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

Framework versions

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