wav2vec2-large-xls-r-300m-hsb-v2

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - HSB dataset. It achieves the following results on the evaluation set:

• Loss: 0.5328
• Wer: 0.4596

Evaluation Commands

1. To evaluate on mozilla-foundation/common_voice_8_0 with test split

python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-hsb-v2 --dataset mozilla-foundation/common_voice_8_0 --config hsb --split test --log_outputs

2. To evaluate on speech-recognition-community-v2/dev_data

Upper Sorbian (hsb) not found in speech-recognition-community-v2/dev_data

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.00045
• train_batch_size: 16
• 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: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
8.5979	3.23	100	3.5602	1.0
3.303	6.45	200	3.2238	1.0
3.2034	9.68	300	3.2002	0.9888
2.7986	12.9	400	1.2408	0.9210
1.3869	16.13	500	0.7973	0.7462
1.0228	19.35	600	0.6722	0.6788
0.8311	22.58	700	0.6100	0.6150
0.717	25.81	800	0.6236	0.6013
0.6264	29.03	900	0.6031	0.5575
0.5494	32.26	1000	0.5656	0.5309
0.4781	35.48	1100	0.5289	0.4996
0.4311	38.71	1200	0.5375	0.4768
0.3902	41.94	1300	0.5246	0.4703
0.3508	45.16	1400	0.5382	0.4696
0.3199	48.39	1500	0.5328	0.4596

Framework versions

• Transformers 4.16.1
• Pytorch 1.10.0+cu111
• Datasets 1.18.2
• Tokenizers 0.11.0