wav2vec2-large-xls-r-300m-gn-k1

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

Loss: 0.9220
Wer: 0.6631

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_8_0 with test split

python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-gn-k1 --dataset mozilla-foundation/common_voice_8_0 --config gn --split test --log_outputs

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

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.00018
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: 600
num_epochs: 200
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
15.9402	8.32	100	6.9185	1.0
4.6367	16.64	200	3.7416	1.0
3.4337	24.96	300	3.2581	1.0
3.2307	33.32	400	2.8008	1.0
1.3182	41.64	500	0.8359	0.8171
0.409	49.96	600	0.8470	0.8323
0.2573	58.32	700	0.7823	0.7576
0.1969	66.64	800	0.8306	0.7424
0.1469	74.96	900	0.9225	0.7713
0.1172	83.32	1000	0.7903	0.6951
0.1017	91.64	1100	0.8519	0.6921
0.0851	99.96	1200	0.8129	0.6646
0.071	108.32	1300	0.8614	0.7043
0.061	116.64	1400	0.8414	0.6921
0.0552	124.96	1500	0.8649	0.6905
0.0465	133.32	1600	0.8575	0.6646
0.0381	141.64	1700	0.8802	0.6723
0.0338	149.96	1800	0.8731	0.6845
0.0306	158.32	1900	0.9003	0.6585
0.0236	166.64	2000	0.9408	0.6616
0.021	174.96	2100	0.9353	0.6723
0.0212	183.32	2200	0.9269	0.6570
0.0191	191.64	2300	0.9277	0.6662
0.0161	199.96	2400	0.9220	0.6631

Framework versions

Transformers 4.16.2
Pytorch 1.10.0+cu111
Datasets 1.18.3
Tokenizers 0.11.0