wav2vec2-large-xls-r-300m-hsb-v3

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: 0.6549
• Wer: 0.4827

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-v3 --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) language 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
8.8951	3.23	100	3.6396	1.0
3.314	6.45	200	3.2331	1.0
3.1931	9.68	300	3.0947	0.9906
1.7079	12.9	400	0.8865	0.8499
0.6859	16.13	500	0.7994	0.7529
0.4804	19.35	600	0.7783	0.7069
0.3506	22.58	700	0.6904	0.6321
0.2695	25.81	800	0.6519	0.5926
0.222	29.03	900	0.7041	0.5720
0.1828	32.26	1000	0.6608	0.5513
0.1474	35.48	1100	0.7129	0.5319
0.1269	38.71	1200	0.6664	0.5056
0.1077	41.94	1300	0.6712	0.4942
0.0934	45.16	1400	0.6467	0.4879
0.0819	48.39	1500	0.6549	0.4827

Framework versions

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