Introduction
Acoustic scene classification (ASC) systems often face performance degradation when deployed on devices different from those used during training, a phenomenon known as "device mismatch." This occurs because hardware variations—specifically microphone characteristics and signal processing pipelines—create an acoustic gap between training (source) and deployment (target) devices. Standard adaptation techniques, such as fine-tuning on limited target data, often fail to capture the underlying structural relationships between domains. This work addresses device adaptation by mathematically adjusting a pre-trained source model using a Bayesian framework centered on latent variables.
Methodology
The proposed framework focuses on the latent variables ($Z$) within a Deep Neural Network (DNN), representing intermediate embeddings between input features and final classification outputs. The core principle is to treat the source domain's latent variables as a prior distribution, which is then updated using the likelihood of the target data to derive a posterior distribution.
Two primary approaches are explored:
- Bayesian Adaptation via Variational Bayes (BA-VB): This method approximates the full posterior distribution of latent variables using a Gaussian mean-field approximation, minimizing the Kullback-Leibler (KL) divergence between the estimated and true distributions.
- Bayesian Adaptation via Maximum a Posteriori (BA-MAP): A point-estimation approach that incorporates prior knowledge. The authors demonstrate that under a Dirichlet assumption for latent variables (soft outputs), BA-MAP becomes equivalent to Teacher-Student (TS) learning.
Experimental Setup
The methods were evaluated using the DCASE 2020 Challenge Task 1a dataset. The source data consisted of approximately 10,000 audio clips from "Device A," while the target data included 750 clips for each of eight different devices (Device B, C, and s1-s6). The dataset features parallel recordings, where the same acoustic scene was captured simultaneously by different hardware, providing a robust ground truth for comparison.
Results and Analysis
The Bayesian methods demonstrated superior performance compared to baseline and state-of-the-art transfer learning techniques. As shown in the results table below, the BA-VB approach achieved the highest average accuracy.
The BA-VB method provided a 14.47% absolute improvement over the "No transfer" baseline and outperformed specialized algorithms like Relational Knowledge Distillation (RKD) and Selective Pseudo-labeling (SP). The success of BA-VB suggests that modeling the full distribution of latent variables, rather than a single point estimate, provides greater robustness against the noise and hardware variations inherent in different recording devices.
Conclusion
This paper introduces a principled Bayesian framework for device adaptation in acoustic scene classification. By treating intermediate representations as latent variables and employing Variational Bayes, the researchers successfully bridged the performance gap caused by device mismatch. While BA-VB is more computationally intensive than MAP due to the sampling required by the reparameterization trick, its performance gains are significant. Future work will investigate the scalability of this approach to larger tasks and its performance in non-parallel data scenarios.
Summary updated