Coupled Sparse NMF vs. Random Forest Classification for Real Life Acoustic Event Detection

Iwona Sobieraj, Mark D. PlumbleyDCASE 2016

Introduction

Acoustic Event Detection (AED) is the computational task of identifying the presence and temporal boundaries of specific sounds within an audio stream. While monophonic detection—identifying a single sound at a time—is a well-studied problem, real-world environments are inherently polyphonic, featuring multiple overlapping sound sources. This paper investigates two distinct paradigms for polyphonic AED: a dictionary-learning approach based on Coupled Sparse Non-negative Matrix Factorization (CSNMF) and a supervised learning approach using Multi-class Random Forests (MRF). By evaluating these methods on complex, real-life audio data, the researchers aim to identify robust alternatives to data-intensive Deep Learning models.

Methodology

The study compares two primary frameworks designed to handle the complexities of overlapping acoustic events.

Coupled Sparse Non-negative Matrix Factorization (CSNMF)

NMF is a popular technique for decomposing a complex signal into a "dictionary" of basic components and an "activation matrix" that describes when these components are active. The authors extend this into a Coupled Sparse framework. In this model, two dictionaries are learned simultaneously: one for audio features (Mel-spectrograms) and another for activity labels. These dictionaries are "coupled" because they share the same activation matrix, ensuring that the detected acoustic features directly correspond to event labels. To improve isolation of distinct sounds, an $l_1$-norm penalty is applied to enforce sparsity, preventing the model from over-fitting to noise.

Multi-class Random Forest (MRF)

The second approach utilizes an ensemble of 500 decision trees. To capture temporal dynamics, the model is trained on "shingled" 2D spectro-temporal patches—blocks of consecutive audio frames. To address polyphony without complex source separation, the authors employ a "label powerset" strategy: every unique combination of overlapping sounds is treated as a single, distinct class. For a set of $E$ possible events, this results in $2^E$ potential classes, allowing the forest to learn the specific spectral signatures of overlapping sounds directly.

Dataset Characteristics

The systems were rigorously tested using the TUT Sound Events 2016 dataset, which captures audio from two distinct real-world settings: a Residential Area (outdoor) and a Home (indoor).

Table 1: Event classes and instance counts for the Residential Area and Home environments in the TUT Sound Events 2016 dataset.

As shown in Table 1, the Residential Area dataset is highly unbalanced, with "bird singing" and "car passing by" accounting for nearly 74% of all instances. The Home environment, while more balanced in terms of class distribution, presents a higher degree of polyphony and a prevalence of short, impulsive "impact" sounds like dishes clinking or cupboards snapping.

Results and Analysis

Performance Comparison

The Multi-class Random Forest (MRF) emerged as the most effective system across nearly all metrics. According to the researchers' findings, the MRF achieved an average F-score of 44.1%, significantly outperforming both the DCASE 2016 baseline (34.3%) and the CSNMF approach (29.2%).

Table 2: Comparative performance (F-score and Error Rate) of the baseline, CSNMF, and MRF systems across different environments.

As summarized in Table 2, the MRF consistently achieved lower Error Rates (ER) and higher F-scores compared to the other systems. Notably, all models struggled significantly more in the "Home" environment than in the "Residential" area. This performance gap is attributed to the acoustic complexity of indoor environments, where short-duration events often overlap and are easily masked by background noise.

The Role of Temporal Context and Sparsity

The study revealed that the optimal temporal window (the number of "shingled" frames) varies by environment. In outdoor settings, longer windows (4–6 frames) helped capture stable sounds like birdsong. Conversely, in the home environment, longer windows tended to blur the sharp spectral features of impact sounds, leading to decreased accuracy. For the CSNMF approach, the enforcement of sparsity ($\lambda$) proved critical; higher sparsity settings were required in the home environment to filter out noisy activations and isolate overlapping events.

Limitations and Future Outlook

Despite the success of the Random Forest approach, it faces a significant "unseen combination" limitation: it can only detect overlaps it has encountered during training. Furthermore, the extreme class imbalance in the residential dataset led to 0% F-scores for rare classes, as the model prioritized the most frequent events.

Future research directions include moving toward true multi-label classification to handle novel sound combinations and exploring advanced features like scattering transforms to better characterize the transient nature of impact sounds.

Summary updated

BibTeX
@inproceedings{sobieraj2016coupled,
  author = {Iwona Sobieraj and Mark D. Plumbley},
  title = {{Coupled Sparse NMF vs. Random Forest Classification for Real Life Acoustic Event Detection}},
  booktitle = {DCASE},
  year = {2016},
}
Figure