Commit
•
8aaa6d9
1
Parent(s):
98fabba
Update README.md
Browse files
README.md
CHANGED
|
@@ -3,11 +3,9 @@ license: apache-2.0
|
|
| 3 |
base_model: facebook/wav2vec2-large-xlsr-53
|
| 4 |
metrics:
|
| 5 |
- wer
|
| 6 |
-
model-index:
|
| 7 |
-
- name: wav2vec2-xlsr-53-ft-ccv-en-cy
|
| 8 |
-
results: []
|
| 9 |
datasets:
|
| 10 |
- techiaith/commonvoice_16_1_en_cy
|
|
|
|
| 11 |
language:
|
| 12 |
- cy
|
| 13 |
- en
|
|
@@ -62,66 +60,3 @@ def transcribe(audio):
|
|
| 62 |
|
| 63 |
transcribe(<path/or/url/to/any/audiofile>)
|
| 64 |
```
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
## Evaluation
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
According to a balanced English+Welsh test set derived from Common Voice version 16.1, the WER of techiaith/wav2vec2-xlsr-53-ft-cy-en-withlm is **23.79%**
|
| 71 |
-
|
| 72 |
-
However, when evaluated with language specific test sets, the model exhibits a bias to perform better with Welsh.
|
| 73 |
-
|
| 74 |
-
| Common Voice Test Set Language | WER | CER |
|
| 75 |
-
| -------- | --- | --- |
|
| 76 |
-
| EN+CY | 23.79| 9.68 |
|
| 77 |
-
| EN | 34.47 | 14.83 |
|
| 78 |
-
| CY | 12.34 | 3.55 |
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
## Training procedure
|
| 82 |
-
|
| 83 |
-
### Training hyperparameters
|
| 84 |
-
|
| 85 |
-
The following hyperparameters were used during training:
|
| 86 |
-
- learning_rate: 0.0003
|
| 87 |
-
- train_batch_size: 32
|
| 88 |
-
- eval_batch_size: 32
|
| 89 |
-
- seed: 42
|
| 90 |
-
- gradient_accumulation_steps: 2
|
| 91 |
-
- total_train_batch_size: 64
|
| 92 |
-
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 93 |
-
- lr_scheduler_type: linear
|
| 94 |
-
- lr_scheduler_warmup_steps: 800
|
| 95 |
-
- training_steps: 9000
|
| 96 |
-
- mixed_precision_training: Native AMP
|
| 97 |
-
|
| 98 |
-
### Training results
|
| 99 |
-
|
| 100 |
-
| Training Loss | Epoch | Step | Validation Loss | Wer |
|
| 101 |
-
|:-------------:|:-----:|:----:|:---------------:|:------:|
|
| 102 |
-
| 6.0574 | 0.25 | 500 | 2.0297 | 0.9991 |
|
| 103 |
-
| 1.224 | 0.5 | 1000 | 0.5368 | 0.4342 |
|
| 104 |
-
| 0.434 | 0.75 | 1500 | 0.4861 | 0.3891 |
|
| 105 |
-
| 0.3295 | 1.01 | 2000 | 0.4301 | 0.3411 |
|
| 106 |
-
| 0.2739 | 1.26 | 2500 | 0.3818 | 0.3053 |
|
| 107 |
-
| 0.2619 | 1.51 | 3000 | 0.3894 | 0.3060 |
|
| 108 |
-
| 0.2517 | 1.76 | 3500 | 0.3497 | 0.2802 |
|
| 109 |
-
| 0.2244 | 2.01 | 4000 | 0.3519 | 0.2792 |
|
| 110 |
-
| 0.1854 | 2.26 | 4500 | 0.3376 | 0.2718 |
|
| 111 |
-
| 0.1779 | 2.51 | 5000 | 0.3206 | 0.2520 |
|
| 112 |
-
| 0.1749 | 2.77 | 5500 | 0.3169 | 0.2535 |
|
| 113 |
-
| 0.1636 | 3.02 | 6000 | 0.3122 | 0.2465 |
|
| 114 |
-
| 0.137 | 3.27 | 6500 | 0.3054 | 0.2382 |
|
| 115 |
-
| 0.1311 | 3.52 | 7000 | 0.2956 | 0.2280 |
|
| 116 |
-
| 0.1261 | 3.77 | 7500 | 0.2898 | 0.2236 |
|
| 117 |
-
| 0.1187 | 4.02 | 8000 | 0.2847 | 0.2176 |
|
| 118 |
-
| 0.1011 | 4.27 | 8500 | 0.2763 | 0.2124 |
|
| 119 |
-
| 0.0981 | 4.52 | 9000 | 0.2754 | 0.2115 |
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
### Framework versions
|
| 123 |
-
|
| 124 |
-
- Transformers 4.38.2
|
| 125 |
-
- Pytorch 2.2.1+cu121
|
| 126 |
-
- Datasets 2.18.0
|
| 127 |
-
- Tokenizers 0.15.2
|
|
|
|
| 3 |
base_model: facebook/wav2vec2-large-xlsr-53
|
| 4 |
metrics:
|
| 5 |
- wer
|
|
|
|
|
|
|
|
|
|
| 6 |
datasets:
|
| 7 |
- techiaith/commonvoice_16_1_en_cy
|
| 8 |
+
- techiaith/banc-trawsgrifiadau-bangor
|
| 9 |
language:
|
| 10 |
- cy
|
| 11 |
- en
|
|
|
|
| 60 |
|
| 61 |
transcribe(<path/or/url/to/any/audiofile>)
|
| 62 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|