@techreport{oai:ipsj.ixsq.nii.ac.jp:00224441,
 author = {Raufun, Nahr and ino, Suzuki and Atsuhiko, Kai and Raufun, Nahr and ino, Suzuki and Atsuhiko, Kai},
 issue = {44},
 month = {Feb},
 note = {Automatic speech recognition (ASR) of real-world speech recorded in real environment has been a challenge in the field of artificial intelligence (AI). The real environment speech can vary in terms of location, recording medium and devices and so on. In this research, we particularly take interest in recognizing data recorded in university classroom. This real-world classroom situation is simulated by re-recording a small amount of data in classroom by playing through loudspeaker and recording them using low-quality wireless microphone. Previous research on supervised training of ASR indicates the requirement of large-scale transcribed data in target environment. However, it is costly to record and transcribe such amount of data for desired environment. Therefore, we adopt DNN-based data augmentation method for end-to-end ASR model as well as self-supervised-learning (SSL) based feature extraction with implicit end-to-end model to perform ASR task for classroom data. Fine-tuning of SSL-based ASR using target domain data helps achieving 17.9% character error rate for low audibility data., Automatic speech recognition (ASR) of real-world speech recorded in real environment has been a challenge in the field of artificial intelligence (AI). The real environment speech can vary in terms of location, recording medium and devices and so on. In this research, we particularly take interest in recognizing data recorded in university classroom. This real-world classroom situation is simulated by re-recording a small amount of data in classroom by playing through loudspeaker and recording them using low-quality wireless microphone. Previous research on supervised training of ASR indicates the requirement of large-scale transcribed data in target environment. However, it is costly to record and transcribe such amount of data for desired environment. Therefore, we adopt DNN-based data augmentation method for end-to-end ASR model as well as self-supervised-learning (SSL) based feature extraction with implicit end-to-end model to perform ASR task for classroom data. Fine-tuning of SSL-based ASR using target domain data helps achieving 17.9% character error rate for low audibility data.},
 title = {Domain Adaptation for Improving End-to-end ASR Performance of Classroom Speech with Variable Recording Condition},
 year = {2023}
}