Exploring gated recurrent unit autoencoders for efficient representation learning of transient audio: a case study on crash cymbal sounds

This study investigates the use of a Gated Recurrent Unit (GRU) based Autoencoder with Residual Vector Quantization (RVQ) for compressing transient crash cymbal sounds. Crash cymbals contain rapid transients and broad spectral ranges that make them difficult to compress using conventional audio codecs. To address this challenge, a neural audio codec architecture was designed, consisting of a GRU encoder-decoder and an RVQ module for compact latent representation. The model was trained for 100 epochs using the Adam optimizer and a combined L1-L2 reconstruction objective. Evaluation results show that the model achieved stable convergence with consistent performance across the training, validation, and testing datasets, reflected by an average Log-Spectral Distance (LSD) of around 6.2. The reconstructed audio preserved the temporal and spectral structure of the original sound, though minor background artifacts were observed during silent portions of the waveform.