wav2vec2-large-xls-r-300m-or-d5
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - OR dataset. It achieves the following results on the evaluation set:
- Loss: 0.9571
- Wer: 0.5450
Evaluation Commands
- To evaluate on mozilla-foundation/common_voice_8_0 with test split
python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-or-d5 --dataset mozilla-foundation/common_voice_8_0 --config or --split test --log_outputs
- To evaluate on speech-recognition-community-v2/dev_data
python eval.py --model_id DrishtiSharma/wav2vec2-large-xls-r-300m-or-d5 --dataset speech-recognition-community-v2/dev_data --config or --split validation --chunk_length_s 10 --stride_length_s 1
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: 800
- num_epochs: 200
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
9.2958 | 12.5 | 300 | 4.9014 | 1.0 |
3.4065 | 25.0 | 600 | 3.5150 | 1.0 |
1.5402 | 37.5 | 900 | 0.8356 | 0.7249 |
0.6049 | 50.0 | 1200 | 0.7754 | 0.6349 |
0.4074 | 62.5 | 1500 | 0.7994 | 0.6217 |
0.3097 | 75.0 | 1800 | 0.8815 | 0.5985 |
0.2593 | 87.5 | 2100 | 0.8532 | 0.5754 |
0.2097 | 100.0 | 2400 | 0.9077 | 0.5648 |
0.1784 | 112.5 | 2700 | 0.9047 | 0.5668 |
0.1567 | 125.0 | 3000 | 0.9019 | 0.5728 |
0.1315 | 137.5 | 3300 | 0.9295 | 0.5827 |
0.1125 | 150.0 | 3600 | 0.9256 | 0.5681 |
0.1035 | 162.5 | 3900 | 0.9148 | 0.5496 |
0.0901 | 175.0 | 4200 | 0.9480 | 0.5483 |
0.0817 | 187.5 | 4500 | 0.9799 | 0.5516 |
0.079 | 200.0 | 4800 | 0.9571 | 0.5450 |
Framework versions
- Transformers 4.16.2
- Pytorch 1.10.0+cu111
- Datasets 1.18.3
- Tokenizers 0.11.0
- Downloads last month
- 5
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-or-d5
Evaluation results
- Test WER on Common Voice 8self-reported0.579
- Test CER on Common Voice 8self-reported0.157
- Test WER on Robust Speech Event - Dev Dataself-reportedNA
- Test CER on Robust Speech Event - Dev Dataself-reportedNA