Lemswasabi commited on
Commit
7a08f19
1 Parent(s): 8b38a61

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +81 -1
README.md CHANGED
@@ -1,8 +1,88 @@
1
  ---
 
 
 
2
  license: mit
3
  language:
4
  - lb
5
  metrics:
6
  - wer
7
  pipeline_tag: automatic-speech-recognition
8
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ tags:
3
+ - automatic-speech-recognition
4
+ - generated_from_trainer
5
  license: mit
6
  language:
7
  - lb
8
  metrics:
9
  - wer
10
  pipeline_tag: automatic-speech-recognition
11
+
12
+ model-index:
13
+ - name: Lemswasabi/wav2vec2-large-xlsr-53-842h-luxembourgish-8h
14
+ results:
15
+ - task:
16
+ type: automatic-speech-recognition # Required. Example: automatic-speech-recognition
17
+ name: Speech Recognition # Optional. Example: Speech Recognition
18
+ metrics:
19
+ - type: wer
20
+ value: 13.86
21
+ name: Dev WER
22
+ - type: wer
23
+ value: 12.87
24
+ name: Test WER
25
+ - type: cer
26
+ value: 3.11
27
+ name: Dev CER
28
+ - type: cer
29
+ value: 2.91
30
+ name: Test CER
31
+ ---
32
+
33
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
34
+ should probably proofread and complete it, then remove this comment. -->
35
+
36
+ #
37
+
38
+ ## Model description
39
+
40
+ We fine-tuned a wav2vec 2.0 large XLSR-53 checkpoint with 842h of unlabelled Luxembourgish speech
41
+ collected from [RTL.lu](https://www.rtl.lu/). Then the model was fine-tuned on 8h of labelled
42
+ Luxembourgish speech from the same domain.
43
+
44
+ ## Intended uses & limitations
45
+
46
+ More information needed
47
+
48
+ ## Training and evaluation data
49
+
50
+ More information needed
51
+
52
+ ## Training procedure
53
+
54
+ ### Training hyperparameters
55
+
56
+ The following hyperparameters were used during training:
57
+ - learning_rate: 7.5e-05
58
+ - train_batch_size: 3
59
+ - eval_batch_size: 3
60
+ - seed: 42
61
+ - gradient_accumulation_steps: 4
62
+ - total_train_batch_size: 12
63
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
64
+ - lr_scheduler_type: linear
65
+ - lr_scheduler_warmup_steps: 2000
66
+ - num_epochs: 50.0
67
+ - mixed_precision_training: Native AMP
68
+
69
+ ### Framework versions
70
+
71
+ - Transformers 4.20.0.dev0
72
+ - Pytorch 1.11.0+cu113
73
+ - Datasets 2.2.1
74
+ - Tokenizers 0.12.1
75
+
76
+ ## Citation
77
+
78
+ This model is a result of our paper `IMPROVING LUXEMBOURGISH SPEECH RECOGNITION WITH CROSS-LINGUAL SPEECH REPRESENTATIONS` submitted to the [IEEE SLT 2022 workshop](https://slt2022.org/)
79
+
80
+ ```
81
+ @misc{lb-wav2vec2,
82
+ author = {Nguyen, Le Minh and Nayak, Shekhar and Coler, Matt.},
83
+ keywords = {Luxembourgish, multilingual speech recognition, language modelling, wav2vec 2.0 XLSR-53, under-resourced language},
84
+ title = {IMPROVING LUXEMBOURGISH SPEECH RECOGNITION WITH CROSS-LINGUAL SPEECH REPRESENTATIONS},
85
+ year = {2022},
86
+ copyright = {2023 IEEE}
87
+ }
88
+ ```