Edit model card

wav2vec2-large-xls-r-300m-as-v9

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

  • Loss: 1.1679
  • Wer: 0.5761

Evaluation Command

  1. To evaluate on mozilla-foundation/common_voice_8_0 with test split

python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-as-v9 --dataset mozilla-foundation/common_voice_8_0 --config as --split test --log_outputs

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

Assamese (as) language isn't available in speech-recognition-community-v2/dev_data

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.000111
  • 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: 300
  • num_epochs: 200
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
8.3852 10.51 200 3.6402 1.0
3.5374 21.05 400 3.3894 1.0
2.8645 31.56 600 1.3143 0.8303
1.1784 42.1 800 0.9417 0.6661
0.7805 52.62 1000 0.9292 0.6237
0.5973 63.15 1200 0.9489 0.6014
0.4784 73.67 1400 0.9916 0.5962
0.4138 84.21 1600 1.0272 0.6121
0.3491 94.72 1800 1.0412 0.5984
0.3062 105.26 2000 1.0769 0.6005
0.2707 115.77 2200 1.0708 0.5752
0.2459 126.31 2400 1.1285 0.6009
0.2234 136.82 2600 1.1209 0.5949
0.2035 147.36 2800 1.1348 0.5842
0.1876 157.87 3000 1.1480 0.5872
0.1669 168.41 3200 1.1496 0.5838
0.1595 178.92 3400 1.1721 0.5778
0.1505 189.46 3600 1.1654 0.5744
0.1486 199.97 3800 1.1679 0.5761

Framework versions

  • Transformers 4.16.1
  • Pytorch 1.10.0+cu111
  • Datasets 1.18.2
  • Tokenizers 0.11.0
Downloads last month
17
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-as-v9

Evaluation results