XLS-R-300M Uzbek CV8
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - UZ dataset. It achieves the following results on the validation set:
- Loss: 0.3063
- Wer: 0.3852
- Cer: 0.0777
Model description
For a description of the model architecture, see facebook/wav2vec2-xls-r-300m
The model vocabulary consists of the Modern Latin alphabet for Uzbek, with punctuation removed. Note that the characters <β> and <β> do not count as punctuation, as <β> modifies <o> and <g>, and <β> indicates the glottal stop or a long vowel.
The decoder uses a kenlm language model built on common_voice text.
Intended uses & limitations
This model is expected to be of some utility for low-fidelity use cases such as:
- Draft video captions
- Indexing of recorded broadcasts
The model is not reliable enough to use as a substitute for live captions for accessibility purposes, and it should not be used in a manner that would infringe the privacy of any of the contributors to the Common Voice dataset nor any other speakers.
Training and evaluation data
The 50% of the train
common voice official split was used as training data. The 50% of the official dev
split was used as validation data, and the full test
set was used for final evaluation of the model without LM, while the model with LM was evaluated only on 500 examples from the test
set.
The kenlm language model was compiled from the target sentences of the train + other dataset splits.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 3e-05
- train_batch_size: 32
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 100.0
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
---|---|---|---|---|---|
3.1401 | 3.25 | 500 | 3.1146 | 1.0 | 1.0 |
2.7484 | 6.49 | 1000 | 2.2842 | 1.0065 | 0.7069 |
1.0899 | 9.74 | 1500 | 0.5414 | 0.6125 | 0.1351 |
0.9465 | 12.99 | 2000 | 0.4566 | 0.5635 | 0.1223 |
0.8771 | 16.23 | 2500 | 0.4212 | 0.5366 | 0.1161 |
0.8346 | 19.48 | 3000 | 0.3994 | 0.5144 | 0.1102 |
0.8127 | 22.73 | 3500 | 0.3819 | 0.4944 | 0.1051 |
0.7833 | 25.97 | 4000 | 0.3705 | 0.4798 | 0.1011 |
0.7603 | 29.22 | 4500 | 0.3661 | 0.4704 | 0.0992 |
0.7424 | 32.47 | 5000 | 0.3529 | 0.4577 | 0.0957 |
0.7251 | 35.71 | 5500 | 0.3410 | 0.4473 | 0.0928 |
0.7106 | 38.96 | 6000 | 0.3401 | 0.4428 | 0.0919 |
0.7027 | 42.21 | 6500 | 0.3355 | 0.4353 | 0.0905 |
0.6927 | 45.45 | 7000 | 0.3308 | 0.4296 | 0.0885 |
0.6828 | 48.7 | 7500 | 0.3246 | 0.4204 | 0.0863 |
0.6706 | 51.95 | 8000 | 0.3250 | 0.4233 | 0.0868 |
0.6629 | 55.19 | 8500 | 0.3264 | 0.4159 | 0.0849 |
0.6556 | 58.44 | 9000 | 0.3213 | 0.4100 | 0.0835 |
0.6484 | 61.69 | 9500 | 0.3182 | 0.4124 | 0.0837 |
0.6407 | 64.93 | 10000 | 0.3171 | 0.4050 | 0.0825 |
0.6375 | 68.18 | 10500 | 0.3150 | 0.4039 | 0.0822 |
0.6363 | 71.43 | 11000 | 0.3129 | 0.3991 | 0.0810 |
0.6307 | 74.67 | 11500 | 0.3114 | 0.3986 | 0.0807 |
0.6232 | 77.92 | 12000 | 0.3103 | 0.3895 | 0.0790 |
0.6216 | 81.17 | 12500 | 0.3086 | 0.3891 | 0.0790 |
0.6174 | 84.41 | 13000 | 0.3082 | 0.3881 | 0.0785 |
0.6196 | 87.66 | 13500 | 0.3059 | 0.3875 | 0.0782 |
0.6174 | 90.91 | 14000 | 0.3084 | 0.3862 | 0.0780 |
0.6169 | 94.16 | 14500 | 0.3070 | 0.3860 | 0.0779 |
0.6166 | 97.4 | 15000 | 0.3066 | 0.3855 | 0.0778 |
Framework versions
- Transformers 4.16.2
- Pytorch 1.10.2+cu102
- Datasets 1.18.3
- Tokenizers 0.11.0
- Downloads last month
- 207
Model tree for lucio/xls-r-uzbek-cv8
Base model
facebook/wav2vec2-xls-r-300mDataset used to train lucio/xls-r-uzbek-cv8
Spaces using lucio/xls-r-uzbek-cv8 2
Evaluation results
- Test WER (with LM) on Common Voice 8self-reported15.065
- Test CER (with LM) on Common Voice 8self-reported3.077
- Test WER (no LM) on Common Voice 8self-reported32.880
- Test CER (no LM) on Common Voice 8self-reported6.530