Alice Baird

The INTERSPEECH 2016 Computational Paralinguistics Challenge: Deception, Sincerity & Native Language

The INTERSPEECH 2016 Computational Paralinguistics Challenge addresses three different problems for the first time in research competition under well-defined conditions: classification of deceptive vs. non-deceptive speech, the estimation of the degree of sincerity, and the identification of the native language out of 11 L1 classes of English L2 speakers. In this paper, we describe these sub-challenges, their conditions, and the baseline feature extraction and classifiers, as provided to the participants.

Active Learning for Bird Sounds Classification

It has been shown that automatic bird sound recognition can be an extremely useful tool for ornithologist and ecologists, allowing for a deeper understanding of; mating, evolution, local biodiversity and even climate change. For a robust and efficient recognition model, a large amount of labelled data is needed, requiring a time consuming and costly effort by expert-human annotators. To reduce this, we introduce for the first time, active learning, for automatic selection of the most informative data for training the recognition model. Experimental results show that our proposed; sparse-instance-based and least-confidence-score-based active learning methods reduce respectively 16.0% and 35.2% human annotated samples than compared to passive learning methods, achieving an acceptable performance (unweighted average recall > 85%), when recognising the sound of 60 different species of birds.

Perception of Paralinguistic Traits in Synthesized Voices

Along with the rise of artificial intelligence and the internet-of-things, synthesized voices are now common in daily--life, providing us with guidance, assistance, and even companionship. From formant to concatenative synthesis, the synthesized voice continues to be defined by the same traits we prescribe to ourselves. When the recorded voice is synthesized, does our perception of its new machine embodiment change, and can we consider an alternative, more inclusive form? To begin evaluating the impact of aesthetic design, this study presents a first--step perception test to explore the paralinguistic traits of the synthesized voice. Using a corpus of 13 synthesized voices, constructed from acoustic concatenative speech synthesis, we assessed the response of 23 listeners from differing cultural backgrounds. To evaluate if perception shifts from the defined traits, we asked listeners to assigned traits of age, gender, accent origin, and human--likeness. Results present a difference in perception for age and human--likeness across voices, and a general agreement across listeners for both gender and accent origin. Connections found between age, gender and human--likeness call for further exploration into a more participatory and inclusive synthesized vocal identity.

Active learning for bird sound classification via a kernel-based extreme learning machine

In recent years, research fields, including ecology, bioacoustics, signal processing, and machine learning, have made bird sound recognition a part of their focus. This has led to significant advancements within the field of ornithology, such as improved understanding of evolution, local biodiversity, mating rituals, and even the implications and realities associated to climate change. The volume of unlabeled bird sound data is now overwhelming, and comparatively little exploration is being made into methods for how best to handle them. In this study, two active learning (AL) methods are proposed, sparse-instance-based active learning (SI-AL), and least-confidence-score-based active learning (LCS-AL), both effectively reducing the need for expert human annotation. To both of these AL paradigms, a kernel-based extreme learning machine (KELM) is then integrated, and a comparison is made to the conventional support vector machine (SVM). Experimental results demonstrate that, when the classifier capacity is improved from an unweighted average recall of 60%-80%, KELM can outperform SVM even when a limited proportion of human annotations are used from the pool of data in both cases of SI-AL (minimum 34.5% vs minimum 59.0%) and LCS-AL (minimum 17.3% vs minimum 28.4%).

Speech analysis for health: Current state-of-the-art and the increasing impact of deep learning

Due to the complex and intricate nature associated with their production, the acoustic-prosodic properties of a speech signal are modulated with a range of health related effects. There is an active and growing area of machine learning research in this speech and health domain, focusing on developing paradigms to objectively extract and measure such effects. Concurrently, deep learning is transforming intelligent signal analysis, such that machines are now reaching near human capabilities in a range of recognition and analysis tasks. Herein, we review current state-of-the-art approaches with speech-based health detection, placing a particular focus on the impact of deep learning within this domain. Based on this overview, it is evident while that deep learning based solutions be become more present in the literature, it has not had the same overall dominating effect seen in other related fields. In this regard, we suggest some possible research directions aimed at fully leveraging the advantages that deep learning can offer speech-based health detection.