It is a well-established practice to build a robust system for sound event detection by training supervised deep learning models on large datasets, but audio data collection and labeling are often challenging and require large amounts of effort. This paper proposes a workflow based on few-shot metric learning for emergency siren detection performed in steps: prototypical networks are trained on publicly available sources or synthetic data in multiple combinations, and at inference time, the best knowledge learned in associating a sound with its class representation is transferred to identify ambulance sirens, given only a few instances for the prototype computation. Performance is evaluated on siren recordings acquired by sensors inside and outside the cabin of an equipped car, investigating the contribution of filtering techniques for background noise reduction. The results show the effectiveness of the proposed approach, achieving AUPRC scores equal to 0.86 and 0.91 in unfiltered and filtered conditions, respectively, outperforming a convolutional baseline model with and without fine-tuning for domain adaptation. Extensive experiments conducted on several recording sensor placements prove that few-shot learning is a reliable technique even in real-world scenarios and gives valuable insights for developing an in-car emergency vehicle detection system.

In the recent years, several supervised and unsupervised approaches to fall detection have been presented in the literature. These are generally based on a corpus of examples of human falls that are, though, hard to collect. For this reason, fall detection algorithms should be designed to gather as much information as possible from the few available data related to the type of events to be detected. The one-shot learning paradigm for expert systems training seems to naturally match these constraints, and this inspired the novel Siamese Neural Network (SNN) architecture for human fall detection proposed in this contribution. Acoustic data are employed as input, and the twin convolutional autoencoders composing the SNN are trained to perform a suitable metric learning in the audio domain and, thus, extract robust features to be used in the final classification stage. A large acoustic dataset has been recorded in three real rooms with different floor types and human falls performed by four volunteers, and then adopted for experiments. Obtained results show that the proposed approach, which only relies on two real human fall events in the training phase, achieves a F $_1$ -Measure of 93.58% during testing, remarkably outperforming the recent supervised and unsupervised state-of-art techniques selected for comparison.

Audio equalization is an active research topic aiming at improving the audio quality of a loudspeaker system by correcting the overall frequency response using linear filters. The estimation of their coefficients is not an easy task, especially in binaural and multipoint scenarios, due to the contribution of multiple impulse responses to each listening point. This paper presents a deep learning approach for tuning filter coefficients employing three different neural networks architectures—the Multilayer Perceptron, the Convolutional Neural Network, and the Convolutional Autoencoder. Suitable loss functions are proposed for each architecture, and are formulated in terms of spectral Euclidean distance. The experiments were conducted in the automotive scenario, considering several loudspeakers and microphones. The obtained results show that deep learning techniques give superior performance compared to baseline methods, achieving almost flat magnitude frequency response.

This paper describes a novel Deep Learning method for the design of IIR parametric filters for automatic multipoint audio equalization, that is the task of improving the sound quality of a listening environment at multiple listening points employing multiple loudspeakers. The filters are designed to approximate the inverse of the RIR and achieve almost flat magnitude response. A simple and effective neural architecture, named BiasNet, is proposed to determine the IIR equalizer parameters. This novel architecture is conceived for optimization and, as such, is able to produce optimal IIR equalizer parameters at its output, after training, with no input required. In absence of input, the presence of learnable non-zero bias terms ensures that the network works properly. An output scaling method is used to obtain accurate tuning of the IIR filters center frequency, quality factor and gain. All layers involved in the proposed method are shown to be differentiable, allowing backpropagation to optimize the network weights and achieve, after a number of training iterations, the optimal output according to a given RIR. The parameters are optimized with respect to a loss function based on a spectral distance between the measured and desired magnitude response, and a regularization term is used to keep the same microphone-loudspeaker energy balance after equalization. Two experimental scenarios are employed, a room and a car cabin, with several loudspeakers. The performance of the proposed method improves over the baseline techniques and achieves an almost flat band at a lower computational cost.