DrishtiSharma/wav2vec2-xls-r-300m-pa-IN-r5

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

• Loss: 0.8881
• Wer: 0.4175

Evaluation Commands

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

python eval.py --model_id DrishtiSharma/wav2vec2-xls-r-300m-pa-IN-r5 --dataset mozilla-foundation/common_voice_8_0 --config pa-IN --split test --log_outputs

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

Punjabi 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: 32
• 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: 2000
• num_epochs: 200.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
10.695	18.52	500	3.5681	1.0
3.2718	37.04	1000	2.3081	0.9643
0.8727	55.56	1500	0.7227	0.5147
0.3349	74.07	2000	0.7498	0.4959
0.2134	92.59	2500	0.7779	0.4720
0.1445	111.11	3000	0.8120	0.4594
0.1057	129.63	3500	0.8225	0.4610
0.0826	148.15	4000	0.8307	0.4351
0.0639	166.67	4500	0.8967	0.4316
0.0528	185.19	5000	0.8875	0.4238

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0