wav2vec2-large-xls-r-300m-sat-final

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

• Loss: 0.8012
• Wer: 0.3815

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-sat-final --dataset mozilla-foundation/common_voice_8_0 --config sat --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-sat-final --dataset speech-recognition-community-v2/dev_data --config sat --split validation --chunk_length_s 10 --stride_length_s 1

Note: Santali (Ol Chiki) language not found in speech-recognition-community-v2/dev_data

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0004
• 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: 170
• num_epochs: 200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
10.6317	33.29	100	2.8629	1.0
2.047	66.57	200	0.9516	0.5703
0.4475	99.86	300	0.8539	0.3896
0.0716	133.29	400	0.8277	0.3454
0.047	166.57	500	0.7597	0.3655
0.0249	199.86	600	0.8012	0.3815

Framework versions

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