Lattice-Free MMI Adaptation Of Self-Supervised Pretrained Acoustic Models

In this work, we propose lattice-free MMI (LFMMI) for supervised adaptation\nof self-supervised pretrained acoustic model. We pretrain a Transformer model\non thousand hours of untranscribed Librispeech data followed by supervised\nadaptation with LFMMI on three different datasets. Our results show that\nfine-tuning with LFMMI, we consistently obtain relative WER improvements of 10%\nand 35.3% on the clean and other test sets of Librispeech (100h), 10.8% on\nSwitchboard (300h), and 4.3% on Swahili (38h) and 4.4% on Tagalog (84h)\ncompared to the baseline trained only with supervised data.\n