wav2vec2-large-xls-r-300m-hi-cv8-b2

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

• Loss: 0.7322
• Wer: 0.3469

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-hi-cv8-b2 --dataset mozilla-foundation/common_voice_8_0 --config hi --split test --log_outputs

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

Hindi language isn't available in speech-recognition-community-v2/dev_data

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.00025
• 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: 700
• num_epochs: 35
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
9.6226	1.04	200	3.8855	1.0
3.4678	2.07	400	3.4283	1.0
2.3668	3.11	600	1.0743	0.7175
0.7308	4.15	800	0.7663	0.5498
0.4985	5.18	1000	0.6957	0.5001
0.3817	6.22	1200	0.6932	0.4866
0.3281	7.25	1400	0.7034	0.4983
0.2752	8.29	1600	0.6588	0.4606
0.2475	9.33	1800	0.6514	0.4328
0.219	10.36	2000	0.6396	0.4176
0.2036	11.4	2200	0.6867	0.4162
0.1793	12.44	2400	0.6943	0.4196
0.1724	13.47	2600	0.6862	0.4260
0.1554	14.51	2800	0.7615	0.4222
0.151	15.54	3000	0.7058	0.4110
0.1335	16.58	3200	0.7172	0.3986
0.1326	17.62	3400	0.7182	0.3923
0.1225	18.65	3600	0.6995	0.3910
0.1146	19.69	3800	0.7075	0.3875
0.108	20.73	4000	0.7297	0.3858
0.1048	21.76	4200	0.7413	0.3850
0.0979	22.8	4400	0.7452	0.3793
0.0946	23.83	4600	0.7436	0.3759
0.0897	24.87	4800	0.7289	0.3754
0.0854	25.91	5000	0.7271	0.3667
0.0803	26.94	5200	0.7378	0.3656
0.0752	27.98	5400	0.7488	0.3680
0.0718	29.02	5600	0.7185	0.3619
0.0702	30.05	5800	0.7428	0.3554
0.0653	31.09	6000	0.7447	0.3559
0.0638	32.12	6200	0.7327	0.3523
0.058	33.16	6400	0.7339	0.3488
0.0594	34.2	6600	0.7322	0.3469

Framework versions

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