Elamur Research⌕
☰
home / papers / 2026 / Authorship Verification of Transcribed German-Language Videos

Authorship Verification of Transcribed German-Language Videos

Paper · 2026 · Oren Halvani and Sophie Titze

Summary

Authorship verification (AV) decides whether two texts come from the same person. Most AV work studies written English, so this paper tests it on spoken German instead. The authors collected single-speaker monologue videos from three domains: financial advice, maths tutoring and DIY. Each domain has 50 speakers with two videos each, for 300 videos and 150 speakers in total. They transcribed the audio with faster-whisper (Whisper large-v3) and cut each transcript to a central segment of about 5,000 characters. They also anonymised names, removed fillers, artifacts and recurring intro/outro phrases, and masked topical words with POSNoise so that only style remains. Ten AV methods from 2017–2026 were adapted to German and evaluated on author-disjoint splits. These include StyloSpeaker, the only one built for transcripts, and the current SOTA LambdaG. Traditional approaches built on character and token n-grams performed best, reaching up to 88% accuracy and 90% AUC. Transformer-based methods were clearly weaker on every corpus. The study shows that classic, transparent stylometry remains competitive for transcribed, non-English data.

Key points

  • Corpora:
    • Three domain corpora (finance, maths tutoring, DIY), each with 50 speakers × 2 videos of at least five minutes.
    • Speakers were split 40/60 between train and test with no overlap, giving 40 training and 60 test cases per domain, balanced between Y and N.
  • Pipeline:
    • Videos were downloaded with pytubefix and transcribed with faster-whisper large-v3, with German fixed as the language and VAD on.
    • Person names were found through self-introduction rules plus person-only NER; channel names came from video metadata.
  • Topic masking: POSNoise was chosen over TextDistortion because its part-of-speech-specific placeholders keep more stylistic detail than uniform asterisks.
  • Methods evaluated: COAV, SiamBERT, FeVecDiff, DV, CSS, MLSR, MStyleDistance, StyloSpeaker, RSP and LambdaG. Most needed adjustment for German and for the small training sets.
  • Results and outlook:
    • Simple n-gram methods led with up to 88% accuracy and 90% AUC; transformer models lagged on all three corpora.
    • The authors ask whether this ranking holds for typologically different languages, since the strongest methods depend on POS sequences and word order.

Interesting ideas

  • I1. Swap content words for placeholders tagged with their POS, so a model can only use function words, punctuation and word order. This exposes whether a system's decisions secretly depend on topic.
  • I2. Build negative pairs by shuffling authors into a circle and pairing each with its neighbour. This gives exactly one positive and one negative per author, keeps classes balanced and reuses every document without duplication.
  • I3. Split data by author rather than by document, and give the test side the larger share. This blocks identity leakage and makes the evaluation more trustworthy when data is scarce.
  • I4. Take the middle of each transcript and remove phrases that the same speaker repeats across videos. This deletes greetings and sign-offs that would otherwise act as shortcuts.
  • I5. Cheap n-gram methods beat transformer models under small, shifted data, which shows that simple transparent baselines must always be reported.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): run the evaluation suite on POSNoise-style masked inputs as a negative control, checking that structural decisions survive without lexical topic cues.
  • I2 (learn): use the same circular trick to make balanced contrastive pairs (same/different referent, same/different event) for coreference or frame-matching tests.
  • I3 (learn): split silver and gold data by source text or story, not by sentence, so held-out scores measure generalisation rather than memorised documents.
  • I4 (learn): when building training or test sets, find and strip recurring boilerplate (titles, standard openings of fables or school tasks) so learned rules cannot latch onto it.
  • I5 (tree): keep benchmarking the perceptron parser against gparse on small or out-of-domain treebanks, since the explainable model may win exactly where data is thin.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

Also on the byline: Fraunhofer Institute for Secure Information Technology.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

Cite

Cite as Oren Halvani and Sophie Titze. 2026. Authorship Verification of Transcribed German-Language Videos. arXiv:2607.29168.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2607.29168 · arXiv 2607.29168

@misc{halvani2026authorship,
  title = {{Authorship Verification of Transcribed German-Language Videos}},
  author = {Oren Halvani and Sophie Titze},
  year = {2026},
  doi = {10.48550/arxiv.2607.29168},
  eprint = {2607.29168},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2607.29168}
}

References

20 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. M. Koppel, J. Schler, S. Argamon, and Y. Winter, “The “Fundamental Problem” of Authorship Attribution,” English Studies, vol. 93, no. 3, pp. 284–291, 2012. find
  2. B. Stein, N. Lipka, and S. M. z. Eissen, “Meta Analysis within Authorship Verification,” in Proceedings of the 2008 19th International Conference on Database and Expert Systems Application, ser. DEXA ’08. USA: IEEE Computer Society, 2008, p. 34–39. find
  3. L. Lundmark, F. Johansson, B. Pelzer, L. Kaati, and J. Fernquist, “The applicability of authorship verification to swedish discussion forums,” in 2020 IEEE International Conference on Big Data (IEEE BigData 2020), Atlanta, GA, USA, December 10-13, 2020, X. Wu, C. Jermaine, L. Xiong, X. Hu, O. Kotevska, S. Lu, W. Xu, S. Aluru, C. Zhai, E. Al-Masri, Z. Chen, and J. Saltz, Eds. IEEE, 2020, pp. 1805–1812. find
  4. M. Stavngaard, A. Sørensen, S. Lorenzen, N. Hjuler, and S. Alstrup, “Detecting Ghostwriters in High Schools,” in 27th European Symposium on Artificial Neural Networks, ESANN 2019, Bruges, Belgium, April 24-26, 2019, 2019. ↗
  5. T. J. Neal, K. Sundararajan, and D. L. Woodard, “Exploiting Linguistic Style as a Cognitive Biometric for Continuous Verification,” in 2018 International Conference on Biometrics, ICB 2018, Gold Coast, Australia, February 20-23, 2018. IEEE, 2018, pp. 270–276. find
  6. C. Aggazzotti, N. Andrews, and E. A. Smith, “Can Authorship Attribution Models Distinguish Speakers in Speech Transcripts?” Transactions of the Association for Computational Linguistics, vol. 12, pp. 875–891, 07 2024. ↗
  7. C. Aggazzotti and E. A. Smith, “A Stylometric Analysis of Speaker Attribution from Speech Transcripts,” arXiv preprint arXiv:2512.13667, 2025. ↗
  8. C. Aggazzotti, M. Wiesner, E. A. Smith, and N. Andrews, “The Impact of Automatic Speech Transcription on Speaker Attribution,” Transactions of the Association for Computational Linguistics, vol. 13, pp. 1578–1596, 11 2025. ↗
  9. C. Cieri, D. Miller, and K. Walker, “The Fisher Corpus: a Resource for the Next Generations of Speech-to-Text,” in Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC’04), M. T. Lino, M. F. Xavier, F. Ferreira, R. Costa, and R. Silva, Eds. Lisbon, Portugal: European Language Resources Association (ELRA), May 2004.
  10. O. Halvani and L. Graner, “POSNoise: An Effective Countermeasure Against Topic Biases in Authorship Analysis,” in Proceedings of the 16th International Conference on Availability, Reliability and Security, ser. ARES ’21. New York, USA: Association for Computing Machinery, 2021. find
  11. E. Stamatatos, “Authorship Attribution Using Text Distortion,” in Proceedings of the 15th Conference of the European Chapter of the Association for the Computational Linguistics, EACL 2017, April 3-7, 2017, Valencia, Spain. The Association for Computer Linguistics, 2017. find
  12. O. Halvani, C. Winter, and L. Graner, “On the Usefulness of Compression Models for Authorship Verification,” in Proceedings of the 12th International Conference on Availability, Reliability and Security, ser. ARES ’17. New York, NY, USA: ACM, 2017, pp. 54:1–54:10. find
  13. A. Nini, O. Halvani, L. Graner, S. Titze, V. Gherardi, and S. Ishihara, “Grammar as a Behavioral Biometric: Using Cognitively Motivated Grammar Models for Authorship Verification,” 2024. [Online]. Available: https://arxiv.org/abs/2403.08462v1 ↗
  14. J. Weerasinghe, R. Singh, and R. Greenstadt, “Feature Vector Difference based Authorship Verification for Open-World Settings,” in Proceedings of the Working Notes of CLEF 2021 - Conference and Labs of the Evaluation Forum, Bucharest, Romania, September 21st - to - 24th, 2021, ser. CEUR Workshop Proceedings, G. Faggioli, N. Ferro, A. Joly, M. Maistro, and F. Piroi, Eds., vol. 2936. CEUR-WS.org, 2021, pp. 2201–2207. find
  15. S. Corbara, A. Moreo, and F. Sebastiani, “Same or Different? Diff-Vectors for Authorship Analysis,” ACM Trans. Knowl. Discov. Data, vol. 18, no. 1, Sep. 2023. ↗
  16. B. van Leeuwen, S. Bhulai, and R. van der Mei, “Combining Style and Semantics for Robust Authorship Verification,” Machine Learning with Applications, vol. 22, p. 100732, 2025. ↗
  17. J. Kim, H. Zhang, and D. Jurgens, “Leveraging Multilingual Training for Authorship Representation: Enhancing Generalization across Languages and Domains,” in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025. find
  18. J. Qiu, J. Zhu, A. Patel, M. Apidianaki, and C. Callison-Burch, “mStyleDistance: Multilingual Style Embeddings and their Evaluation,” in Findings of the Association for Computational Linguistics: ACL 2025, 2025, pp. 16 917–16 931. ↗
  19. P. Zeng, P. Alipoormolabashi, J. Mun, G. Dey, N. Soni, N. Balasubramanian, O. Rambow, and H. Schwartz, “Residualized Similarity for Faithfully Explainable Authorship Verification,” in Findings of the Association for Computational Linguistics: EMNLP 2025, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng, Eds. Suzhou, China: Association for Computational Linguistics, Nov. 2025, pp. 15 824– 15 837. find
  20. A. Nini, O. Halvani, L. Graner, S. Titze, V. Gherardi, and S. Ishihara, “Grammar as a Behavioral Biometric: Using Cognitively Motivated Grammar Models for Authorship Verification,” Humanities and Social Sciences Communications, vol. 13, no. 1, p. 455, 2026. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/halvani-2026-authorship-verification-transcribed-german-language/