Towards Improving Speaker Distance Estimation through Generative Impulse Response Augmentation

A team of researchers (Ratnarajah, Ergezer, Nair, Athi) has published a paper accepted at the GenDA 2025 workshop (an ICASSP 2025 satellite event) that tackles the Room Acoustics and Speaker Distance Estimation (SDE) Challenge. Their approach leverages generative augmentation of room impulse responses (RIRs) to overcome sparse training data. They used the open-source FastRIR generator, conditioned only on speaker and listener spatial coordinates, to synthesize realistic RIRs. A custom quality filter ensured the generated RIRs aligned with challenge-specific RIR distributions, and hyperparameter optimization was applied during model fine-tuning.

The results are striking: the mean absolute error (MAE) for distance estimation dropped from 1.66m to 0.6m in GWA-style rooms (a 64% improvement) and from 2.18m to 0.69m in Treble rooms (a 68% improvement). The gains were especially pronounced at medium to long distances, where traditional models struggle. This work demonstrates that generative RIR augmentation is a practical, low-cost way to boost SDE performance without collecting massive real-world datasets—opening up applications in smart speakers, hearing aids, and acoustic monitoring.