Externe Publikationen
Refine
Year of publication
- 2021 (7) (remove)
Document Type
- conference proceeding (article) (4)
- Article (2)
- Part of a Book (1)
Language
- English (7)
Has Fulltext
- no (7)
Is part of the Bibliography
- no (7)
Keywords
- Crowdsourcing (1)
- Found data (1)
- Likability ratings (1)
- North Sea (1)
- Ranking (1)
- Sequence modelling (1)
- Speech quality (1)
- Untertitelung (1)
- VoIP (1)
- ablation study (1)
Institute
- Fakultät Informatik und Mathematik (7) (remove)
Begutachtungsstatus
- peer-reviewed (5)
It remains difficult to segregate pelagic habitats since structuring processes are dynamic on a wide range of scales and clear boundaries in the open ocean are non-existent. However, to improve our knowledge about existing ecological niches and the processes shaping the enormous diversity of marine plankton, we need a better understanding of the driving forces behind plankton patchiness. Here we describe a new machine-learning method to detect and quantify pelagic habitats based on hydrographic measurements. An Autoencoder learns two-dimensional, meaningful representations of higher-dimensional micro-habitats, which are characterized by a variety of biotic and abiotic measurements from a high-speed ROTV. Subsequently, we apply a density-based clustering algorithm to group similar micro-habitats into associated pelagic macro-habitats in the German Bight of the North Sea. Three distinct macro-habitats, a “surface mixed layer,” a “bottom layer,” and an exceptionally “productive layer” are consistently identified, each with its distinct plankton community. We provide evidence that the model detects relevant features like the doming of the thermocline within an Offshore Wind Farm or the presence of a tidal mixing front.
Dubbing, i.e., the lip-synchronous translation and revoicing of audio-visual media into a target language from a different source language, is essential for the full-fledged reception of foreign audio-visual media, be it movies, instructional videos or short social media clips. In this paper, we objectify influences on the ‘dubbability’ of translations, i.e., how well a translation would be synchronously revoiceable to the lips on screen. We explore the value of traditional heuristics used in evaluating the qualitative aspects, in particular matching bilabial consonants and the jaw opening while producing vowels, and control for quantity, i.e., that translations are similar to the source in length. We perform an ablation study using an adversarial neural classifier which is trained to differentiate “true” dubbing translations from machine translations. While we are able to confirm the value of matching lip closure in dubbing, we find that the opening angle of the jaw as determined by the realized vowel may be less relevant than frequently considered in audio-visual translation.
In this paper, we present a study in which a robot initiates interactions with people passing by in an in-the-wild scenario. The robot adapts the loudness of its voice dynamically to the distance of the respective person approached, thus indicating who it is talking to. It furthermore tracks people based on information on body orientation and eye gaze and adapts the text produced based on people's distance autonomously. Our study shows that the adaptation of the loudness of its voice is perceived as personalization by the participants and that the likelihood that they stop by and interact with the robot increases when the robot incrementally adjusts its behavior.
We present an open source plugin for live subtitling in the popular open source video conferencing software BigBlueBut-ton. Our plugin decodes each speaker’s audio stream separately and in parallel, thereby obliviating the need for speaker di-arization and seamlessly handling overlapped talk. Any Kaldi-compatible nnet3 model can be used with our plugin and we demonstrate it using freely available TDNN-HMM-based ASR models for English and German. Our subtitles can be used as they are (e.g., in loud environments) or can form the basis for further NLP processes. Our tool can also simplify the collection of remotely recorded multi-party dialogue corpora.
We present a fully automatic solution for German video subtitling, with a focus on lecture videos. We rely entirely on open source models and scripts for German ASR, automatic punctuation reconstruction and subtitle segmentation. All training scripts, 1000h of German speech training data, pre-trained models and the final subtitling program are publicly available. It can readily be integrated into lecture video platforms such as Lecture2Go. The automatically generated subtitles can also serve as a basis to make the video material more accessible (e.g. via search, keyword clouds, and the like) or for further manual revision, potentially helping in significantly speeding up manual work. A particular challenge that we observe in lectures are technical terms that are frequent in a particular lecture, but infrequent in a typical language model and that might be out of vocabulary for a general purpose ASR. We approach this challenge by extracting texts from accompanying lecture slides to adapt the language model of our TDNN-HMM based ASR system. We demonstrate the usability of the full system and its generated subtitles and evaluate on a dataset of manually transcribed lectures with an average of 26.3% WER.
Speech quality and likability is a multi-faceted phenomenon consisting of a combination of perceptory features that cannot easily be computed nor weighed automatically. Yet, it is often easy to decide which of two voices one likes better, even though it would be hard to describe why, or to name the underlying basic perceptory features. Although likability is inherently subjective and individual preferences differ, generalizations are useful and there is often a broad intersubjective consensus about whether one speaker is more likeable than another. We present a methodology to efficiently create a likability ranking for many speakers from crowdsourced pairwise likability ratings which focuses manual rating effort on pairs of similar quality using an active sampling technique. Using this methodology, we collected pairwise likability ratings for many speakers (>220) from many raters (>160). We analyze listener preferences by correlating the resulting ranking with various acoustic and prosodic features. We also present a neural network that is able to model the complexity of listener preferences and the underlying temporal evolution of features. The recurrent neural network achieves remarkably high performance in estimating the pairwise decisions and an ablation study points toward the criticality of modeling temporal aspects in speech quality assessment.
The SMOOTH-robot is a mobile robot that-due to its modularity-combines a relatively low price with the possibility to be used for a large variety of tasks in a wide range of domains. In this article, we demonstrate the potential of the SMOOTH-robot through three use cases, two of which were performed in elderly care homes. The robot is designed so that it can either make itself ready or be quickly changed by staff to perform different tasks. We carefully considered important design parameters such as the appearance, intended and unintended interactions with users, and the technical complexity, in order to achieve high acceptability and a sufficient degree of utilization of the robot. Three demonstrated use cases indicate that such a robot could contribute to an improved work environment, having the potential to free resources of care staff which could be allocated to actual care-giving tasks. Moreover, the SMOOTH-robot can be used in many other domains, as we will also exemplify in this article.