xls-r-asr_xh-run5

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the asr_xh dataset. It achieves the following results:

• Wer (Validation): 54.36%
• Wer (Test): 54.94%

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer (Train)
14.5246	0.64	50	8.0044	1.0
5.7225	1.27	100	4.2174	1.0
3.7885	1.91	150	3.2739	1.0
3.1939	2.55	200	3.0574	1.0
3.0792	3.18	250	3.0012	1.0
2.8734	3.82	300	2.4681	1.0
1.735	4.46	350	0.9881	0.9242
0.933	5.1	400	0.6460	0.7483
0.6435	5.73	450	0.6392	0.7668
0.5548	6.37	500	0.5045	0.6725
0.4812	7.01	550	0.5334	0.6667
0.3797	7.64	600	0.4583	0.6135
0.3542	8.28	650	0.4543	0.5811
0.2959	8.92	700	0.5105	0.6466
0.2686	9.55	750	0.4518	0.5948
0.22	10.19	800	0.4408	0.5870
0.2155	10.83	850	0.4228	0.5513
0.1755	11.46	900	0.4862	0.5708
0.1777	12.1	950	0.4304	0.5212
0.1627	12.74	1000	0.4358	0.5445
0.1444	13.38	1050	0.4418	0.5157
0.1364	14.01	1100	0.4711	0.5348
0.1213	14.65	1150	0.4952	0.5546

Framework versions

• Transformers 4.28.0
• Pytorch 2.0.1+cu118
• Datasets 2.14.4
• Tokenizers 0.13.3