Edit model card

xls-asr-vi-40h

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common voice 7.0 vi & private dataset. It achieves the following results on the evaluation set (Without Language Model):

  • Loss: 1.1177
  • Wer: 60.58

Evaluation

Please run the eval.py file

!python eval_custom.py --model_id geninhu/xls-asr-vi-40h --dataset mozilla-foundation/common_voice_7_0 --config vi --split test

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 5e-06
  • 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: 1500
  • num_epochs: 50.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
23.3878 0.93 1500 21.9179 1.0
8.8862 1.85 3000 6.0599 1.0
4.3701 2.78 4500 4.3837 1.0
4.113 3.7 6000 4.2698 0.9982
3.9666 4.63 7500 3.9726 0.9989
3.5965 5.56 9000 3.7124 0.9975
3.3944 6.48 10500 3.5005 1.0057
3.304 7.41 12000 3.3710 1.0043
3.2482 8.33 13500 3.4201 1.0155
3.212 9.26 15000 3.3732 1.0151
3.1778 10.19 16500 3.2763 1.0009
3.1027 11.11 18000 3.1943 1.0025
2.9905 12.04 19500 2.8082 0.9703
2.7095 12.96 21000 2.4993 0.9302
2.4862 13.89 22500 2.3072 0.9140
2.3271 14.81 24000 2.1398 0.8949
2.1968 15.74 25500 2.0594 0.8817
2.111 16.67 27000 1.9404 0.8630
2.0387 17.59 28500 1.8895 0.8497
1.9504 18.52 30000 1.7961 0.8315
1.9039 19.44 31500 1.7433 0.8213
1.8342 20.37 33000 1.6790 0.7994
1.7824 21.3 34500 1.6291 0.7825
1.7359 22.22 36000 1.5783 0.7706
1.7053 23.15 37500 1.5248 0.7492
1.6504 24.07 39000 1.4930 0.7406
1.6263 25.0 40500 1.4572 0.7348
1.5893 25.93 42000 1.4202 0.7161
1.5669 26.85 43500 1.3987 0.7143
1.5277 27.78 45000 1.3512 0.6991
1.501 28.7 46500 1.3320 0.6879
1.4781 29.63 48000 1.3112 0.6788
1.4477 30.56 49500 1.2850 0.6657
1.4483 31.48 51000 1.2813 0.6633
1.4065 32.41 52500 1.2475 0.6541
1.3779 33.33 54000 1.2244 0.6503
1.3788 34.26 55500 1.2116 0.6407
1.3428 35.19 57000 1.1938 0.6352
1.3453 36.11 58500 1.1927 0.6340
1.3137 37.04 60000 1.1699 0.6252
1.2984 37.96 61500 1.1666 0.6229
1.2927 38.89 63000 1.1585 0.6188
1.2919 39.81 64500 1.1618 0.6190
1.293 40.74 66000 1.1479 0.6181
1.2853 41.67 67500 1.1423 0.6202
1.2687 42.59 69000 1.1315 0.6131
1.2603 43.52 70500 1.1333 0.6128
1.2577 44.44 72000 1.1191 0.6079
1.2435 45.37 73500 1.1177 0.6079
1.251 46.3 75000 1.1211 0.6092
1.2482 47.22 76500 1.1177 0.6060
1.2422 48.15 78000 1.1227 0.6097
1.2485 49.07 79500 1.1187 0.6071
1.2425 50.0 81000 1.1177 0.6058

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.17.1.dev0
  • Tokenizers 0.11.0
Downloads last month
13
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 geninhu/xls-asr-vi-40h

Evaluation results

  • Test WER (with Language model) on Common Voice 7.0
    self-reported
    56.570