YAML Metadata
Error:
"datasets[1]" must be a string
Serbian wav2vec2-xls-r-300m-sr-cv8
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: 1.7302
- Wer: 0.4825
- Cer: 0.1847
Evaluation on mozilla-foundation/common_voice_8_0 gave the following results:
- WER: 0.48530097993467103
- CER: 0.18413288165227845
Evaluation on speech-recognition-community-v2/dev_data gave the following results:
- WER: 0.9718373107518604
- CER: 0.8302740620263108
The model can be evaluated using the attached eval.py
script:
python eval.py --model_id comodoro/wav2vec2-xls-r-300m-sr-cv8 --dataset mozilla-foundation/common-voice_8_0 --split test --config sr
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 300
- num_epochs: 800
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
---|---|---|---|---|---|
5.6536 | 15.0 | 1200 | 2.9744 | 1.0 | 1.0 |
2.7935 | 30.0 | 2400 | 1.6613 | 0.8998 | 0.4670 |
1.6538 | 45.0 | 3600 | 0.9248 | 0.6918 | 0.2699 |
1.2446 | 60.0 | 4800 | 0.9151 | 0.6452 | 0.2398 |
1.0766 | 75.0 | 6000 | 0.9110 | 0.5995 | 0.2207 |
0.9548 | 90.0 | 7200 | 1.0273 | 0.5921 | 0.2149 |
0.8919 | 105.0 | 8400 | 0.9929 | 0.5646 | 0.2117 |
0.8185 | 120.0 | 9600 | 1.0850 | 0.5483 | 0.2069 |
0.7692 | 135.0 | 10800 | 1.1001 | 0.5394 | 0.2055 |
0.7249 | 150.0 | 12000 | 1.1018 | 0.5380 | 0.1958 |
0.6786 | 165.0 | 13200 | 1.1344 | 0.5114 | 0.1941 |
0.6432 | 180.0 | 14400 | 1.1516 | 0.5054 | 0.1905 |
0.6009 | 195.0 | 15600 | 1.3149 | 0.5324 | 0.1991 |
0.5773 | 210.0 | 16800 | 1.2468 | 0.5124 | 0.1903 |
0.559 | 225.0 | 18000 | 1.2186 | 0.4956 | 0.1922 |
0.5298 | 240.0 | 19200 | 1.4483 | 0.5333 | 0.2085 |
0.5136 | 255.0 | 20400 | 1.2871 | 0.4802 | 0.1846 |
0.4824 | 270.0 | 21600 | 1.2891 | 0.4974 | 0.1885 |
0.4669 | 285.0 | 22800 | 1.3283 | 0.4942 | 0.1878 |
0.4511 | 300.0 | 24000 | 1.4502 | 0.5002 | 0.1994 |
0.4337 | 315.0 | 25200 | 1.4714 | 0.5035 | 0.1911 |
0.4221 | 330.0 | 26400 | 1.4971 | 0.5124 | 0.1962 |
0.3994 | 345.0 | 27600 | 1.4473 | 0.5007 | 0.1920 |
0.3892 | 360.0 | 28800 | 1.3904 | 0.4937 | 0.1887 |
0.373 | 375.0 | 30000 | 1.4971 | 0.4946 | 0.1902 |
0.3657 | 390.0 | 31200 | 1.4208 | 0.4900 | 0.1821 |
0.3559 | 405.0 | 32400 | 1.4648 | 0.4895 | 0.1835 |
0.3476 | 420.0 | 33600 | 1.4848 | 0.4946 | 0.1829 |
0.3276 | 435.0 | 34800 | 1.5597 | 0.4979 | 0.1873 |
0.3193 | 450.0 | 36000 | 1.7329 | 0.5040 | 0.1980 |
0.3078 | 465.0 | 37200 | 1.6379 | 0.4937 | 0.1882 |
0.3058 | 480.0 | 38400 | 1.5878 | 0.4942 | 0.1921 |
0.2987 | 495.0 | 39600 | 1.5590 | 0.4811 | 0.1846 |
0.2931 | 510.0 | 40800 | 1.6001 | 0.4825 | 0.1849 |
0.276 | 525.0 | 42000 | 1.7388 | 0.4942 | 0.1918 |
0.2702 | 540.0 | 43200 | 1.7037 | 0.4839 | 0.1866 |
0.2619 | 555.0 | 44400 | 1.6704 | 0.4755 | 0.1840 |
0.262 | 570.0 | 45600 | 1.6042 | 0.4751 | 0.1865 |
0.2528 | 585.0 | 46800 | 1.6402 | 0.4821 | 0.1865 |
0.2442 | 600.0 | 48000 | 1.6693 | 0.4886 | 0.1862 |
0.244 | 615.0 | 49200 | 1.6203 | 0.4765 | 0.1792 |
0.2388 | 630.0 | 50400 | 1.6829 | 0.4830 | 0.1828 |
0.2362 | 645.0 | 51600 | 1.8100 | 0.4928 | 0.1888 |
0.2224 | 660.0 | 52800 | 1.7746 | 0.4932 | 0.1899 |
0.2218 | 675.0 | 54000 | 1.7752 | 0.4946 | 0.1901 |
0.2201 | 690.0 | 55200 | 1.6775 | 0.4788 | 0.1844 |
0.2147 | 705.0 | 56400 | 1.7085 | 0.4844 | 0.1851 |
0.2103 | 720.0 | 57600 | 1.7624 | 0.4848 | 0.1864 |
0.2101 | 735.0 | 58800 | 1.7213 | 0.4783 | 0.1835 |
0.1983 | 750.0 | 60000 | 1.7452 | 0.4848 | 0.1856 |
0.2015 | 765.0 | 61200 | 1.7525 | 0.4872 | 0.1869 |
0.1969 | 780.0 | 62400 | 1.7443 | 0.4844 | 0.1852 |
0.2043 | 795.0 | 63600 | 1.7302 | 0.4825 | 0.1847 |
Framework versions
- Transformers 4.16.2
- Pytorch 1.10.1+cu102
- Datasets 1.18.3
- Tokenizers 0.11.0
- Downloads last month
- 15
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.
Evaluation results
- Test WER on Common Voice 8.0self-reported48.530
- Test WER on Robust Speech Event - Dev Dataself-reported97.430
- Test WER on Robust Speech Event - Test Dataself-reported96.690