Tsubaki (Germany)

Ensemble learning is a useful technique to combine several models to improve classification performance. Previous research on speech classification shows the benefit of ensemble learning over single models; however, there is no systematic evaluation on the accommodating single model performance in ensemble learning for speech classification. There is also no detailed report on the usefulness of ensemble learning over task-specific acoustic feature. We evaluate performance-weighted ensemble learning by taking into account the previous single model performance for speech classification tasks, including speech emotion recognition, laughter type classification, gender prediction and age prediction. We compare different weighting schemes based on unweighted and weighted accuracies, in which we also reported our results using these metrics. Results on six tasks and eleven datasets show diverse findings on the effectiveness of performance-weighted ensemble learning over other ensemble methods and single models.

<p>Abusive content in online social networks is a well-known problem that can cause serious psychological harm and incite hatred. The ability to upload audio data increases the importance of developing methods to detect abusive content in speech recordings. However, simply transferring the mechanisms from written abuse detection would ignore relevant information such as emotion and tone. In addition, many current algorithms require training in the specific language for which they are being used. This paper proposes to use acoustic and prosodic features to classify abusive content. We used the ADIMA data set, which contains recordings from ten Indic languages, and trained different models in multilingual and cross-lingual settings. Our results show that it is possible to classify abusive and non-abusive content using only acoustic and prosodic features. The most important and influential features are discussed.</p>

Abstract Facial behaviour forms an important cue for understanding human affect. While a large number of existing approaches successfully recognize affect from facial behaviours occurring in daily life, facial behaviours displayed in work settings have not been investigated to a great extent. This paper presents the first study that systematically investigates the influence of spatial and temporal facial behaviours on human affect recognition in work-like settings. We first introduce a new multi-site data collection protocol for acquiring human behavioural data under various simulated working conditions. Then, we propose a deep learning-based framework that leverages both spatio-temporal facial behavioural cues and background information for workers’ affect recognition. We conduct extensive experiments to evaluate the impact of spatial, temporal and contextual information for models that learn to recognize affect in work-like settings. Our experimental results show that (i) workers’ affective states can be inferred from their facial behaviours; (ii) models pre-trained on naturalistic datasets prove useful for predicting affect from facial behaviours in work-like settings; and (iii) task type and task setting influence the affect recognition performance.

With the advent of 'heavy Artificial Intelligence' – big data, deep learning, and ubiquitous use of the internet, ethical considerations are widely dealt with in public discussions and governmental bodies. Within Computational Paralinguistics with its manifold topics and possible applications (modelling of long-term, medium-term, and short-term traits and states such as personality, emotion, or speech pathology), we have not yet seen that many contributions. In this article, we try to set the scene by (1) giving a short overview of ethics and privacy, (2) describing the field of Computational Paralinguistics, its history and exemplary use cases, as well as (de-)anonymisation and peculiarities of speech and text data, and (3) proposing rules for good practice in the field, such as choosing the right performance measure, and accounting for representativity and interpretability.

OPINION article Front. Digit. Health, 26 June 2020 | https://doi.org/10.3389/fdgth.2020.00005

Abstract An advanced unsteady multi-hole pressure probe is developed in cooperation between the Chair of Aerodynamics and Fluid Mechanics of the Technical University of Munich, the probe manufacturer Vectoflow GmbH and the sensor producer fos4X GmbH. The application of additive manufacturing enables a wide variation in probe geometries. The spatial characteristics of the unsteady probe are determined in the aerodynamic calibration in a known steady uniform jet. Furthermore, the acoustic system inside the channels of the probe is investigated experimentally in the dynamic calibration. Both aerodynamic and dynamic calibration ensure an accurate reconstruction of the velocity components. Measurements of the near wake of a circular cylinder have been recorded with the unsteady probe and compared to data from the literature and corresponding hotwire anemometry measurements. Further improvements to enhance the range of applications of the probe have been initiated: This includes the miniaturization of the probe. Moreover, an unsteady differential fiber-optic sensor and its optic and passive measurement principle are introduced which shall improve the main dynamic specifications of the probe. By combining the additive manufactured pressure probe design and the novel fiber-optic sensor, more accurate measurements of high frequency flow phenomena are expected.

Speech emotion recognition (SER) is a powerful tool for endowing computers with the capacity to process information about the affective states of users in human-machine interactions. Recent research has shown the effectiveness of graph embedding-based subspace learning and extreme learning machine applied to SER, but there are still various drawbacks in these two techniques that limit their application. Regarding subspace learning, the change from linearity to nonlinearity is usually achieved through kernelization, whereas extreme learning machines only take label information into consideration at the output layer. In order to overcome these drawbacks, this paper leverages extreme learning machines for dimensionality reduction and proposes a novel framework to combine spectral regression-based subspace learning and extreme learning machines. The proposed framework contains three stages-data mapping, graph decomposition, and regression. At the data mapping stage, various mapping strategies provide different views of the samples. At the graph decomposition stage, specifically designed embedding graphs provide a possibility to better represent the structure of data through generating virtual coordinates. Finally, at the regression stage, dimension-reduced mappings are achieved by connecting the virtual coordinates and data mapping. Using this framework, we propose several novel dimensionality reduction algorithms, apply them to SER tasks, and compare their performance to relevant state-of-the-art methods. Our results on several paralinguistic corpora show that our proposed techniques lead to significant improvements.