This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - DE dataset. It achieves the following results on the evaluation set:

  • Loss: 0.1768
  • Wer: 0.2016

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: 3.4
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
5.7531 0.04 500 5.4564 1.0
2.9882 0.08 1000 3.0041 1.0
2.1953 0.13 1500 1.1723 0.7121
1.2406 0.17 2000 0.3656 0.3623
1.1294 0.21 2500 0.2843 0.2926
1.0731 0.25 3000 0.2554 0.2664
1.051 0.3 3500 0.2387 0.2535
1.0479 0.34 4000 0.2345 0.2512
1.0026 0.38 4500 0.2270 0.2452
0.9921 0.42 5000 0.2212 0.2353
0.9839 0.47 5500 0.2141 0.2330
0.9907 0.51 6000 0.2122 0.2334
0.9788 0.55 6500 0.2114 0.2270
0.9687 0.59 7000 0.2066 0.2323
0.9777 0.64 7500 0.2033 0.2237
0.9476 0.68 8000 0.2020 0.2194
0.9625 0.72 8500 0.1977 0.2191
0.9497 0.76 9000 0.1976 0.2175
0.9781 0.81 9500 0.1956 0.2159
0.9552 0.85 10000 0.1958 0.2191
0.9345 0.89 10500 0.1964 0.2158
0.9528 0.93 11000 0.1926 0.2154
0.9502 0.98 11500 0.1953 0.2149
0.9358 1.02 12000 0.1927 0.2167
0.941 1.06 12500 0.1901 0.2115
0.9287 1.1 13000 0.1936 0.2090
0.9491 1.15 13500 0.1900 0.2104
0.9478 1.19 14000 0.1931 0.2120
0.946 1.23 14500 0.1914 0.2134
0.9499 1.27 15000 0.1931 0.2173
0.9346 1.32 15500 0.1913 0.2105
0.9509 1.36 16000 0.1902 0.2137
0.9294 1.4 16500 0.1895 0.2086
0.9418 1.44 17000 0.1913 0.2183
0.9302 1.49 17500 0.1884 0.2114
0.9418 1.53 18000 0.1894 0.2108
0.9363 1.57 18500 0.1886 0.2132
0.9338 1.61 19000 0.1856 0.2078
0.9185 1.66 19500 0.1852 0.2056
0.9216 1.7 20000 0.1874 0.2095
0.9176 1.74 20500 0.1873 0.2078
0.9288 1.78 21000 0.1865 0.2097
0.9278 1.83 21500 0.1869 0.2100
0.9295 1.87 22000 0.1878 0.2095
0.9221 1.91 22500 0.1852 0.2121
0.924 1.95 23000 0.1855 0.2042
0.9104 2.0 23500 0.1858 0.2105
0.9284 2.04 24000 0.1850 0.2080
0.9162 2.08 24500 0.1839 0.2045
0.9111 2.12 25000 0.1838 0.2080
0.91 2.17 25500 0.1889 0.2106
0.9152 2.21 26000 0.1856 0.2026
0.9209 2.25 26500 0.1891 0.2133
0.9094 2.29 27000 0.1857 0.2089
0.9065 2.34 27500 0.1840 0.2052
0.9156 2.38 28000 0.1833 0.2062
0.8986 2.42 28500 0.1789 0.2001
0.9045 2.46 29000 0.1769 0.2022
0.9039 2.51 29500 0.1819 0.2073
0.9145 2.55 30000 0.1828 0.2063
0.9081 2.59 30500 0.1811 0.2049
0.9252 2.63 31000 0.1833 0.2086
0.8957 2.68 31500 0.1795 0.2083
0.891 2.72 32000 0.1809 0.2058
0.9023 2.76 32500 0.1812 0.2061
0.8918 2.8 33000 0.1775 0.1997
0.8852 2.85 33500 0.1790 0.1997
0.8928 2.89 34000 0.1767 0.2013
0.9079 2.93 34500 0.1735 0.1986
0.9032 2.97 35000 0.1793 0.2024
0.9018 3.02 35500 0.1778 0.2027
0.8846 3.06 36000 0.1776 0.2046
0.8848 3.1 36500 0.1812 0.2064
0.9062 3.14 37000 0.1800 0.2018
0.9011 3.19 37500 0.1783 0.2049
0.8996 3.23 38000 0.1810 0.2036
0.893 3.27 38500 0.1805 0.2056
0.897 3.31 39000 0.1773 0.2035
0.8992 3.36 39500 0.1804 0.2054
0.8987 3.4 40000 0.1768 0.2016

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.17.1.dev0
  • Tokenizers 0.11.0

Evaluation Commands

  1. To evaluate on mozilla-foundation/common_voice_7_0 with split test
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-300m-german-de --dataset mozilla-foundation/common_voice_7_0 --config de --split test --log_outputs
  1. To evaluate on test dev data
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-300m-german-de --dataset speech-recognition-community-v2/dev_data --config de --split validation --chunk_length_s 5.0 --stride_length_s 1.0
Downloads last month
40
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 AndrewMcDowell/wav2vec2-xls-r-300m-german-de

Evaluation results