Design of a Multi-scenario Audio Adaptive Coding and Denoising System with Evaluation Framework based on Random Forests and Short-Time Frequency Transform

This paper employs entropy weighting, random forest regression, short-time Fourier transform (STFT), Mel-spectrum analysis and other processing techniques to investigate optimisation of storage efficiency, signal fidelity and encoding/decoding efficiency in multi-scenario audio processing, thereby providing technical support for intelligent audio processing. To achieve multidimensional performance evaluation, a three-dimensional assessment framework encompassing storage efficiency, signal fidelity, coding efficiency, and scenario adaptability is constructed. The Entropy Weighting Method quantifies the weighting of each metric, while incorporating scenario-specific requirements to assign differentiated weights, thereby establishing a multi-scenario adaptive evaluation system. At the parameter dynamic decision level, time-frequency features such as spectral entropy, dynamic range, and spectral complexity are extracted via STFT. Key features are selected through LASSO to optimise the random forest regression model, while incorporating signal type and device performance parameters to achieve adaptive adjustment of sampling rate, bitrate, and encoding format. For noisy signals, preprocessing first eliminates data interference. Features are then extracted using Mel-spectrum and cepstral coefficients to identify background noise, burst noise, and other types, applying corresponding filtering algorithms for suppression. Experimental validation demonstrates this approach effectively enhances storage compression efficiency and signal processing accuracy, adapting to diverse audio processing requirements across multiple scenarios.