This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - DE dataset. It achieves the following results on the evaluation set:
- Loss: 0.1355
- Wer: 0.1532
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 7.5e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- 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: 2000
- num_epochs: 2.5
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
1.0826 | 0.07 | 1000 | 0.4637 | 0.4654 |
1.118 | 0.15 | 2000 | 0.2595 | 0.2687 |
1.1268 | 0.22 | 3000 | 0.2635 | 0.2661 |
1.0919 | 0.29 | 4000 | 0.2417 | 0.2566 |
1.1013 | 0.37 | 5000 | 0.2414 | 0.2567 |
1.0898 | 0.44 | 6000 | 0.2546 | 0.2731 |
1.0808 | 0.51 | 7000 | 0.2399 | 0.2535 |
1.0719 | 0.59 | 8000 | 0.2353 | 0.2528 |
1.0446 | 0.66 | 9000 | 0.2427 | 0.2545 |
1.0347 | 0.73 | 10000 | 0.2266 | 0.2402 |
1.0457 | 0.81 | 11000 | 0.2290 | 0.2448 |
1.0124 | 0.88 | 12000 | 0.2295 | 0.2448 |
1.025 | 0.95 | 13000 | 0.2138 | 0.2345 |
1.0107 | 1.03 | 14000 | 0.2108 | 0.2294 |
0.9758 | 1.1 | 15000 | 0.2019 | 0.2204 |
0.9547 | 1.17 | 16000 | 0.2000 | 0.2178 |
0.986 | 1.25 | 17000 | 0.2018 | 0.2200 |
0.9588 | 1.32 | 18000 | 0.1992 | 0.2138 |
0.9413 | 1.39 | 19000 | 0.1898 | 0.2049 |
0.9339 | 1.47 | 20000 | 0.1874 | 0.2056 |
0.9268 | 1.54 | 21000 | 0.1797 | 0.1976 |
0.9194 | 1.61 | 22000 | 0.1743 | 0.1905 |
0.8987 | 1.69 | 23000 | 0.1738 | 0.1932 |
0.8884 | 1.76 | 24000 | 0.1703 | 0.1873 |
0.8939 | 1.83 | 25000 | 0.1633 | 0.1831 |
0.8629 | 1.91 | 26000 | 0.1549 | 0.1750 |
0.8607 | 1.98 | 27000 | 0.1550 | 0.1738 |
0.8316 | 2.05 | 28000 | 0.1512 | 0.1709 |
0.8321 | 2.13 | 29000 | 0.1481 | 0.1657 |
0.825 | 2.2 | 30000 | 0.1446 | 0.1627 |
0.8115 | 2.27 | 31000 | 0.1396 | 0.1583 |
0.7959 | 2.35 | 32000 | 0.1389 | 0.1569 |
0.7835 | 2.42 | 33000 | 0.1362 | 0.1545 |
0.7959 | 2.49 | 34000 | 0.1355 | 0.1531 |
Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 1.18.2.dev0
- Tokenizers 0.11.0
Evaluation Commands
- To evaluate on
mozilla-foundation/common_voice_8_0
with splittest
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-1B-german --dataset mozilla-foundation/common_voice_8_0 --config de --split test --log_outputs
- To evaluate on test dev data
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-1B-german --dataset speech-recognition-community-v2/dev_data --config de --split validation --chunk_length_s 5.0 --stride_length_s 1.0
- Downloads last month
- 45
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 AndrewMcDowell/wav2vec2-xls-r-1B-german
Evaluation results
- Test WER on Common Voice 8self-reported15.250
- Test CER on Common Voice 8self-reported3.780
- Test WER on Robust Speech Event - Dev Dataself-reported35.290
- Test CER on Robust Speech Event - Dev Dataself-reported13.830
- Test WER on Robust Speech Event - Test Dataself-reported36.200