The IACAS-Thinkit System for CHiME-7 Challenge

• DOI: 10.21437/chime.2023-5
• Publication Date: 2023-12-04T10:29:15Z
• Authors: Lingxuan Ye; Haitian Lu; Gaofeng Cheng; Yifan Chen; Zengqiang Shang; Xuyuan Li

This paper reports the IACAS-Thinkit's system for 7th CHiME challenge's task 1: distant automatic speech transcription and segmentation with multiple recording devices.Our includes training data augmentation, target speaker voice activity detection (TS-VAD) based diarization (SD), time-domain speakerbeam single channel extraction (TSE), guided source separation (GSS) multichannel WavLM recognition.Evaluated on CHiME-7 evaluation set, our main track achieves 25.0% macro-average Diarizationattributed Word Error Rate (DA-WER), an absolute reduction of 30.27% over baseline system; far-field acoustic robustness sub-track 20.5% macroaverage DA-WER, 13.75% system.