Update README.md
Browse files
README.md
CHANGED
|
@@ -1,132 +1,180 @@
|
|
| 1 |
---
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
tags:
|
| 4 |
-
- automatic-speech-recognition
|
| 5 |
-
-
|
| 6 |
-
-
|
| 7 |
datasets:
|
| 8 |
-
-
|
| 9 |
model-index:
|
| 10 |
-
- name:
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
---
|
| 13 |
|
| 14 |
-
|
| 15 |
-
should probably proofread and complete it, then remove this comment. -->
|
| 16 |
|
| 17 |
-
|
| 18 |
|
| 19 |
-
This model
|
| 20 |
-
It achieves the following results on the evaluation set:
|
| 21 |
-
- Loss: 0.2089
|
| 22 |
-
- Wer: 0.2954
|
| 23 |
-
- Cer: 0.0953
|
| 24 |
|
| 25 |
-
|
| 26 |
|
| 27 |
-
|
| 28 |
|
| 29 |
-
##
|
| 30 |
|
| 31 |
-
|
|
|
|
|
|
|
| 32 |
|
| 33 |
-
##
|
| 34 |
|
| 35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
## Training procedure
|
| 38 |
|
|
|
|
|
|
|
| 39 |
### Training hyperparameters
|
| 40 |
|
| 41 |
The following hyperparameters were used during training:
|
| 42 |
-
|
| 43 |
-
-
|
| 44 |
-
-
|
| 45 |
-
-
|
| 46 |
-
-
|
| 47 |
-
-
|
| 48 |
-
-
|
| 49 |
-
-
|
| 50 |
-
-
|
| 51 |
-
-
|
| 52 |
-
-
|
|
|
|
| 53 |
|
| 54 |
### Training results
|
| 55 |
|
| 56 |
-
| Training Loss | Epoch | Step | Validation Loss |
|
| 57 |
-
|
| 58 |
-
|
|
| 59 |
-
|
|
| 60 |
-
|
|
| 61 |
-
|
|
| 62 |
-
|
|
| 63 |
-
|
|
| 64 |
-
|
|
| 65 |
-
|
|
| 66 |
-
|
|
| 67 |
-
|
|
| 68 |
-
|
|
| 69 |
-
|
|
| 70 |
-
|
|
| 71 |
-
|
|
| 72 |
-
|
|
| 73 |
-
|
|
| 74 |
-
|
|
| 75 |
-
|
|
| 76 |
-
|
|
| 77 |
-
|
|
| 78 |
-
|
|
| 79 |
-
|
|
| 80 |
-
|
|
| 81 |
-
|
|
| 82 |
-
|
|
| 83 |
-
|
|
| 84 |
-
|
|
| 85 |
-
|
|
| 86 |
-
|
|
| 87 |
-
|
|
| 88 |
-
|
|
| 89 |
-
|
|
| 90 |
-
|
|
| 91 |
-
|
|
| 92 |
-
|
|
| 93 |
-
|
|
| 94 |
-
|
|
| 95 |
-
|
|
| 96 |
-
|
|
| 97 |
-
|
|
| 98 |
-
|
|
| 99 |
-
|
|
| 100 |
-
|
|
| 101 |
-
|
|
| 102 |
-
|
|
| 103 |
-
|
|
| 104 |
-
|
|
| 105 |
-
|
|
| 106 |
-
|
|
| 107 |
-
|
|
| 108 |
-
|
|
| 109 |
-
|
|
| 110 |
-
|
|
| 111 |
-
|
|
| 112 |
-
|
|
| 113 |
-
|
|
| 114 |
-
|
|
| 115 |
-
|
|
| 116 |
-
|
|
| 117 |
-
|
|
| 118 |
-
|
|
| 119 |
-
|
|
| 120 |
-
|
|
| 121 |
-
|
|
| 122 |
-
|
|
| 123 |
-
|
|
| 124 |
-
|
|
| 125 |
-
|
|
| 126 |
-
|
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
|
| 131 |
- Transformers 4.17.0.dev0
|
| 132 |
- Pytorch 1.10.2+cu102
|
|
|
|
| 1 |
---
|
| 2 |
+
language: ko
|
| 3 |
license: apache-2.0
|
| 4 |
tags:
|
| 5 |
+
- automatic-speech-recognition
|
| 6 |
+
- generated_from_trainer
|
| 7 |
+
- robust-speech-event
|
| 8 |
datasets:
|
| 9 |
+
- kresnik/zeroth_korean
|
| 10 |
model-index:
|
| 11 |
+
- name: Wav2Vec2 XLS-R 300M Korean LM
|
| 12 |
+
results:
|
| 13 |
+
- task:
|
| 14 |
+
name: Automatic Speech Recognition
|
| 15 |
+
type: automatic-speech-recognition
|
| 16 |
+
dataset:
|
| 17 |
+
name: Zeroth Korean
|
| 18 |
+
type: kresnik/zeroth_korean
|
| 19 |
+
args: clean
|
| 20 |
+
metrics:
|
| 21 |
+
- name: Test WER
|
| 22 |
+
type: wer
|
| 23 |
+
value: 30.94
|
| 24 |
+
- name: Test CER
|
| 25 |
+
type: cer
|
| 26 |
+
value: 7.97
|
| 27 |
+
- task:
|
| 28 |
+
name: Automatic Speech Recognition
|
| 29 |
+
type: automatic-speech-recognition
|
| 30 |
+
dataset:
|
| 31 |
+
name: Robust Speech Event - Dev Data
|
| 32 |
+
type: speech-recognition-community-v2/dev_data
|
| 33 |
+
args: ko
|
| 34 |
+
metrics:
|
| 35 |
+
- name: Test WER
|
| 36 |
+
type: wer
|
| 37 |
+
value: 68.34
|
| 38 |
+
- name: Test CER
|
| 39 |
+
type: cer
|
| 40 |
+
value: 37.08
|
| 41 |
---
|
| 42 |
|
| 43 |
+
# Wav2Vec2 XLS-R 300M Korean LM
|
|
|
|
| 44 |
|
| 45 |
+
Wav2Vec2 XLS-R 300M Korean LM is an automatic speech recognition model based on the [XLS-R](https://arxiv.org/abs/2111.09296) architecture. This model is a fine-tuned version of [Wav2Vec2-XLS-R-300M](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the [Zeroth Korean](https://huggingface.co/datasets/kresnik/zeroth_korean) dataset. A 5-gram Language model, trained on the Korean subset of [Open Subtitles](https://huggingface.co/datasets/open_subtitles), was then subsequently added to this model.
|
| 46 |
|
| 47 |
+
This model was trained using HuggingFace's PyTorch framework and is part of the [Robust Speech Challenge Event](https://discuss.huggingface.co/t/open-to-the-community-robust-speech-recognition-challenge/13614) organized by HuggingFace. All training was done on a Tesla V100, sponsored by OVH.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
+
All necessary scripts used for training could be found in the [Files and versions](https://huggingface.co/w11wo/wav2vec2-xls-r-300m-korean-lm/tree/main) tab, as well as the [Training metrics](https://huggingface.co/w11wo/wav2vec2-xls-r-300m-korean-lm/tensorboard) logged via Tensorboard.
|
| 50 |
|
| 51 |
+
As for the N-gram language model training, we followed the [blog post tutorial](https://huggingface.co/blog/wav2vec2-with-ngram) provided by HuggingFace.
|
| 52 |
|
| 53 |
+
## Model
|
| 54 |
|
| 55 |
+
| Model | #params | Arch. | Training/Validation data (text) |
|
| 56 |
+
| ------------------------------- | ------- | ----- | ------------------------------- |
|
| 57 |
+
| `wav2vec2-xls-r-300m-korean-lm` | 300M | XLS-R | `Zeroth Korean` Dataset |
|
| 58 |
|
| 59 |
+
## Evaluation Results
|
| 60 |
|
| 61 |
+
The model achieves the following results on evaluation without a language model:
|
| 62 |
+
|
| 63 |
+
| Dataset | WER | CER |
|
| 64 |
+
| -------------------------------- | ------ | ------ |
|
| 65 |
+
| `Zeroth Korean` | 29.54% | 9.53% |
|
| 66 |
+
| `Robust Speech Event - Dev Data` | 76.26% | 38.67% |
|
| 67 |
+
|
| 68 |
+
With the addition of the language model, it achieves the following results:
|
| 69 |
+
|
| 70 |
+
| Dataset | WER | CER |
|
| 71 |
+
| -------------------------------- | ------ | ------ |
|
| 72 |
+
| `Zeroth Korean` | 30.94% | 7.97% |
|
| 73 |
+
| `Robust Speech Event - Dev Data` | 68.34% | 37.08% |
|
| 74 |
|
| 75 |
## Training procedure
|
| 76 |
|
| 77 |
+
The training process did not involve the addition of a language model. The following results were simply lifted from the original automatic speech recognition [model training](https://huggingface.co/w11wo/wav2vec2-xls-r-300m-korean).
|
| 78 |
+
|
| 79 |
### Training hyperparameters
|
| 80 |
|
| 81 |
The following hyperparameters were used during training:
|
| 82 |
+
|
| 83 |
+
- `learning_rate`: 7.5e-05
|
| 84 |
+
- `train_batch_size`: 8
|
| 85 |
+
- `eval_batch_size`: 8
|
| 86 |
+
- `seed`: 42
|
| 87 |
+
- `gradient_accumulation_steps`: 4
|
| 88 |
+
- `total_train_batch_size`: 32
|
| 89 |
+
- `optimizer`: Adam with `betas=(0.9, 0.999)` and `epsilon=1e-08`
|
| 90 |
+
- `lr_scheduler_type`: linear
|
| 91 |
+
- `lr_scheduler_warmup_steps`: 2000
|
| 92 |
+
- `num_epochs`: 50.0
|
| 93 |
+
- `mixed_precision_training`: Native AMP
|
| 94 |
|
| 95 |
### Training results
|
| 96 |
|
| 97 |
+
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
|
| 98 |
+
| :-----------: | :---: | :---: | :-------------: | :----: | :----: |
|
| 99 |
+
| 19.7138 | 0.72 | 500 | 19.6427 | 1.0 | 1.0 |
|
| 100 |
+
| 4.8039 | 1.44 | 1000 | 4.7842 | 1.0 | 1.0 |
|
| 101 |
+
| 4.5619 | 2.16 | 1500 | 4.5608 | 0.9992 | 0.9598 |
|
| 102 |
+
| 4.254 | 2.88 | 2000 | 4.2729 | 0.9955 | 0.9063 |
|
| 103 |
+
| 4.1905 | 3.6 | 2500 | 4.2257 | 0.9903 | 0.8758 |
|
| 104 |
+
| 4.0683 | 4.32 | 3000 | 3.9294 | 0.9937 | 0.7911 |
|
| 105 |
+
| 3.486 | 5.04 | 3500 | 2.7045 | 1.0012 | 0.5934 |
|
| 106 |
+
| 2.946 | 5.75 | 4000 | 1.9691 | 0.9425 | 0.4634 |
|
| 107 |
+
| 2.634 | 6.47 | 4500 | 1.5212 | 0.8807 | 0.3850 |
|
| 108 |
+
| 2.4066 | 7.19 | 5000 | 1.2551 | 0.8177 | 0.3601 |
|
| 109 |
+
| 2.2651 | 7.91 | 5500 | 1.0423 | 0.7650 | 0.3039 |
|
| 110 |
+
| 2.1828 | 8.63 | 6000 | 0.9599 | 0.7273 | 0.3106 |
|
| 111 |
+
| 2.1023 | 9.35 | 6500 | 0.9482 | 0.7161 | 0.3063 |
|
| 112 |
+
| 2.0536 | 10.07 | 7000 | 0.8242 | 0.6767 | 0.2860 |
|
| 113 |
+
| 1.9803 | 10.79 | 7500 | 0.7643 | 0.6563 | 0.2637 |
|
| 114 |
+
| 1.9468 | 11.51 | 8000 | 0.7319 | 0.6441 | 0.2505 |
|
| 115 |
+
| 1.9178 | 12.23 | 8500 | 0.6937 | 0.6320 | 0.2489 |
|
| 116 |
+
| 1.8515 | 12.95 | 9000 | 0.6443 | 0.6053 | 0.2196 |
|
| 117 |
+
| 1.8083 | 13.67 | 9500 | 0.6286 | 0.6122 | 0.2148 |
|
| 118 |
+
| 1.819 | 14.39 | 10000 | 0.6015 | 0.5986 | 0.2074 |
|
| 119 |
+
| 1.7684 | 15.11 | 10500 | 0.5682 | 0.5741 | 0.1982 |
|
| 120 |
+
| 1.7195 | 15.83 | 11000 | 0.5385 | 0.5592 | 0.2007 |
|
| 121 |
+
| 1.7044 | 16.55 | 11500 | 0.5362 | 0.5524 | 0.2097 |
|
| 122 |
+
| 1.6879 | 17.27 | 12000 | 0.5119 | 0.5489 | 0.2083 |
|
| 123 |
+
| 1.656 | 17.98 | 12500 | 0.4990 | 0.5362 | 0.1968 |
|
| 124 |
+
| 1.6122 | 18.7 | 13000 | 0.4561 | 0.5092 | 0.1900 |
|
| 125 |
+
| 1.5919 | 19.42 | 13500 | 0.4778 | 0.5225 | 0.1975 |
|
| 126 |
+
| 1.5896 | 20.14 | 14000 | 0.4563 | 0.5098 | 0.1859 |
|
| 127 |
+
| 1.5589 | 20.86 | 14500 | 0.4362 | 0.4940 | 0.1725 |
|
| 128 |
+
| 1.5353 | 21.58 | 15000 | 0.4140 | 0.4826 | 0.1580 |
|
| 129 |
+
| 1.5441 | 22.3 | 15500 | 0.4031 | 0.4742 | 0.1550 |
|
| 130 |
+
| 1.5116 | 23.02 | 16000 | 0.3916 | 0.4748 | 0.1545 |
|
| 131 |
+
| 1.4731 | 23.74 | 16500 | 0.3841 | 0.4810 | 0.1542 |
|
| 132 |
+
| 1.4647 | 24.46 | 17000 | 0.3752 | 0.4524 | 0.1475 |
|
| 133 |
+
| 1.4328 | 25.18 | 17500 | 0.3587 | 0.4476 | 0.1461 |
|
| 134 |
+
| 1.4129 | 25.9 | 18000 | 0.3429 | 0.4242 | 0.1366 |
|
| 135 |
+
| 1.4062 | 26.62 | 18500 | 0.3450 | 0.4251 | 0.1355 |
|
| 136 |
+
| 1.3928 | 27.34 | 19000 | 0.3297 | 0.4145 | 0.1322 |
|
| 137 |
+
| 1.3906 | 28.06 | 19500 | 0.3210 | 0.4185 | 0.1336 |
|
| 138 |
+
| 1.358 | 28.78 | 20000 | 0.3131 | 0.3970 | 0.1275 |
|
| 139 |
+
| 1.3445 | 29.5 | 20500 | 0.3069 | 0.3920 | 0.1276 |
|
| 140 |
+
| 1.3159 | 30.22 | 21000 | 0.3035 | 0.3961 | 0.1255 |
|
| 141 |
+
| 1.3044 | 30.93 | 21500 | 0.2952 | 0.3854 | 0.1242 |
|
| 142 |
+
| 1.3034 | 31.65 | 22000 | 0.2966 | 0.3772 | 0.1227 |
|
| 143 |
+
| 1.2963 | 32.37 | 22500 | 0.2844 | 0.3706 | 0.1208 |
|
| 144 |
+
| 1.2765 | 33.09 | 23000 | 0.2841 | 0.3567 | 0.1173 |
|
| 145 |
+
| 1.2438 | 33.81 | 23500 | 0.2734 | 0.3552 | 0.1137 |
|
| 146 |
+
| 1.2487 | 34.53 | 24000 | 0.2703 | 0.3502 | 0.1118 |
|
| 147 |
+
| 1.2249 | 35.25 | 24500 | 0.2650 | 0.3484 | 0.1142 |
|
| 148 |
+
| 1.2229 | 35.97 | 25000 | 0.2584 | 0.3374 | 0.1097 |
|
| 149 |
+
| 1.2374 | 36.69 | 25500 | 0.2568 | 0.3337 | 0.1095 |
|
| 150 |
+
| 1.2153 | 37.41 | 26000 | 0.2494 | 0.3327 | 0.1071 |
|
| 151 |
+
| 1.1925 | 38.13 | 26500 | 0.2518 | 0.3366 | 0.1077 |
|
| 152 |
+
| 1.1908 | 38.85 | 27000 | 0.2437 | 0.3272 | 0.1057 |
|
| 153 |
+
| 1.1858 | 39.57 | 27500 | 0.2396 | 0.3265 | 0.1044 |
|
| 154 |
+
| 1.1808 | 40.29 | 28000 | 0.2373 | 0.3156 | 0.1028 |
|
| 155 |
+
| 1.1842 | 41.01 | 28500 | 0.2356 | 0.3152 | 0.1026 |
|
| 156 |
+
| 1.1668 | 41.73 | 29000 | 0.2319 | 0.3188 | 0.1025 |
|
| 157 |
+
| 1.1448 | 42.45 | 29500 | 0.2293 | 0.3099 | 0.0995 |
|
| 158 |
+
| 1.1327 | 43.17 | 30000 | 0.2265 | 0.3047 | 0.0979 |
|
| 159 |
+
| 1.1307 | 43.88 | 30500 | 0.2222 | 0.3078 | 0.0989 |
|
| 160 |
+
| 1.1419 | 44.6 | 31000 | 0.2215 | 0.3038 | 0.0981 |
|
| 161 |
+
| 1.1231 | 45.32 | 31500 | 0.2193 | 0.3013 | 0.0972 |
|
| 162 |
+
| 1.139 | 46.04 | 32000 | 0.2162 | 0.3007 | 0.0968 |
|
| 163 |
+
| 1.1114 | 46.76 | 32500 | 0.2122 | 0.2982 | 0.0960 |
|
| 164 |
+
| 1.111 | 47.48 | 33000 | 0.2125 | 0.2946 | 0.0948 |
|
| 165 |
+
| 1.0982 | 48.2 | 33500 | 0.2099 | 0.2957 | 0.0953 |
|
| 166 |
+
| 1.109 | 48.92 | 34000 | 0.2092 | 0.2955 | 0.0955 |
|
| 167 |
+
| 1.0905 | 49.64 | 34500 | 0.2088 | 0.2954 | 0.0953 |
|
| 168 |
+
|
| 169 |
+
## Disclaimer
|
| 170 |
+
|
| 171 |
+
Do consider the biases which came from pre-training datasets that may be carried over into the results of this model.
|
| 172 |
+
|
| 173 |
+
## Authors
|
| 174 |
+
|
| 175 |
+
Wav2Vec2 XLS-R 300M Korean LM was trained and evaluated by [Wilson Wongso](https://w11wo.github.io/). All computation and development are done on OVH Cloud.
|
| 176 |
+
|
| 177 |
+
## Framework versions
|
| 178 |
|
| 179 |
- Transformers 4.17.0.dev0
|
| 180 |
- Pytorch 1.10.2+cu102
|