Edit model card

wav2vec2-large-xls-r-300m-sat-a3

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: 0.8961
  • Wer: 0.3976

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-a3 --dataset mozilla-foundation/common_voice_8_0 --config sat --split test --log_outputs

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

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: 200
  • num_epochs: 200
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
11.1266 33.29 100 2.8577 1.0
2.1549 66.57 200 1.0799 0.5542
0.5628 99.86 300 0.7973 0.4016
0.0779 133.29 400 0.8424 0.4177
0.0404 166.57 500 0.9048 0.4137
0.0212 199.86 600 0.8961 0.3976

Framework versions

  • Transformers 4.16.2
  • Pytorch 1.10.0+cu111
  • Datasets 1.18.3
  • Tokenizers 0.11.0
Downloads last month
11
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train DrishtiSharma/wav2vec2-large-xls-r-300m-sat-a3

Evaluation results