Introduction
Acoustic Scene Classification (ASC) is the computational task of identifying the environment in which an audio recording was made—for instance, distinguishing between an "office," a "park," or a "metro station." The primary challenge in ASC lies in developing effective audio representations. While traditional approaches relied on hand-crafted features like Mel-frequency cepstral coefficients (MFCCs), contemporary research has shifted toward feature learning. However, two significant hurdles remain: scaling issues where low-energy background sounds are overshadowed during optimization, and representation gaps between low-level time-frequency data and mid-level learned features.
Methodology
This research introduces two primary technical contributions to address these challenges: Task-driven Nonnegative Matrix Factorization with Adaptive Scaling (TNMF-AS) and a hybrid multi-view Deep Neural Network (DNN-M).
Task-driven NMF with Adaptive Scaling (TNMF-AS)
Nonnegative Matrix Factorization (NMF) decomposes audio into a dictionary of recurring sounds and their corresponding activations. Task-driven NMF (TNMF) jointly learns this dictionary and the classifier. The authors propose Adaptive Scaling, a technique analogous to Batch Normalization. By tracking the mean and variance of activations across mini-batches, the model ensures that all dictionary components contribute equitably to the classification decision, regardless of their original energy levels. This prevents the omission of subtle but discriminative acoustic cues.
Hybrid DNN-M Architecture
The proposed DNN-M architecture utilizes a "multi-view" approach to bridge the representation gap. As shown in Figure 1, the network processes two parallel inputs:
- Branch A: Unsupervised NMF activation features (mid-level representation).
- Branch B: Constant-Q Transform (CQT) representations (low-level time-frequency data).
These branches are merged via concatenation in the deeper layers of the network. This design leverages NMF as a pre-trained feature extractor while ensuring that the CQT branch preserves raw information.
Experimental Setup
The systems were evaluated using the DCASE 2017 challenge dataset (Task 1), which comprises 13 hours of urban audio recorded with binaural microphones. The dataset includes 4,680 10-second recordings across 15 distinct environment classes, such as bus, cafe, forest, and library.
Results and Analysis
The experimental results demonstrate that the proposed methods significantly outperform the DCASE 2017 baseline.
As summarized in the results table, the TNMF-AS system achieved an accuracy of 88.2% (at K=512), a notable improvement over the standard TNMF (86.8%). This validates the importance of feature normalization during the learning loop. The hybrid DNN-M model reached 87.9% accuracy, outperforming individual branches and confirming the complementary nature of low-level and mid-level data.
The final system fusion, combining the log-probabilities of TNMF-AS and DNN-M using stereo audio, achieved a peak accuracy of 90.1%, compared to the baseline of 77.2%. Analysis of the confusion matrix indicates near-perfect identification for environments like "bus" and "cafe," though some confusion persists between similar scenes like "office" and "library" or "park" and "residential area."
Conclusion
This work highlights the efficacy of adaptive scaling in supervised NMF and the benefits of hybrid architectures for acoustic scene classification. By balancing feature contributions and integrating multiple data views, the proposed methods achieve state-of-the-art performance on the DCASE 2017 benchmark. Future research will explore end-to-end architectures that jointly learn NMF representations and DNN layers to further enhance discriminative power in complex acoustic environments.
Summary updated