wav2vec2-large-xls-r-300m-or-d5

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

• Loss: 0.9571
• Wer: 0.5450

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-or-d5 --dataset mozilla-foundation/common_voice_8_0 --config or --split test --log_outputs

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

python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-or-d5 --dataset speech-recognition-community-v2/dev_data --config or --split validation --chunk_length_s 10 --stride_length_s 1

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: 800
• num_epochs: 200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
9.2958	12.5	300	4.9014	1.0
3.4065	25.0	600	3.5150	1.0
1.5402	37.5	900	0.8356	0.7249
0.6049	50.0	1200	0.7754	0.6349
0.4074	62.5	1500	0.7994	0.6217
0.3097	75.0	1800	0.8815	0.5985
0.2593	87.5	2100	0.8532	0.5754
0.2097	100.0	2400	0.9077	0.5648
0.1784	112.5	2700	0.9047	0.5668
0.1567	125.0	3000	0.9019	0.5728
0.1315	137.5	3300	0.9295	0.5827
0.1125	150.0	3600	0.9256	0.5681
0.1035	162.5	3900	0.9148	0.5496
0.0901	175.0	4200	0.9480	0.5483
0.0817	187.5	4500	0.9799	0.5516
0.079	200.0	4800	0.9571	0.5450

Framework versions

• Transformers 4.16.2
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.11.0