spectrum distortion factor Psyzn of each frame according to the spectrum states of model training audio data and real label audio data, which is beneficial to more accurately measuring the performance of the model in different spectrum states, and improves the performance of the model. And the model training module adopts a recurrent neural network technology, so that the time sequence information of the audio signals is better captured, and the adaptability of the noise reduction model to a complex audio scene is improved. The system fully considers under-fitting and over-fitting states which may occur in training data processing. When the diversity of training data is processed, the system not only pays attention to the diversity of noise, but also considers the change of audio signals in different speakers, env