Date: Wednesday, 16th January 2019 at 2pm Place: LORIA, room A008 Speaker: Md Sahidullah (Inria – LORIA, Multispeech team)Title: Speaker embeddings: from i-vector to x-vector and beyondAbstract:
Speaker recognition is the task of recognizing a human from his/her voice. The state-of-the-art speaker recognition technology uses a speaker embedding method for representing a speech utterance of arbitrary length in the form of a fixed-dimensional vector. The recent advancements in deep neural network (DNN) research have enabled the development of robust and efficient speaker embedding techniques. In this talk, I will first provide a brief overview of speaker recognition basics. It will be followed by the description of the conventional speaker embedding method popularly known as i-vector. Then I will present various attempts to develop speech signal representations with DNN-based discriminative training. I will explain the recently introduced x-vector embedding which showed promising speaker recognition performance. This talk will end with a discussion on potential future directions in the speaker embedding research including our ongoing work.
Angeliki Monnier on Tuesday, May 28
Date: Tuesday, 28th May 2019 at 2pm Place: LORIA, room A008 Speaker: Angeliki Monnier (Université de Lorraine – CREM)
Title: Understanding ICT uses as “habitus”?
This presentation will address the uses of information and communication technologies (ICT), particularly social media, as “habitus”, inspired by Pierre Bourdieu’s Field Theory. My thesis draws on an empirical study, conducted on the professional network LinkedIn, which focused on discussion groups held by Greek and French expatriates, respectively.
More particularly, analysis has shown that discussions are not conducted in the same way between Greek and French expatriates. For the former, the question is the image of the country, tarnished by the economic crisis. For the latter, it is the direct contact with the Other, on the background of the French colonial past, which colors the discussions.
My argument is that the representations and “patterns of thought” that determine how Greek and French expatriates get involved in a discussion group within a professional social media are socio-historically determined and could be associated to distinct habitus. Following Pierre Bourdieu’s theory, these divergences could be related to the positions that Greece and France hold within the international migration field.
Title: Typometrics: from treebanks to typological universals
The emergence of treebanks of more than 70 languages in the same annotation scheme with the Universal Dependencies (UD) project puts empiricism in syntactic typology on a new footing. The analysis of a treebank not only confirms the existence or the absence of certain syntactic constructions on large quantities of natural text, it also makes it possible to measure their frequencies.
First, it becomes possible to compare the typological family trees of the language families in synchrony, based on the presence of certain typological features, with automatic hierarchizations, using for example a dendrogram of the frequency similarity of syntactic links. This compares the actual use of constructions in different languages.
Second, UD treebanks can be transformed to generalize qualitative syntactic universals to Greenberg: generally, typological universals declare or can be interpreted as the impossibility (or statistical rarity) of languages with certain properties. For example, Greenberg’s Universal 6, « All languages with dominant VSO order have SVO as an alternative or as the only alternative basic order. » excludes the existence of a language with only the VSO word order. I will show how a new form of visualization of word placement trends, in a scatterplot, allows an interpretation as quantitative universal — whose special case are qualitative universals.
The presentation will also show the necessary preparations for a typological interpretation of the UD annotations. In particular, treebank annotation and transformation tools will be presented. The study of word order requires a slightly different surface syntax annotation than the UD analysis. We will present an alternative annotation format, SUD (Surface-syntactic Universal Dependencies), on which our quantitative typology (or typometrics) studies are based, as well as the UD <-> SUD conversion tools.
Christine Evers on Thursday, October 31
Date: Thursday, 31st October 2019 at 10am Place: LORIA, room A008 Speaker: Christine Evers, Imperial College London
Title: Self-localization and mapping in dynamic, acoustic environments
Sound is used in nature to detect, identify and track salient events, to navigate, and to self-localise. The ability to make sense of acoustic signals is therefore a fundamental prerequisite for robots and autonomous systems. Audition – the ability to hear – excels particularly in scenarios where lighting conditions are poor, where salient events are outside of the line-of-sight, and in crowded environments where sensors, such as LIDAR and RADAR, are unreliable. However, due to the challenges affecting acoustic signals, robot audition has, to date, received only limited attention in the research community.
This talk focuses on acoustic scene mapping for robot audition in order to address the questions: “What is around me?” and “Where am I?”. The first part of the talk addresses the practical challenges of tracking moving sound sources when microphone arrays are integrated on moving platforms. The second part addresses the self-localization of the moving microphone array in the acoustic scene map when prior knowledge of the array position and orientation are unavailable. To conclude, extensions, including multi-modal sensor fusion, and future directions are discussed.