The article discusses the development of a new generative model for estimating room impulse responses (RIR), which is crucial in understanding sound propagation in real-world environments. The authors propose an alternate generator architecture for this purpose, using an autoencoder with residual quantization. They then apply this to the RIR estimation problem, casting it as a reference-conditioned autoregressive token generation task. The system is shown to be preferable to other baselines across various evaluation metrics, indicating its potential for effective RIR estimation.
Publication date: 8 Nov 2023
Project Page: https://sh-lee97.github.io/neural-ir-est
Paper: https://arxiv.org/pdf/2311.02581
