@inproceedings{SchenkeBaumann, author = {Schenke, Diana Marie and Baumann, Timo}, title = {Controlled Diversity: Length-optimized Natural Language Generation}, series = {Proceedings of the ISCA/ITG Workshop on Diversity in Large Speech and Language Models, February 20, 2025 in Berlin, Germany}, booktitle = {Proceedings of the ISCA/ITG Workshop on Diversity in Large Speech and Language Models, February 20, 2025 in Berlin, Germany}, editor = {M{\"o}ller, Sebastian and Knoeferle, Pia and Schulte, Britta and Feldhus, Nils}, doi = {10.48550/arXiv.2502.19347}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:898-opus4-80534}, pages = {8}, abstract = {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 Training on a dataset containing the model's own responses eliminates this issue.}, language = {en} } @unpublished{AmanpourBaumannHofmann, author = {Amanpour, Ali and Baumann, Timo and Hofmann, Ulrich G.}, title = {Optimizing Neural Data Analysis: Determining Minimum Recording Length for Unambigous Signal Processing}, doi = {10.35096/othr/pub-8060}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:898-opus4-80606}, pages = {5}, abstract = {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.}, language = {en} } @incollection{MoralesBaumannPerezetal., author = {Morales, Juan A. and Baumann, Timo and P{\´e}rez, Jos{\´e} L. and Peinado, Antonio M. and G{\´o}mez, {\´A}ngel M.}, title = {Implementaci{\`o}n de un reconocedor distribuido de voz en tiempo real sobre IP}, series = {Actas de las IV Jornadas en Tecnolog{\´i}as del Habla}, booktitle = {Actas de las IV Jornadas en Tecnolog{\´i}as del Habla}, abstract = {En este trabajo se presenta una implementaci{\´o}n de unsistema de reconocimiento distribuido del habla en tiem-po real para su aplicaci{\´o}n en un entorno de Internet. Desa-rrollado como una aplicaci{\´o}n cliente-servidor, el clien-te hace uso del front-end est{\´a}ndar definido por la ETSI.Incluye un detector de voz para s{\´o}lo enviar informaci{\´o}ncuando el locutor habla y la transmisi{\´o}n de la informaci{\´o}nse hace de acuerdo al RFC 3557. El servidor realiza el re-conocimiento utilizando programaci{\´o}n din{\´a}mica, incluyea su vez un detector de voz y utiliza la t{\´e}cnica de mitiga-ci{\´o}n de p{\´e}rdida de paquetes propuesta en el est{\´a}ndar de laETSI. La modularidad del dise{\~n}o permite la utilizaci{\´o}n decualquier otro reconocedor sin que por ello se vean afec-tados los clientes. Las pruebas reales de funcionamientopara la aplicaci{\´o}n particular desarrollada han demostradouna alta fiabilidad en varias condiciones de transmisi{\´o}n.}, language = {es} } @article{NigelGWardVegaBaumann, author = {Nigel G. Ward, and Vega, Alejandro and Baumann, Timo}, title = {Prosodic and Temporal Features for Language Modeling for Dialog}, series = {Speech Communication}, volume = {54}, journal = {Speech Communication}, number = {2}, publisher = {ELSEVIER}, doi = {10.1016/j.specom.2011.07.009}, pages = {161 -- 174}, abstract = {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.}, language = {en} } @incollection{LohmannEichhornBaumann, author = {Lohmann, Kris and Eichhorn, Ole and Baumann, Timo}, title = {Generating Situated Assisting Utterances to Facilitate Tactile-Map Understanding: A Prototype System}, series = {Proceedings of the Third Workshop on Speech and Language Processing for Assistive Technologies (SLPAT), Montreal, Canada, June 7-8, 2012}, booktitle = {Proceedings of the Third Workshop on Speech and Language Processing for Assistive Technologies (SLPAT), Montreal, Canada, June 7-8, 2012}, publisher = {Association for Computational Linguistics}, pages = {56 -- 65}, abstract = {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.}, language = {en} } @inproceedings{Baumann, author = {Baumann, Timo}, title = {Feedback in Adaptive Interactive Storytelling}, series = {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}, booktitle = {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}, address = {Stevenson, USA}, isbn = {978-1-62276-759-5}, abstract = {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.}, language = {en} } @inproceedings{Baumann, author = {Baumann, Timo}, title = {Integrating Prosodic Modelling with Incremental Speech Recognition}, series = {Proceedings of DiaHolmia, the 13th International Workshop on the Semantics and Pragmatics of Dialogue (SEMDIAL 2009), 24-26 JUNE, 2009, Sttockhom, Sweden}, booktitle = {Proceedings of DiaHolmia, the 13th International Workshop on the Semantics and Pragmatics of Dialogue (SEMDIAL 2009), 24-26 JUNE, 2009, Sttockhom, Sweden}, publisher = {Royal Institute of Technology}, pages = {143 -- 144}, abstract = {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.}, language = {en} } @masterthesis{Baumann, type = {Bachelor Thesis}, author = {Baumann, Timo}, title = {Automatische Erkennung von Akzentuierungen und Phrasierungen in Sprachsynthesekorpora}, address = {Hamburg}, language = {de} } @inproceedings{PenzkoferBaumann, author = {Penzkofer, Vinzent and Baumann, Timo}, title = {Evaluating and Fine-Tuning Retrieval-Augmented Language Models to Generate Text With Accurate Citations}, series = {Proceedings of the 20th Conference on Natural Language Processing (KONVENS 2024), September 10-13, 2024, Vienna, Austria}, booktitle = {Proceedings of the 20th Conference on Natural Language Processing (KONVENS 2024), September 10-13, 2024, Vienna, Austria}, publisher = {Association for Computational Linguistics}, organization = {Austrian Research Institute for Artificial Intelligence}, pages = {57 -- 64}, abstract = {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.}, language = {en} } @inproceedings{BaumannEllerGagarina, author = {Baumann, Timo and Eller, Korbinian and Gagarina, Natalia}, title = {BERT-based Annotation of Oral Texts Elicited via Multilingual Assessment Instrument for Narratives}, series = {Proceedings of the the 6th Workshop on Narrative Understanding (WNU), November 12th to November 16th 2024, Miami, Florida, USA}, booktitle = {Proceedings of the the 6th Workshop on Narrative Understanding (WNU), November 12th to November 16th 2024, Miami, Florida, USA}, editor = {Lal, Yash Kumar and Clark, Elizabeth and Iyyer, Mohit and Chaturvedi, Snigdha and Brei, Anneliese and Brahman, Faeze and Chandu, Khyathi Raghavi}, publisher = {Association for Computational Linguistics}, address = {Stroudsburg, PA, USA}, doi = {10.18653/v1/2024.wnu-1.16}, pages = {99 -- 104}, abstract = {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.}, language = {en} } @inproceedings{MuehlhausenGomezLaueretal., author = {M{\"u}hlhausen, Sara and Gomez, Sarah and Lauer, Norina and Baumann, Timo}, title = {Cross lingual transfer learning does not improve aphasic speech recognition}, series = {Elektronische Sprachsignalverarbeitung 2025: Tagungsband der 36. Konferenz Halle/Saale, 05.-07. M{\"A}RZ 2025}, booktitle = {Elektronische Sprachsignalverarbeitung 2025: Tagungsband der 36. Konferenz Halle/Saale, 05.-07. M{\"A}RZ 2025}, editor = {Grawunder, Sven}, publisher = {TUDpress}, address = {Dresden}, isbn = {978-3-95908-803-9}, issn = {0940-6832}, doi = {10.35096/othr/pub-8051}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:898-opus4-80518}, pages = {8}, abstract = {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.}, subject = {Automatische Spracherkennung}, language = {en} } @incollection{BaumannHusseinMeyerSickendiek, author = {Baumann, Timo and Hussein, Hussein and Meyer-Sickendiek, Burkhard}, title = {Free Verse Prosodies: Identifying and Classifying Spoken Poetry Using Literary and Computational Perspectives (Rhythmicalizer)}, series = {Mixing Methods: Practical Insights from the Humanities in the Digital Age}, booktitle = {Mixing Methods: Practical Insights from the Humanities in the Digital Age}, editor = {Schneider, Birgit and L{\"o}ffler, Beate and Mager, Tino and Hein, Carola}, publisher = {Bielefeld University Press}, address = {Bielefeld}, doi = {10.1515/9783839469132-018}, pages = {167 -- 186}, abstract = {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.}, language = {en} } @inproceedings{BaumannBussAttereretal., author = {Baumann, Timo and Buß, Okko and Atterer, Michaela and Schlangen, David}, title = {Evaluating the Potential Utility of ASR N-Best Lists for Incremental Spoken Dialogue Systems}, series = {Proceedings of Interspeech 2009 : speech and intelligence ; 6 - 10 September, 2009, Brighton, UK}, booktitle = {Proceedings of Interspeech 2009 : speech and intelligence ; 6 - 10 September, 2009, Brighton, UK}, publisher = {ISCA}, address = {Brighton, UK}, doi = {10.21437/Interspeech.2009-318}, pages = {1031 -- 1034}, abstract = {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.}, language = {en} }