### Describe the bug
Hi everyone,
I am training a speech recognition model us…ing SpeechBrain, but I noticed that while the loss is decreasing, my Word Error Rate (WER) remains high. Even after multiple epochs, it does not reduce below a certain threshold.
Training setup almost the same as example provided in recipes/CommonVoice/transformers/conformer-large.yaml except some data path and language specific parameters.
Training setup almost the same as example provided in recipes/CommonVoice/transformers/conformer-large.yaml:
• Model: transformers
• Dataset: CommonVoice corpus
Observations:
• Training and validation loss improve, but WER remains stuck at 100%.
• Increasing epochs does not help significantly.
• Model predictions contain a lot of substitutions/deletions/insertions.
• Some outputs are completely blank or garbage text.
Things I have tried:
✅ Adjusting learning rate
✅ Data augmentation (speed perturbation, noise addition)
✅ Checking label alignment in transcripts
✅ Changing beam search parameters in decoding
✅ Trying different architectures
Has anyone else faced this issue? Are there specific hyperparameters or decoding techniques that helped reduce WER in SpeechBrain?
I am using it for uzbek language of CommonVoice, which has about 400 hours of data.
Would appreciate any insights! Thanks.
### Expected behaviour
WER should lower as training
### To Reproduce
torchrun --standalone --nproc_per_node=6 python train.py hprams/conformer-large.yaml
### Environment Details
6 GPUS with 12GB RAM
### Relevant Log Output
```shell
Logs:
epoch: 1, lr: 8.32e-07, steps: 27 - train loss: 3.42e+02 - valid loss: 3.67e+02, valid ACC: 9.25e-05, valid WER: 1.62e+03, valid CER: 1.03e+03
epoch: 2, lr: 1.70e-06, steps: 54 - train loss: 2.82e+02 - valid loss: 2.51e+02, valid ACC: 1.05e-04, valid WER: 1.51e+03, valid CER: 1.05e+03
epoch: 3, lr: 2.56e-06, steps: 81 - train loss: 1.85e+02 - valid loss: 1.36e+02, valid ACC: 7.07e-02, valid WER: 2.65e+02, valid CER: 2.56e+02
epoch: 4, lr: 3.42e-06, steps: 108 - train loss: 1.17e+02 - valid loss: 1.13e+02, valid ACC: 7.07e-02, valid WER: 100.00, valid CER: 99.68
epoch: 5, lr: 4.29e-06, steps: 135 - train loss: 1.09e+02 - valid loss: 1.10e+02, valid ACC: 7.67e-02, valid WER: 1.00e+02, valid CER: 99.28
epoch: 6, lr: 5.15e-06, steps: 162 - train loss: 1.10e+02 - valid loss: 1.08e+02, valid ACC: 8.07e-02, valid WER: 1.00e+02, valid CER: 99.28
epoch: 7, lr: 6.02e-06, steps: 189 - train loss: 1.06e+02 - valid loss: 1.06e+02, valid ACC: 8.32e-02, valid WER: 1.00e+02, valid CER: 99.28
epoch: 8, lr: 6.88e-06, steps: 216 - train loss: 98.01 - valid loss: 1.05e+02, valid ACC: 8.38e-02, valid WER: 1.00e+02, valid CER: 99.28
epoch: 9, lr: 7.74e-06, steps: 243 - train loss: 1.09e+02 - valid loss: 1.04e+02, valid ACC: 8.42e-02, valid WER: 1.00e+02, valid CER: 99.28
epoch: 10, lr: 8.61e-06, steps: 270 - train loss: 1.02e+02 - valid loss: 1.03e+02, valid ACC: 8.62e-02, valid WER: 99.99, valid CER: 98.42
epoch: 11, lr: 9.47e-06, steps: 297 - train loss: 1.05e+02 - valid loss: 1.02e+02, valid ACC: 8.71e-02, valid WER: 100.00, valid CER: 97.57
epoch: 12, lr: 1.03e-05, steps: 324 - train loss: 96.42 - valid loss: 1.02e+02, valid ACC: 8.47e-02, valid WER: 100.00, valid CER: 97.19
epoch: 13, lr: 1.12e-05, steps: 351 - train loss: 96.79 - valid loss: 1.01e+02, valid ACC: 8.70e-02, valid WER: 100.00, valid CER: 97.20
epoch: 14, lr: 1.21e-05, steps: 378 - train loss: 96.04 - valid loss: 99.98, valid ACC: 9.07e-02, valid WER: 1.00e+02, valid CER: 97.51
```
### Additional Context
_No response_