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

Loss: 0.8729
Wer: 0.4942

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-mr-v2 --dataset mozilla-foundation/common_voice_8_0 --config mr --split test --log_outputs

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

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

Note: Marathi language not found in speech-recognition-community-v2/dev_data!

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.000333
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: 1000
num_epochs: 200
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
8.4934	9.09	200	3.7326	1.0
3.4234	18.18	400	3.3383	0.9996
3.2628	27.27	600	2.7482	0.9992
1.7743	36.36	800	0.6755	0.6787
1.0346	45.45	1000	0.6067	0.6193
0.8137	54.55	1200	0.6228	0.5612
0.6637	63.64	1400	0.5976	0.5495
0.5563	72.73	1600	0.7009	0.5383
0.4844	81.82	1800	0.6662	0.5287
0.4057	90.91	2000	0.6911	0.5303
0.3582	100.0	2200	0.7207	0.5327
0.3163	109.09	2400	0.7107	0.5118
0.2761	118.18	2600	0.7538	0.5118
0.2415	127.27	2800	0.7850	0.5178
0.2127	136.36	3000	0.8016	0.5034
0.1873	145.45	3200	0.8302	0.5187
0.1723	154.55	3400	0.9085	0.5223
0.1498	163.64	3600	0.8396	0.5126
0.1425	172.73	3800	0.8776	0.5094
0.1258	181.82	4000	0.8651	0.5014
0.117	190.91	4200	0.8772	0.4970
0.1093	200.0	4400	0.8729	0.4942

Framework versions

Transformers 4.16.1
Pytorch 1.10.0+cu111
Datasets 1.18.2
Tokenizers 0.11.0