Survey Talk
home
Survey Talk
Nancy F. Chen
A*STAR, Singapore
Title: Summarizing Conversations: From Meetings to Social Media Chats
Survey Talk 1 Monday 19 September, 13:10-13:40 KST at R116-118Abstract Human-to-human conversation is a dynamic process of information exchange between multiple parties that unfolds and evolves over time. It remains one of the most natural means of how humans pass down knowledge, increase mutual understanding, convey emotions, and collaborate with one another. However, it is nontrivial to automatically distill such unstructured information into summaries useful for downstream tasks. In this talk we will examine recent approaches and applications in addition to discussing future trends.
Esther Klabbers
ReadSpeaker, USA
Title: Speech Synthesis in Conversational AI
Survey Talk 2 Monday 19 September, 13:50-14:20 KST at R116-118Abstract Conversational AI is a type of artificial intelligence that enables consumers to interact with computer applications the way they would with other humans. State-of-the-art neural network speech synthesis is capable of generating expressive speech that sounds intelligible and natural, but there are several issues to consider when developing speech synthesis for conversational AI applications. In this survey talk we will give an overview of some of these considerations pertaining to training data selection, acoustic model architectures, and evaluation of the generated speech in the context of conversational AI.
Theodora Chaspari
Texas A&M University
Title: Emotion and Affective Speech Processing/Technology
Survey Talk 3 Wednesday 21 September, 12:10-12:40 KST at R116-118Abstract In the era of Internet of Things (IoT) and artificial intelligence (AI), voice-enabled technologies promise a unique opportunity for bettering our lives via human-centered services. Affective speech processing is central to such technologies, since speech can be a valuable indicator of emotion, communicative intent, and social behavior. However, the implementation and broad adoption of speech-enabled affective technologies comes with various computational and societal challenges, that yield from the complex and heterogeneous data spaces, limited availability of reliably labelled data, inherently high inter-individual variability, and the risk of leaking of personally identifiable information. This talk will overview the major challenges in designing robust speech emotion recognition models for supporting human-centered affective technologies, present recent advances on privacy preservation, fairness, and explainability, and discuss thoughts on rendering these technologies trustworthy to the stakeholders and the general public.
Nicolas Cummins
King's College London, Thymia
Title: Machine Learning for Speech-based Health Analysis: State-of-the-art and Future Challenges
Survey Talk 4 Wednesday 21 September, 12:50-13:20 KST at R116-118Abstract Speech is a unique and rich health signal: no other signal contains its singular combination of cognitive, neuromuscular and physiological information. This presentation will first describe how our voice is a tacit communicator of our health. Then, it will present an overview of current state-of-the-art in the prediction of health state using speech. It will end by highlighting future challenges in relation to the translation of speech analysis into clinic practise.
Leibny Paola Garcia Perera
Johns Hopkins University, Center for Language and Speech Processing, USA
Title: Current Trends in Diarization
Survey Talk 5 Thursday 22 September, 12:10-12:40 KST at R116-118Abstract The primary goal of speaker diarization is to answer the question “who spoke when” in a recording, identifying regions containing speech and assigning speaker identity labels to each utterance. In this talk, we will explore the state-of-the-art approaches to solve speaker diarization. We will highlight the evolution from cascaded systems to end-to-end diarization and the current improvements. Finally, we will point out which directions, such as multimodal diarization, are evolving rapidly.
Tom Bäckström
Aalto University, Department of Signal Processing and Acoustics
Title: Privacy in Speech Technology
Survey Talk 6 Thursday 22 September, 12:50-13:20 KST at R116-118Abstract Speech signals contain a host of information beyond the intended message, including speaker identity, state of health, affiliations and gender identity, much of which is private information. We have recently come to understand that this exposes users of speech technology to exploitation such as price gauging, discrimination, stalking and extortion. Consequently, research in the area has picked up speed and aims to improve the quality of experience and usability by preserving privacy. This talk introduces the main concepts and progress in the main research areas.
Sponsor
Supported
by
by
