Update README.md
Browse files
README.md
CHANGED
|
@@ -1,139 +1,179 @@
|
|
| 1 |
---
|
| 2 |
-
language:
|
| 3 |
-
- zh-HK
|
| 4 |
license: apache-2.0
|
| 5 |
tags:
|
| 6 |
-
- automatic-speech-recognition
|
| 7 |
-
-
|
| 8 |
-
-
|
| 9 |
datasets:
|
| 10 |
-
- common_voice
|
| 11 |
model-index:
|
| 12 |
-
- name:
|
| 13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
---
|
| 15 |
|
| 16 |
-
|
| 17 |
-
should probably proofread and complete it, then remove this comment. -->
|
| 18 |
|
| 19 |
-
|
| 20 |
|
| 21 |
-
This model
|
| 22 |
-
It achieves the following results on the evaluation set:
|
| 23 |
-
- Loss: 0.8089
|
| 24 |
-
- Wer: 1.2499
|
| 25 |
-
- Cer: 0.3173
|
| 26 |
|
| 27 |
-
|
| 28 |
|
| 29 |
-
|
| 30 |
|
| 31 |
-
##
|
| 32 |
|
| 33 |
-
|
|
|
|
|
|
|
| 34 |
|
| 35 |
-
##
|
| 36 |
|
| 37 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
## Training procedure
|
| 40 |
|
|
|
|
|
|
|
| 41 |
### Training hyperparameters
|
| 42 |
|
| 43 |
The following hyperparameters were used during training:
|
| 44 |
-
|
| 45 |
-
-
|
| 46 |
-
-
|
| 47 |
-
-
|
| 48 |
-
-
|
| 49 |
-
-
|
| 50 |
-
-
|
| 51 |
-
-
|
| 52 |
-
-
|
| 53 |
-
-
|
| 54 |
-
-
|
|
|
|
| 55 |
|
| 56 |
### Training results
|
| 57 |
|
| 58 |
-
| Training Loss | Epoch | Step | Validation Loss |
|
| 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 |
-
|
|
| 132 |
-
|
|
| 133 |
-
|
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 137 |
|
| 138 |
- Transformers 4.17.0.dev0
|
| 139 |
- Pytorch 1.10.2+cu102
|
|
|
|
| 1 |
---
|
| 2 |
+
language: zh-HK
|
|
|
|
| 3 |
license: apache-2.0
|
| 4 |
tags:
|
| 5 |
+
- automatic-speech-recognition
|
| 6 |
+
- generated_from_trainer
|
| 7 |
+
- robust-speech-event
|
| 8 |
datasets:
|
| 9 |
+
- common_voice
|
| 10 |
model-index:
|
| 11 |
+
- name: Wav2Vec2 XLS-R 300M Cantonese (zh-HK) LM
|
| 12 |
+
results:
|
| 13 |
+
- task:
|
| 14 |
+
name: Automatic Speech Recognition
|
| 15 |
+
type: automatic-speech-recognition
|
| 16 |
+
dataset:
|
| 17 |
+
name: Common Voice
|
| 18 |
+
type: common_voice
|
| 19 |
+
args: zh-HK
|
| 20 |
+
metrics:
|
| 21 |
+
- name: Test CER
|
| 22 |
+
type: cer
|
| 23 |
+
value: 12.14
|
| 24 |
+
- task:
|
| 25 |
+
name: Automatic Speech Recognition
|
| 26 |
+
type: automatic-speech-recognition
|
| 27 |
+
dataset:
|
| 28 |
+
name: Robust Speech Event - Dev Data
|
| 29 |
+
type: speech-recognition-community-v2/dev_data
|
| 30 |
+
args: zh-HK
|
| 31 |
+
metrics:
|
| 32 |
+
- name: Test CER
|
| 33 |
+
type: cer
|
| 34 |
+
value: 56.86
|
| 35 |
---
|
| 36 |
|
| 37 |
+
# Wav2Vec2 XLS-R 300M Cantonese (zh-HK) LM
|
|
|
|
| 38 |
|
| 39 |
+
Wav2Vec2 XLS-R 300M Cantonese (zh-HK) 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 `zh-HK` subset of the [Common Voice](https://huggingface.co/datasets/common_voice) dataset. A 5-gram Language model, trained on multiple [PyCantonese](https://pycantonese.org/data.html) corpora, was then subsequently added to this model.
|
| 40 |
|
| 41 |
+
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.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
+
All necessary scripts used for training could be found in the [Files and versions](https://huggingface.co/w11wo/wav2vec2-xls-r-300m-zh-HK-lm-v2/tree/main) tab, as well as the [Training metrics](https://huggingface.co/w11wo/wav2vec2-xls-r-300m-zh-HK-lm-v2/tensorboard) logged via Tensorboard.
|
| 44 |
|
| 45 |
+
As for the N-gram language model training, we followed the [blog post tutorial](https://huggingface.co/blog/wav2vec2-with-ngram) provided by HuggingFace.
|
| 46 |
|
| 47 |
+
## Model
|
| 48 |
|
| 49 |
+
| Model | #params | Arch. | Training/Validation data (text) |
|
| 50 |
+
| --------------------------------- | ------- | ----- | ------------------------------- |
|
| 51 |
+
| `wav2vec2-xls-r-300m-zh-HK-lm-v2` | 300M | XLS-R | `Common Voice zh-HK` Dataset |
|
| 52 |
|
| 53 |
+
## Evaluation Results
|
| 54 |
|
| 55 |
+
The model achieves the following results on evaluation without a language model:
|
| 56 |
+
|
| 57 |
+
| Dataset | CER |
|
| 58 |
+
| -------------------------------- | ------ |
|
| 59 |
+
| `Common Voice` | 31.73% |
|
| 60 |
+
| `Robust Speech Event - Dev Data` | 56.60% |
|
| 61 |
+
|
| 62 |
+
With the addition of the language model, it achieves the following results:
|
| 63 |
+
|
| 64 |
+
| Dataset | CER |
|
| 65 |
+
| -------------------------------- | ------ |
|
| 66 |
+
| `Common Voice` | 12.14% |
|
| 67 |
+
| `Robust Speech Event - Dev Data` | 56.86% |
|
| 68 |
|
| 69 |
## Training procedure
|
| 70 |
|
| 71 |
+
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).
|
| 72 |
+
|
| 73 |
### Training hyperparameters
|
| 74 |
|
| 75 |
The following hyperparameters were used during training:
|
| 76 |
+
|
| 77 |
+
- `learning_rate`: 0.0001
|
| 78 |
+
- `train_batch_size`: 8
|
| 79 |
+
- `eval_batch_size`: 8
|
| 80 |
+
- `seed`: 42
|
| 81 |
+
- `gradient_accumulation_steps`: 4
|
| 82 |
+
- `total_train_batch_size`: 32
|
| 83 |
+
- `optimizer`: Adam with `betas=(0.9, 0.999)` and `epsilon=1e-08`
|
| 84 |
+
- `lr_scheduler_type`: linear
|
| 85 |
+
- `lr_scheduler_warmup_steps`: 2000
|
| 86 |
+
- `num_epochs`: 100.0
|
| 87 |
+
- `mixed_precision_training`: Native AMP
|
| 88 |
|
| 89 |
### Training results
|
| 90 |
|
| 91 |
+
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
|
| 92 |
+
| :-----------: | :---: | :---: | :-------------: | :----: | :----: |
|
| 93 |
+
| 69.8341 | 1.34 | 500 | 80.0722 | 1.0 | 1.0 |
|
| 94 |
+
| 6.6418 | 2.68 | 1000 | 6.6346 | 1.0 | 1.0 |
|
| 95 |
+
| 6.2419 | 4.02 | 1500 | 6.2909 | 1.0 | 1.0 |
|
| 96 |
+
| 6.0813 | 5.36 | 2000 | 6.1150 | 1.0 | 1.0 |
|
| 97 |
+
| 5.9677 | 6.7 | 2500 | 6.0301 | 1.1386 | 1.0028 |
|
| 98 |
+
| 5.9296 | 8.04 | 3000 | 5.8975 | 1.2113 | 1.0058 |
|
| 99 |
+
| 5.6434 | 9.38 | 3500 | 5.5404 | 2.1624 | 1.0171 |
|
| 100 |
+
| 5.1974 | 10.72 | 4000 | 4.5440 | 2.1702 | 0.9366 |
|
| 101 |
+
| 4.3601 | 12.06 | 4500 | 3.3839 | 2.2464 | 0.8998 |
|
| 102 |
+
| 3.9321 | 13.4 | 5000 | 2.8785 | 2.3097 | 0.8400 |
|
| 103 |
+
| 3.6462 | 14.74 | 5500 | 2.5108 | 1.9623 | 0.6663 |
|
| 104 |
+
| 3.5156 | 16.09 | 6000 | 2.2790 | 1.6479 | 0.5706 |
|
| 105 |
+
| 3.32 | 17.43 | 6500 | 2.1450 | 1.8337 | 0.6244 |
|
| 106 |
+
| 3.1918 | 18.77 | 7000 | 1.8536 | 1.9394 | 0.6017 |
|
| 107 |
+
| 3.1139 | 20.11 | 7500 | 1.7205 | 1.9112 | 0.5638 |
|
| 108 |
+
| 2.8995 | 21.45 | 8000 | 1.5478 | 1.0624 | 0.3250 |
|
| 109 |
+
| 2.7572 | 22.79 | 8500 | 1.4068 | 1.1412 | 0.3367 |
|
| 110 |
+
| 2.6881 | 24.13 | 9000 | 1.3312 | 2.0100 | 0.5683 |
|
| 111 |
+
| 2.5993 | 25.47 | 9500 | 1.2553 | 2.0039 | 0.6450 |
|
| 112 |
+
| 2.5304 | 26.81 | 10000 | 1.2422 | 2.0394 | 0.5789 |
|
| 113 |
+
| 2.4352 | 28.15 | 10500 | 1.1582 | 1.9970 | 0.5507 |
|
| 114 |
+
| 2.3795 | 29.49 | 11000 | 1.1160 | 1.8255 | 0.4844 |
|
| 115 |
+
| 2.3287 | 30.83 | 11500 | 1.0775 | 1.4123 | 0.3780 |
|
| 116 |
+
| 2.2622 | 32.17 | 12000 | 1.0704 | 1.7445 | 0.4894 |
|
| 117 |
+
| 2.2225 | 33.51 | 12500 | 1.0272 | 1.7237 | 0.5058 |
|
| 118 |
+
| 2.1843 | 34.85 | 13000 | 0.9756 | 1.8042 | 0.5028 |
|
| 119 |
+
| 2.1 | 36.19 | 13500 | 0.9527 | 1.8909 | 0.6055 |
|
| 120 |
+
| 2.0741 | 37.53 | 14000 | 0.9418 | 1.9026 | 0.5880 |
|
| 121 |
+
| 2.0179 | 38.87 | 14500 | 0.9363 | 1.7977 | 0.5246 |
|
| 122 |
+
| 2.0615 | 40.21 | 15000 | 0.9635 | 1.8112 | 0.5599 |
|
| 123 |
+
| 1.9448 | 41.55 | 15500 | 0.9249 | 1.7250 | 0.4914 |
|
| 124 |
+
| 1.8966 | 42.89 | 16000 | 0.9023 | 1.5829 | 0.4319 |
|
| 125 |
+
| 1.8662 | 44.24 | 16500 | 0.9002 | 1.4833 | 0.4230 |
|
| 126 |
+
| 1.8136 | 45.58 | 17000 | 0.9076 | 1.1828 | 0.2987 |
|
| 127 |
+
| 1.7908 | 46.92 | 17500 | 0.8774 | 1.5773 | 0.4258 |
|
| 128 |
+
| 1.7354 | 48.26 | 18000 | 0.8727 | 1.5037 | 0.4024 |
|
| 129 |
+
| 1.6739 | 49.6 | 18500 | 0.8636 | 1.1239 | 0.2789 |
|
| 130 |
+
| 1.6457 | 50.94 | 19000 | 0.8516 | 1.2269 | 0.3104 |
|
| 131 |
+
| 1.5847 | 52.28 | 19500 | 0.8399 | 1.3309 | 0.3360 |
|
| 132 |
+
| 1.5971 | 53.62 | 20000 | 0.8441 | 1.3153 | 0.3335 |
|
| 133 |
+
| 1.602 | 54.96 | 20500 | 0.8590 | 1.2932 | 0.3433 |
|
| 134 |
+
| 1.5063 | 56.3 | 21000 | 0.8334 | 1.1312 | 0.2875 |
|
| 135 |
+
| 1.4631 | 57.64 | 21500 | 0.8474 | 1.1698 | 0.2999 |
|
| 136 |
+
| 1.4997 | 58.98 | 22000 | 0.8638 | 1.4279 | 0.3854 |
|
| 137 |
+
| 1.4301 | 60.32 | 22500 | 0.8550 | 1.2737 | 0.3300 |
|
| 138 |
+
| 1.3798 | 61.66 | 23000 | 0.8266 | 1.1802 | 0.2934 |
|
| 139 |
+
| 1.3454 | 63.0 | 23500 | 0.8235 | 1.3816 | 0.3711 |
|
| 140 |
+
| 1.3678 | 64.34 | 24000 | 0.8550 | 1.6427 | 0.5035 |
|
| 141 |
+
| 1.3761 | 65.68 | 24500 | 0.8510 | 1.6709 | 0.4907 |
|
| 142 |
+
| 1.2668 | 67.02 | 25000 | 0.8515 | 1.5842 | 0.4505 |
|
| 143 |
+
| 1.2835 | 68.36 | 25500 | 0.8283 | 1.5353 | 0.4221 |
|
| 144 |
+
| 1.2961 | 69.7 | 26000 | 0.8339 | 1.5743 | 0.4369 |
|
| 145 |
+
| 1.2656 | 71.05 | 26500 | 0.8331 | 1.5331 | 0.4217 |
|
| 146 |
+
| 1.2556 | 72.39 | 27000 | 0.8242 | 1.4708 | 0.4109 |
|
| 147 |
+
| 1.2043 | 73.73 | 27500 | 0.8245 | 1.4469 | 0.4031 |
|
| 148 |
+
| 1.2722 | 75.07 | 28000 | 0.8202 | 1.4924 | 0.4096 |
|
| 149 |
+
| 1.202 | 76.41 | 28500 | 0.8290 | 1.3807 | 0.3719 |
|
| 150 |
+
| 1.1679 | 77.75 | 29000 | 0.8195 | 1.4097 | 0.3749 |
|
| 151 |
+
| 1.1967 | 79.09 | 29500 | 0.8059 | 1.2074 | 0.3077 |
|
| 152 |
+
| 1.1241 | 80.43 | 30000 | 0.8137 | 1.2451 | 0.3270 |
|
| 153 |
+
| 1.1414 | 81.77 | 30500 | 0.8117 | 1.2031 | 0.3121 |
|
| 154 |
+
| 1.132 | 83.11 | 31000 | 0.8234 | 1.4266 | 0.3901 |
|
| 155 |
+
| 1.0982 | 84.45 | 31500 | 0.8064 | 1.3712 | 0.3607 |
|
| 156 |
+
| 1.0797 | 85.79 | 32000 | 0.8167 | 1.3356 | 0.3562 |
|
| 157 |
+
| 1.0119 | 87.13 | 32500 | 0.8215 | 1.2754 | 0.3268 |
|
| 158 |
+
| 1.0216 | 88.47 | 33000 | 0.8163 | 1.2512 | 0.3184 |
|
| 159 |
+
| 1.0375 | 89.81 | 33500 | 0.8137 | 1.2685 | 0.3290 |
|
| 160 |
+
| 0.9794 | 91.15 | 34000 | 0.8220 | 1.2724 | 0.3255 |
|
| 161 |
+
| 1.0207 | 92.49 | 34500 | 0.8165 | 1.2906 | 0.3361 |
|
| 162 |
+
| 1.0169 | 93.83 | 35000 | 0.8153 | 1.2819 | 0.3305 |
|
| 163 |
+
| 1.0127 | 95.17 | 35500 | 0.8187 | 1.2832 | 0.3252 |
|
| 164 |
+
| 0.9978 | 96.51 | 36000 | 0.8111 | 1.2612 | 0.3210 |
|
| 165 |
+
| 0.9923 | 97.85 | 36500 | 0.8076 | 1.2278 | 0.3122 |
|
| 166 |
+
| 1.0451 | 99.2 | 37000 | 0.8086 | 1.2451 | 0.3156 |
|
| 167 |
+
|
| 168 |
+
## Disclaimer
|
| 169 |
+
|
| 170 |
+
Do consider the biases which came from pre-training datasets that may be carried over into the results of this model.
|
| 171 |
+
|
| 172 |
+
## Authors
|
| 173 |
+
|
| 174 |
+
Wav2Vec2 XLS-R 300M Cantonese (zh-HK) LM was trained and evaluated by [Wilson Wongso](https://w11wo.github.io/). All computation and development are done on OVH Cloud.
|
| 175 |
+
|
| 176 |
+
## Framework versions
|
| 177 |
|
| 178 |
- Transformers 4.17.0.dev0
|
| 179 |
- Pytorch 1.10.2+cu102
|