wav2vec2-large-xls-r-300m-bas-v1

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

• Loss: 0.5997
• Wer: 0.3870

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-bas-v1 --dataset mozilla-foundation/common_voice_8_0 --config bas --split test --log_outputs

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

Basaa (bas) language isn't available in speech-recognition-community-v2/dev_data

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.000111
• 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: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
12.7076	5.26	200	3.6361	1.0
3.1657	10.52	400	3.0101	1.0
2.3987	15.78	600	0.9125	0.6774
1.0079	21.05	800	0.6477	0.5352
0.7392	26.31	1000	0.5432	0.4929
0.6114	31.57	1200	0.5498	0.4639
0.5222	36.83	1400	0.5220	0.4561
0.4648	42.1	1600	0.5586	0.4289
0.4103	47.36	1800	0.5337	0.4082
0.3692	52.62	2000	0.5421	0.3861
0.3403	57.88	2200	0.5549	0.4096
0.3011	63.16	2400	0.5833	0.3925
0.2932	68.42	2600	0.5674	0.3815
0.2696	73.68	2800	0.5734	0.3889
0.2496	78.94	3000	0.5968	0.3985
0.2289	84.21	3200	0.5888	0.3893
0.2091	89.47	3400	0.5849	0.3852
0.2005	94.73	3600	0.5938	0.3875
0.1876	99.99	3800	0.5997	0.3870

Framework versions

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