patrickvonplaten/wav2vec2-large-lv60h-100h-2nd-try

The model was trained on librispeech-clean-train.100 with following hyper-parameters: "Clean" and "Other" The model is trained on two GPUs: Titan RTX and 16 GPUs with a total batch size of ca.750 seconds with linear decaying learning rate with 3000 warmup steps . The results are a bit worse than those reported in the Appendix in Table 3 of the original paper . Check: https://://wandb.ai/patrickvonplaten/huggingface/reports/Project-Dashboard--Vmllddzo0OTI0OTc?accessToken=8azw8iyxnbiqd4y,