saattrupdan commited on
Commit
a5fced5
1 Parent(s): 9707c2d

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +33 -73
README.md CHANGED
@@ -1,87 +1,47 @@
1
  ---
 
 
2
  license: apache-2.0
3
- tags:
4
- - generated_from_trainer
5
  datasets:
6
- - common_voice
 
 
7
  model-index:
8
  - name: alvenir-wav2vec2-base-cv8-da
9
- results: []
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
10
  ---
11
 
12
- <!-- This model card has been generated automatically according to the information the Trainer had access to. You
13
- should probably proofread and complete it, then remove this comment. -->
14
-
15
- # alvenir-wav2vec2-base-cv8-da
16
-
17
- This model is a fine-tuned version of [Alvenir/wav2vec2-base-da](https://huggingface.co/Alvenir/wav2vec2-base-da) on the common_voice dataset.
18
- It achieves the following results on the evaluation set:
19
- - Loss: 1.1833
20
- - Wer: 0.4851
21
 
22
  ## Model description
23
 
24
- More information needed
25
-
26
- ## Intended uses & limitations
27
-
28
- More information needed
29
-
30
- ## Training and evaluation data
31
-
32
- More information needed
33
-
34
- ## Training procedure
35
-
36
- ### Training hyperparameters
37
-
38
- The following hyperparameters were used during training:
39
- - learning_rate: 4e-05
40
- - train_batch_size: 8
41
- - eval_batch_size: 8
42
- - seed: 42
43
- - gradient_accumulation_steps: 4
44
- - total_train_batch_size: 32
45
- - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
46
- - lr_scheduler_type: linear
47
- - lr_scheduler_warmup_steps: 500
48
- - num_epochs: 500
49
- - mixed_precision_training: Native AMP
50
-
51
- ### Training results
52
 
53
- | Training Loss | Epoch | Step | Validation Loss | Wer |
54
- |:-------------:|:------:|:----:|:---------------:|:------:|
55
- | 3.1193 | 5.55 | 300 | 2.9589 | 1.0 |
56
- | 2.7263 | 11.11 | 600 | 1.9051 | 0.9900 |
57
- | 0.8586 | 16.66 | 900 | 0.8672 | 0.6476 |
58
- | 0.5162 | 22.22 | 1200 | 0.8564 | 0.5739 |
59
- | 0.4339 | 27.77 | 1500 | 0.8644 | 0.5396 |
60
- | 0.3467 | 33.33 | 1800 | 0.8947 | 0.5382 |
61
- | 0.2742 | 38.88 | 2100 | 0.9070 | 0.5184 |
62
- | 0.2675 | 44.44 | 2400 | 0.8916 | 0.5070 |
63
- | 0.2362 | 49.99 | 2700 | 0.9895 | 0.5108 |
64
- | 0.2076 | 55.55 | 3000 | 0.9703 | 0.5054 |
65
- | 0.1944 | 61.11 | 3300 | 0.9893 | 0.5014 |
66
- | 0.1617 | 66.66 | 3600 | 1.0011 | 0.5080 |
67
- | 0.1634 | 72.22 | 3900 | 1.0013 | 0.4959 |
68
- | 0.1488 | 77.77 | 4200 | 1.0574 | 0.5 |
69
- | 0.1447 | 83.33 | 4500 | 1.0825 | 0.4917 |
70
- | 0.126 | 88.88 | 4800 | 1.0953 | 0.4949 |
71
- | 0.1326 | 94.44 | 5100 | 1.1292 | 0.4925 |
72
- | 0.1229 | 99.99 | 5400 | 1.1357 | 0.4998 |
73
- | 0.1216 | 105.55 | 5700 | 1.1957 | 0.4933 |
74
- | 0.1153 | 111.11 | 6000 | 1.1763 | 0.4847 |
75
- | 0.1054 | 116.66 | 6300 | 1.2280 | 0.5012 |
76
- | 0.1056 | 122.22 | 6600 | 1.1631 | 0.4878 |
77
- | 0.1009 | 127.77 | 6900 | 1.1889 | 0.4918 |
78
- | 0.0936 | 133.33 | 7200 | 1.2309 | 0.4916 |
79
- | 0.1 | 138.88 | 7500 | 1.1833 | 0.4851 |
80
 
 
81
 
82
- ### Framework versions
83
 
84
- - Transformers 4.16.2
85
- - Pytorch 1.10.2+cu102
86
- - Datasets 1.18.3
87
- - Tokenizers 0.11.0
 
1
  ---
2
+ language:
3
+ - da
4
  license: apache-2.0
5
+ tasks:
6
+ - automatic-speech-recognition
7
  datasets:
8
+ - common_voice_8_0
9
+ metrics:
10
+ - wer
11
  model-index:
12
  - name: alvenir-wav2vec2-base-cv8-da
13
+ results:
14
+ - task:
15
+ type: automatic-speech-recognition
16
+ dataset:
17
+ type: Alvenir/alvenir_asr_da_eval
18
+ name: Alvenir ASR test dataset
19
+ metrics:
20
+ - type: wer
21
+ value: 41.08
22
+ - task:
23
+ type: automatic-speech-recognition
24
+ dataset:
25
+ type: mozilla-foundation/common_voice_8_0
26
+ args: da
27
+ name: Danish Common Voice 8.0
28
+ metrics:
29
+ - type: wer
30
+ value: 46.05
31
  ---
32
 
33
+ # Alvenir-Wav2vec2-base-CV8-da
 
 
 
 
 
 
 
 
34
 
35
  ## Model description
36
 
37
+ This model is a fine-tuned version of the Danish acoustic model [Alvenir/wav2vec2-base-da](https://huggingface.co/Alvenir/wav2vec2-base-da) on the Danish part of [Common Voice 8.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_8_0), containing ~6 crowdsourced hours of read-aloud Danish speech.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39
 
40
+ ## Performance
41
 
42
+ The model achieves the following WER scores (lower is better):
43
 
44
+ | **Dataset** | **WER without LM** | **WER with 5-gram LM** |
45
+ | :---: | ---: | ---: |
46
+ | [Danish part of Common Voice 8.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_8_0/viewer/da/train) | 46.05 | xx.xx |
47
+ | [Alvenir test set](https://huggingface.co/datasets/Alvenir/alvenir_asr_da_eval) | 41.08 | xx.xx |