TY - CHAP A1 - Saha, Debjoy A1 - Nayak, Shravan A1 - Baumann, Timo T1 - Merkel Podcast Corpus: A Multimodal Dataset Compiled from 16 Years of Angela Merkel's Weekly Video Podcasts T2 - Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), Marseille, 20-25 June 2022 N2 - We introduce the Merkel Podcast Corpus, an audio-visual-text corpus in German collected from 16 years of (almost) weekly Internet podcasts of former German chancellor Angela Merkel. To the best of our knowledge, this is the first single speaker corpus in the German language consisting of audio, visual and text modalities of comparable size and temporal extent. We describe the methods used with which we have collected and edited the data which involves downloading the videos, transcripts and other metadata, forced alignment, performing active speaker recognition and face detection to finally curate the single speaker dataset consisting of utterances spoken by Angela Merkel. The proposed pipeline is general and can be used to curate other datasets of similar nature, such as talk show contents. Through various statistical analyses and applications of the dataset in talking face generation and TTS, we show the utility of the dataset. We argue that it is a valuable contribution to the research community, in particular, due to its realistic and challenging material at the boundary between prepared and spontaneous speech. Accepted at LREC 2022 KW - corpus KW - speaker diarization KW - multi-modal KW - forced alignment KW - single-speaker KW - cross-modal learning KW - German Y1 - 2022 UR - http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.270.pdf N1 - Preprint unter: https://arxiv.org/abs/2205.12194 SP - 2536 EP - 2540 ER - TY - CHAP A1 - Baumann, Timo A1 - Schlangen, David ED - Fingscheidt, T. T1 - The INPROTK 2012 release BT - A toolkit for incremental spoken dialogue processing T2 - Sprachkommunikation 2012 : Beiträge zur 10. ITG-Fachtagung vom 26. bis 28. September 2012 in Braunschweig N2 - We describe the 2012 release of INPROTK1, our “Incremental Processing Toolkit“ which combines a powerful and extensible architecture for incremental processing with components for incremental speech recognition and, new to this release, incremental speech synthesis. These components work domainindependently; we also provide example implementations of higher-level components such as natural language understanding and dialogue management that are somewhat more tied to a particular domain. The toolkit is accompanied by evaluation tools for analysing timing behaviour, and we highlight some timing results on conversational speech input in this paper. We offer our toolkit to foster research in this new and exciting area, which promises to help increase the naturalness of behaviours that can be modelled in such systems. KW - Speech synthesis Y1 - 2012 UR - https://ieeexplore.ieee.org/document/6309600 SN - 978-3-8007-3455-9 SP - 147 EP - 150 PB - VDE-Verl CY - Berlin ; Offenbach ER - TY - GEN A1 - Baumann, Timo A1 - Geiser, Dorothee A1 - Menzel, Wolfgang A1 - Mohr, Mario A1 - Neef, Svenja A1 - Nykamp, Sören A1 - Rokita, Nils T1 - Concurrent Sub-turn Interaction Specification and Dialogue Management with an Application to Interactive Storytelling T2 - GSCL Workshop Gesprochene Sprache und Sprachverarbeitung (GSS), Darmstadt, Germany, 2013 N2 - Conventional dialogue management centers around an interaction style that is best described as a ping-pong game, with full turns being the units at which speech is delivered (and expected) by the system, which greatly simplifies the interaction management, delivery and understanding components of the system. While the resulting mode of interaction works well for task-based systems, it is insufficient for more conversational interaction styles, where content is delivered and grounded in units finer than full turns (Poesio and Traum 1997) and where turns are delivered concurrently by both interlocutors and hence overlap more frequently. One domain with particularly frequent overlapping contributions is interactive storytelling: a storyteller that is responsive to listeners will integrate their feedback immediately while still speaking a current contribution, and listener’s remarks or propositions regarding the story will typically be uttered immediately when the related content is delivered. We present our work on a dialogue manager that leverages recent advances in incremental speech delivery and reception (Baumann 2013) to provide for an interactive and concurrent storytelling experience. Our system uses an interaction graph, which uses a word-by-word granularity and allows to specify for individual stretches of speech where (and with what content) users may interrupt/comment, how this is interpreted, and how it is integrated into the storytelling process, thus providing for in-utterance alternatives to be spoken even without audibly interrupting the system’s ongoing utterance. The system’s interaction graph is specified in an XML language and may be hand-crafted by a story designer but can also be automatically generated. In our current system, speech recognition results are interpreted only when the user utterance is finished; however, we plan to integrate incremental speech recognition and understanding capabilities, and to immediately react to the start of user contributions. Y1 - 2013 CY - Darmstadt, Germany ER - TY - CHAP A1 - Schenke, Diana Marie A1 - Baumann, Timo ED - Möller, Sebastian ED - Knoeferle, Pia ED - Schulte, Britta ED - Feldhus, Nils T1 - Controlled Diversity: Length-optimized Natural Language Generation T2 - Proceedings of the ISCA/ITG Workshop on Diversity in Large Speech and Language Models, February 20, 2025 in Berlin, Germany N2 - LLMs are not generally able to adjust the length of their outputs based on strict length requirements, a capability that would improve their usefulness in applications that require adherence to diverse user and system requirements. We present an approach to train LLMs to acquire this capability by augmenting existing data and applying existing fine-tuning techniques, which we compare based on the trained models’ adherence to the length requirement and overall response quality relative to the baseline model. Our results demonstrate that these techniques can be successfully applied to train LLMs to adhere to length requirements, with the trained models generating texts which better align to the length requirements. Our results indicate that our method may change the response quality when using training data that was not generated by the baseline model. This allows simultaneous alignment to another training objective in certain scenarios, but is undesirable otherwise. Training on a dataset containing the model’s own responses eliminates this issue. Y1 - 2025 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:898-opus4-80534 ER - TY - INPR A1 - Amanpour, Ali A1 - Baumann, Timo A1 - Hofmann, Ulrich G. T1 - Optimizing Neural Data Analysis: Determining Minimum Recording Length for Unambigous Signal Processing N2 - Advanced silicon electrode arrays facilitate the recording of thousands of neurons, generating extensive neural data that imposes a significant burden on researchers and processing algorithms. Thus, real-time analysis pipelines are gaining increasing value, while at the same time having to deal with non-stationary and noisy data. We intend to apply Machine Learning (ML) algorithms to a dense set of recordings from rat brains in order to prepare a functional atlas, correlating neuronal signals with anatomical position. While doing so, we needed to decide on a rational way which recording snippet length would best represent the original, longer source sequence and thus suffices to be further processed for anatomical correlation. We implemented an algorithm to evaluate the spectral information of systematically length varied records based on similarity to the original record. For our dataset a recording duration of 3 seconds satisfied moderate requirements across all channels, thus allowing us to reduce computational load for ongoing ML classification of microprobe sourced electrophysiologic signals. Y1 - 2025 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:898-opus4-80606 ER - TY - CHAP A1 - Morales, Juan A. A1 - Baumann, Timo A1 - Pérez, José L. A1 - Peinado, Antonio M. A1 - Gómez, Ángel M. T1 - Implementaciòn de un reconocedor distribuido de voz en tiempo real sobre IP T2 - Actas de las IV Jornadas en Tecnologías del Habla N2 - En este trabajo se presenta una implementación de unsistema de reconocimiento distribuido del habla en tiem-po real para su aplicación en un entorno de Internet. Desa-rrollado como una aplicación cliente-servidor, el clien-te hace uso del front-end estándar definido por la ETSI.Incluye un detector de voz para sólo enviar informacióncuando el locutor habla y la transmisión de la informaciónse hace de acuerdo al RFC 3557. El servidor realiza el re-conocimiento utilizando programación dinámica, incluyea su vez un detector de voz y utiliza la técnica de mitiga-ción de pérdida de paquetes propuesta en el estándar de laETSI. La modularidad del diseño permite la utilización decualquier otro reconocedor sin que por ello se vean afec-tados los clientes. Las pruebas reales de funcionamientopara la aplicación particular desarrollada han demostradouna alta fiabilidad en varias condiciones de transmisión. Y1 - 2006 UR - https://www.researchgate.net/publication/319702803_IMPLEMENTACION_DE_UN_RECONOCEDOR_DISTRIBUIDO_DE_VOZ_EN_TIEMPO_REAL_SOBRE_IP ER - TY - JOUR A1 - Nigel G. Ward, A1 - Vega, Alejandro A1 - Baumann, Timo T1 - Prosodic and Temporal Features for Language Modeling for Dialog JF - Speech Communication N2 - If we can model the cognitive and communicative processes underlying speech, we should be able to better predict what a speaker will do. With this idea as inspiration, we examine a number of prosodic and timing features as potential sources of information on what words the speaker is likely to say next. In spontaneous dialog we find that word probabilities do vary with such features. Using perplexity as the metric, the most informative of these included recent speaking rate, volume, and pitch, and time until end of utterance. Using simple combinations of such features to augment trigram language models gave up to a 8.4% perplexity benefit on the Switchboard corpus, and up to a 1.0% relative reduction in word error rate (0.3% absolute) on the Verbmobil II corpus. KW - Dialog dynamics KW - Dialog state KW - Prosody KW - Interlocutor behavior KW - Word probabilities KW - Prediction KW - Perplexity KW - Speech recognition KW - Switchboard corpus KW - Verbmobil corpus Y1 - 2012 U6 - https://doi.org/10.1016/j.specom.2011.07.009 VL - 54 IS - 2 SP - 161 EP - 174 PB - ELSEVIER ER - TY - CHAP A1 - Lohmann, Kris A1 - Eichhorn, Ole A1 - Baumann, Timo T1 - Generating Situated Assisting Utterances to Facilitate Tactile-Map Understanding: A Prototype System T2 - Proceedings of the Third Workshop on Speech and Language Processing for Assistive Technologies (SLPAT), Montreal, Canada, June 7–8, 2012 N2 - Tactile maps are important substitutes for visual maps for blind and visually impaired people and the efficiency of tactile-map reading can largely be improved by giving assisting utterances that make use of spatial language. In this paper, we elaborate earlier ideas for a system that generates such utterances and present a prototype implementation based on a semantic conceptualization of the movements that the map user performs. A worked example shows the plausibility of the solution and the output that the prototype generates given input derived from experimental data. Y1 - 2012 UR - https://aclanthology.org/W12-2908.pdf SP - 56 EP - 65 PB - Association for Computational Linguistics ER - TY - CHAP A1 - Baumann, Timo T1 - Feedback in Adaptive Interactive Storytelling T2 - 13th annual conference of the International Speech Communication Association 2012 (INTERSPEECH 2012): Portland, Oregon, USA, 9 - 13 September 2012, Interdisciplinary Workshop on Feedback Behaviors in Dialog N2 - Telling stories is different from reading out text: a speakerreponds to the listener’s feedback and incorporates thisinto the ongoing talk. However, current computer sys-tems are unable to do this and instead just non-attentivelyread out a text, disregarding all feedback (or the absencethereof). I propose and discuss an idea for a small researchproject and a plan for how an attentive listening storytellercan be built. KW - Incrementality KW - Feedback KW - Storytelling KW - Adaptation KW - Prosody Y1 - 2012 UR - https://www.researchgate.net/publication/236879509_Feedback_in_Adaptive_Interactive_Storytelling SN - 978-1-62276-759-5 CY - Stevenson, USA ER - TY - CHAP A1 - Baumann, Timo T1 - Integrating Prosodic Modelling with Incremental Speech Recognition T2 - Proceedings of DiaHolmia, the 13th International Workshop on the Semantics and Pragmatics of Dialogue (SEMDIAL 2009), 24-26 JUNE, 2009, Sttockhom, Sweden N2 - We describe ongoing and proposed work concerning incremental prosody extraction and classification for a spoken dialogue system. The system described will be tightly integrated with the SDS's speech recogntion which also works incrementally. The proposed architecture should allow for more control over the user interaction experience, for example allowing more precise and timely end-of-utterance vs. hesitation distinction, and auditive or visual back-channel generation. Y1 - 2009 UR - https://www.isca-speech.org/archive_v0/diaholmia_2009/dho9_143.html SP - 143 EP - 144 PB - Royal Institute of Technology ER - TY - THES A1 - Baumann, Timo T1 - Automatische Erkennung von Akzentuierungen und Phrasierungen in Sprachsynthesekorpora Y1 - 2007 UR - https://nats-www.informatik.uni-hamburg.de/pub/User/TimoBaumann/Publications/da.pdf N1 - Univ. Hamburg, Diplomarbeit CY - Hamburg ER - TY - CHAP A1 - Penzkofer, Vinzent A1 - Baumann, Timo T1 - Evaluating and Fine-Tuning Retrieval-Augmented Language Models to Generate Text With Accurate Citations T2 - Proceedings of the 20th Conference on Natural Language Processing (KONVENS 2024), September 10-13, 2024, Vienna, Austria N2 - Retrieval Augmented Generation (RAG) is becoming an essential tool for easily accessing large amounts of textual information. However, it is often challenging to determine whether the information in a given response originates from the retrieved context, the training, or is a result of hallucination. Our contribution in this area is twofold. Firstly, we demonstrate how existing datasets for information retrieval evaluation can be used to assess the ability of Large Language Models (LLMs) to correctly identify relevantsources. Our findings indicate that there are notable discrepancies in the performance of different current LLMs in this task. Secondly, we utilise the datasets and metrics for citation evaluation to enhance the citation quality of small open-weight LLMs through fine-tuning. We achieve significant performance gains in this task, matching the results of much larger models. Y1 - 2024 UR - https://aclanthology.org/2024.konvens-main.6.pdf SP - 57 EP - 64 PB - Association for Computational Linguistics ER - TY - CHAP A1 - Baumann, Timo A1 - Eller, Korbinian A1 - Gagarina, Natalia ED - Lal, Yash Kumar ED - Clark, Elizabeth ED - Iyyer, Mohit ED - Chaturvedi, Snigdha ED - Brei, Anneliese ED - Brahman, Faeze ED - Chandu, Khyathi Raghavi T1 - BERT-based Annotation of Oral Texts Elicited via Multilingual Assessment Instrument for Narratives T2 - Proceedings of the the 6th Workshop on Narrative Understanding (WNU), November 12th to November 16th 2024, Miami, Florida, USA N2 - We investigate how NLP can help annotate the structure and complexity of oral narrative texts elicited via the Multilingual Assessment Instrument for Narratives (MAIN). MAIN is a theory-based tool designed to evaluate the narrative abilities of children who are learning one or more languages from birth or early in their development. It provides a standardized way to measure how well children can comprehend and produce stories across different languages and referential norms for children between 3 and 12 years old. MAIN has been adapted to over ninety languages and is used in over 65 countries. The MAIN analysis focuses on story structure and story complexity which are typically evaluated manually based on scoring sheets. We here investigate the automation of this process using BERT-based classification which already yields promising results. Y1 - 2024 U6 - https://doi.org/10.18653/v1/2024.wnu-1.16 SP - 99 EP - 104 PB - Association for Computational Linguistics CY - Stroudsburg, PA, USA ER - TY - CHAP A1 - Mühlhausen, Sara A1 - Gomez, Sarah A1 - Lauer, Norina A1 - Baumann, Timo ED - Grawunder, Sven T1 - Cross lingual transfer learning does not improve aphasic speech recognition T2 - Elektronische Sprachsignalverarbeitung 2025: Tagungsband der 36. Konferenz Halle/Saale, 05.–07. MÄRZ 2025 N2 - In addressing the particular linguistic challenges posed by patients suffering from aphasia, a language disorder, this paper proposes a fine-tuning approach to enhance the speech recognition capabilities of existing models. The available aphasic research data in German is highly limited. To address this constraint, we propose a cross-lingual transfer approach to utilize English data to improve performance in German. This advancement aims to support the development of a therapy platform tailored for patients with aphasia. For the base speech recognition model, we choose to use OpenAI’s Whisper model, and for fine-tuning, we make use of TalkBank’s AphasiaBank. The experimental findings demonstrate that the transcription of aphasic audio with Whisper is less successful than non-aphasic audio. However, fine-tuning the transcription in the respective language resulted in an enhancement of its quality. In contrast, fine-tuning the transcription in another language and expecting a transfer of the learned aphasic speech properties led to a deterioration in its quality. KW - Automatische Spracherkennung KW - Sprachdialogsystem KW - Aphasie Y1 - 2025 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:898-opus4-80518 UR - https://www.essv.de/pdf/2025_77_84.pdf SN - 978-3-95908-803-9 SN - 0940-6832 PB - TUDpress CY - Dresden ER - TY - CHAP A1 - Baumann, Timo A1 - Hussein, Hussein A1 - Meyer-Sickendiek, Burkhard ED - Schneider, Birgit ED - Löffler, Beate ED - Mager, Tino ED - Hein, Carola T1 - Free Verse Prosodies: Identifying and Classifying Spoken Poetry Using Literary and Computational Perspectives (Rhythmicalizer) T2 - Mixing Methods: Practical Insights from the Humanities in the Digital Age N2 - At least 80% of modern and postmodern poems exhibit neither rhyme nor metrical schemes such as iamb or trochee. However, does this mean that they are free of any rhythmical features?TheUS American research onfree verse prosody claimsthe opposite: Modern poets like Whitman, the Imagists, the Beat poets and contemporary Slam poets have developed a postmetrical idea of prosody, using rhythmical features of everyday language, prose, and musical styles like Jazz or Hip Hop. It has spawned a large and complex variety intheir poetic prosodies which,however,appearto bemuchharderto quantify and regularize than traditional patterns. In our project, we examinethe largest portal for spoken poetry Lyrikline and analysed and classified such rhythmical patterns by using pattern recognition and classification techniques. We integrate a human-in-the-loop approach in which we interleave manual annotation with computational modelling and data-based analysis. Our results are integrated into the website of Lyrikline. Our follow-up project makes our research results available to a wider audience, in particular to high school-level teaching. Y1 - 2023 U6 - https://doi.org/10.1515/9783839469132-018 SP - 167 EP - 186 PB - Bielefeld University Press CY - Bielefeld ER - TY - CHAP A1 - Baumann, Timo A1 - Buß, Okko A1 - Atterer, Michaela A1 - Schlangen, David T1 - Evaluating the Potential Utility of ASR N-Best Lists for Incremental Spoken Dialogue Systems T2 - Proceedings of Interspeech 2009 : speech and intelligence ; 6 - 10 September, 2009, Brighton, UK N2 - The potential of using ASR n-best lists for dialogue systems has often been recognised (if less often realised): it is often the case that even when the top-ranked hypothesis is erroneous, a bet- ter one can be found at a lower rank. In this paper, we describe metrics for evaluating whether the same potential carries over to incremental dialogue systems, where ASR output is consumed and reacted upon while speech is still ongoing. We show that even small N can provide an advantage for semantic process- ing, at a cost of a computational overhead. KW - dialogue systems KW - speech recognition KW - naturallanguage understanding KW - incrementality Y1 - 2009 U6 - https://doi.org/10.21437/Interspeech.2009-318 SP - 1031 EP - 1034 PB - ISCA CY - Brighton, UK ER - TY - CHAP A1 - Baumann, Timo A1 - Atterer, Michaela A1 - Schlangen, David T1 - Assessing and Improving the Performance of Speech Recognition for Incremental Systems T2 - NAACL '09: Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, May 31 - June 5, 2009, Boulder, Colorado, USA N2 - In incremental spoken dialogue systems, par- tial hypotheses about what was said are re- quired even while the utterance is still ongo- ing. We define measures for evaluating the quality of incremental ASR components with respect to the relative correctness of the par- tial hypotheses compared to hypotheses that can optimize over the complete input, the tim- ingof hypothesisformationrelative to the por- tion ofthe inputthey areabout, andhypothesis stability, defined as the number of times they are revised. We show that simple incremen- tal post-processing can improve stability dra- matically, at the cost of timeliness (from 90% of edits of hypotheses being spurious down to 10% at a lag of 320ms). The measures are not independent,and we show how system de- signers can find a desired operating point for their ASR. To our knowledge, we are the first to suggest and examine a variety of measures for assessing incremental ASR and improve performance on this basis. Y1 - 2009 UR - https://dl.acm.org/doi/10.5555/1620754.1620810 SN - 978-1-932432-41-1 SP - 380 EP - 388 PB - Association for Computational Linguistics ER - TY - CHAP A1 - Baumann, Timo A1 - Buß, Okko A1 - Schlangen, David T1 - InproTK in Action: Open-Source Software for Building German-Speaking Incremental Spoken Dialogue Systems T2 - Electronic speech signal processing 2010 : proceedings of the 21st conference, Berlin, 8 - 10 September 2010 N2 - We present INPROTK, a toolkit for building incremental spoken dia-logue systems. Incremental spoken dialogue systems (systems that may react whilethe user’s utterance is ongoing) are a fairly recent research topic and allow for ex-citing new features. Even though toolkits exist that help in building conventionaldialogue systems, INPROTK offers both a tested architecture for building incre-mental SDSs as well as many of the building blocks necessary when building suchsystems. With INPROTK a researcher can avoid many of the technical difficulties,which hopefully further fosters research in this area. Y1 - 2010 UR - https://www.researchgate.net/publication/233540470_InproTK_in_Action_Open-Source_Software_for_Building_German-Speaking_Incremental_Spoken_Dialogue_Systems SP - 204 EP - 211 PB - TUDpress CY - Berlin, Germany ER - TY - CHAP A1 - Atterer, Michaela A1 - Baumann, Timo A1 - Schlangen, David T1 - No Sooner Said Than Done: Testing the Incrementality of Semantic Interpretations of Spontaneous Speech T2 - Proceedings of Interspeech 2009 : 6 - 10 September 2009, Brighton, U.K. N2 - Ideally, a spoken dialogue system should react without much delay to a user’s utterance. Such a system would already select an object, for instance, before the user has finished her utterance about moving this particular object to a particular place. A prerequisite for such a prompt reaction is that semantic representations are built up on the fly and passed on to other modules. Few approaches to incremental semantics construction exist, and, to our knowledge, none of those has been systematically tested on a spontaneous speech corpus. In this paper, we develop measures to test empirically on transcribed spontaneous speech to what extent we can create semantic interpretation on the fly with an incremental semantic chunker that builds a frame semantics. KW - incrementality KW - spoken dialogue systems KW - spontaneous speech KW - evaluation Y1 - 2009 U6 - https://doi.org/10.21437/Interspeech.2009-539 SP - 1855 EP - 1858 PB - International Speech Communication Association CY - Brighton, UK ER - TY - CHAP A1 - Baumann, Timo A1 - Schlangen, David T1 - Predicting the Micro-Timing of User Input for an Incremental Spoken Dialogue System that Completes a User’s Ongoing Turn T2 - SIGDIAL 2011 Conference, 12th annual meeting of the Special Interest Group on Discourse and Dialogue, Co-located with ACL HLT 2011, Portland, Oregon, 17 - 18 June 2011 N2 - We present the novel task of predicting tem-poral features of continuations of user input,while that input is still ongoing. We show that the remaining duration of an ongoing word, aswell as the duration of the next can be predicted reasonably well, and we put this information touse in a system that synchronously completesa user’s speech. While we focus on collaborative completions, the techniques presented here may also be useful for the alignment of back-channels and immediate turn-taking in anincremental SDS, or to synchronously monitorthe user’s speech fluency for other reasons. Y1 - 2011 UR - https://www.researchgate.net/publication/220794517_Predicting_the_Micro-Timing_of_User_Input_for_an_Incremental_Spoken_Dialogue_System_that_Completes_a_User%27s_Ongoing_Turn SN - 978-1-61839-242-8 SP - 120 EP - 129 PB - Curran CY - Red Hook, NY ER -