Login
- Original title:
- Exploring Hybrid CTC/Attention End-to-End Speech Recognition: Adversarial Robustness, Sinc Convolutions, and CTC Segmentation
- Translated title:
- Untersuchung hybrider CTC/Attention Ende-zu-Ende-Spracherkennung: Robustheit, Sinc-Faltungen und CTC-Segmentierung
- Author:
- Kürzinger, Ludwig
- Year:
- 2024
- Document type:
- Dissertation
- Faculty/School:
- TUM School of Computation, Information and Technology
- Institution:
- Mensch-Maschine-Kommunikation (Prof. Rigoll)
- Advisor:
- Rigoll, Gerhard (Prof. Dr.)
- Referee:
- Rigoll, Gerhard (Prof. Dr.); Ney, Hermann (Prof. Dr.)
- Language:
- en
- Subject group:
- DAT Datenverarbeitung, Informatik
- Keywords:
- End-to-End Speech Recognition, Neural Networks, Connectionist Temporal Classification (CTC), Attention Mechanisms, Sinc Convolutions, CTC Segmentation, Forced Alignment
- Translated keywords:
- Ende-zu-Ende Spracherkennung, Neuronale Netze, Connectionist Temporal Classification (CTC), Attention-Mechanismen, Sinc-Faltungen, CTC-Segmentierung, Forced Alignment
- TUM classification:
- DAT 815
- Abstract:
- This dissertation revisits the shift towards end-to-end speech recognition, addressing challenges in the hybrid CTC/attention framework such as parameter optimization, robustness against adversarial noise, integrated preprocessing, and increased data requirements. It incorporates Sinc convolutions, bypassing the need for separate feature extraction, and discusses CTC segmentation for improved speech data alignment, requiring minimal preprocessing, demonstrated with German and Japanese datasets.
- Translated abstract:
- Diese Dissertation untersucht den Übergang zu Ende-zu-Ende Spracherkennungssystemen am Beispiel der hybriden CTC/Attention-Architektur, hinsichtlich Parameteroptimierung, Robustheit gegenüber Adversarial Machine Learning, integrierter Merkmalsextraktion und dem hohen Datenbedarf. Darin werden lernbare Sinc-Faltungen als integraler Bestandteil des neuronalen Netzes diskutiert, sowie CTC-Segmentierung zur Sprachdatenaufbereitung, demonstriert an deutschen und japanischen Datensätzen.
- WWW:
- https://mediatum.ub.tum.de/?id=1726115
- Date of submission:
- 15.11.2023
- Oral examination:
- 24.05.2024
- File size:
- 4111285 bytes
- Pages:
- 165
- Urn (citeable URL):
- https://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:91-diss-20240524-1726115-1-9
- Last change:
- 12.07.2024
BibTeX