Hugging Face's logo

sammy786
/
wav2vec2-xlsr-dhivehi

Automatic Speech Recognition
Transformers
PyTorch
Divehi
wav2vec2
mozilla-foundation/common_voice_8_0
Generated from Trainer
robust-speech-event
model_for_talk
hf-asr-leaderboard
Eval Results
Inference Endpoints
Model card Files Community
1
sammy786 commited on
Commit
e8cab68
•
1 Parent(s): 402259d

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +143 -0
README.md ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - dv
4
+ license: apache-2.0
5
+ tags:
6
+ - automatic-speech-recognition
7
+ - mozilla-foundation/common_voice_8_0
8
+ - generated_from_trainer
9
+ - dv
10
+ - robust-speech-event
11
+ - model_for_talk
12
+ datasets:
13
+ - mozilla-foundation/common_voice_8_0
14
+
15
+ model-index:
16
+ - name: sammy786/wav2vec2-xlsr-dhivehi
17
+ results:
18
+ - task:
19
+ name: Automatic Speech Recognition
20
+ type: automatic-speech-recognition
21
+ dataset:
22
+ name: Common Voice 8
23
+ type: mozilla-foundation/common_voice_8_0
24
+ args: dv
25
+ metrics:
26
+ - name: Test WER
27
+ type: wer
28
+ value: 26.91
29
+ - name: Test CER
30
+ type: cer
31
+ value: 4.02
32
+
33
+ ---
34
+ # sammy786/wav2vec2-xlsr-dhivehi
35
+
36
+ This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - dv dataset.
37
+ It achieves the following results on evaluation set (which is 10 percent of train data set merged with other and dev datasets):
38
+ - Loss: 14.86
39
+ - Wer: 29.32
40
+
41
+ ## Model description
42
+ "facebook/wav2vec2-xls-r-1b" was finetuned.
43
+
44
+ ## Intended uses & limitations
45
+ More information needed
46
+ ## Training and evaluation data
47
+ Training data -
48
+ Common voice Finnish train.tsv, dev.tsv and other.tsv
49
+
50
+ ## Training procedure
51
+ For creating the train dataset, all possible datasets were appended and 90-10 split was used.
52
+
53
+ ### Training hyperparameters
54
+
55
+ The following hyperparameters were used during training:
56
+
57
+ - learning_rate: 0.000045637994662983496
58
+ - train_batch_size: 8
59
+ - eval_batch_size: 16
60
+ - seed: 13
61
+ - gradient_accumulation_steps: 4
62
+ - total_train_batch_size: 32
63
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
64
+ - lr_scheduler_type: cosine_with_restarts
65
+ - lr_scheduler_warmup_steps: 500
66
+ - num_epochs: 30
67
+ - mixed_precision_training: Native AMP
68
+
69
+
70
+ ### Training results
71
+
72
+ | Step | Training Loss | Validation Loss | Wer |
73
+ |-------|---------------|-----------------|----------|
74
+ | 200 | 4.883800 | 3.190218 | 1.000000 |
75
+ | 400 | 1.600100 | 0.497887 | 0.726159 |
76
+ | 600 | 0.928500 | 0.358781 | 0.603892 |
77
+ | 800 | 0.867900 | 0.309132 | 0.570786 |
78
+ | 1000 | 0.743100 | 0.309116 | 0.552954 |
79
+ | 1200 | 0.725100 | 0.266839 | 0.538378 |
80
+ | 1400 | 0.786200 | 0.259797 | 0.535897 |
81
+ | 1600 | 0.655700 | 0.245691 | 0.517290 |
82
+ | 1800 | 0.650500 | 0.246957 | 0.516204 |
83
+ | 2000 | 0.685500 | 0.234808 | 0.516204 |
84
+ | 2200 | 0.487100 | 0.228409 | 0.507753 |
85
+ | 2400 | 0.401300 | 0.221087 | 0.495968 |
86
+ | 2600 | 0.359300 | 0.212476 | 0.489301 |
87
+ | 2800 | 0.347300 | 0.204848 | 0.487750 |
88
+ | 3000 | 0.327000 | 0.203163 | 0.478756 |
89
+ | 3200 | 0.337100 | 0.210235 | 0.487595 |
90
+ | 3400 | 0.308900 | 0.201471 | 0.491316 |
91
+ | 3600 | 0.292600 | 0.192437 | 0.476120 |
92
+ | 3800 | 0.289600 | 0.198398 | 0.468445 |
93
+ | 4000 | 0.290200 | 0.193484 | 0.467204 |
94
+ | 4200 | 0.272600 | 0.193999 | 0.470150 |
95
+ | 4400 | 0.266700 | 0.187384 | 0.460769 |
96
+ | 4600 | 0.253800 | 0.187279 | 0.476663 |
97
+ | 4800 | 0.266400 | 0.197395 | 0.466817 |
98
+ | 5000 | 0.258000 | 0.188920 | 0.456660 |
99
+ | 5200 | 0.237200 | 0.180770 | 0.457358 |
100
+ | 5400 | 0.237900 | 0.178149 | 0.448287 |
101
+ | 5600 | 0.232600 | 0.179827 | 0.461002 |
102
+ | 5800 | 0.228500 | 0.182142 | 0.445185 |
103
+ | 6000 | 0.221000 | 0.173619 | 0.440688 |
104
+ | 6200 | 0.219500 | 0.172291 | 0.442859 |
105
+ | 6400 | 0.219400 | 0.173339 | 0.430609 |
106
+ | 6600 | 0.201900 | 0.177552 | 0.426423 |
107
+ | 6800 | 0.199000 | 0.173157 | 0.429834 |
108
+ | 7000 | 0.200000 | 0.166503 | 0.423709 |
109
+ | 7200 | 0.194600 | 0.171812 | 0.429834 |
110
+ | 7400 | 0.192100 | 0.164989 | 0.420530 |
111
+ | 7600 | 0.185000 | 0.168355 | 0.418825 |
112
+ | 7800 | 0.175100 | 0.168128 | 0.419290 |
113
+ | 8000 | 0.173500 | 0.167959 | 0.424950 |
114
+ | 8200 | 0.172200 | 0.173643 | 0.414793 |
115
+ | 8400 | 0.164200 | 0.167020 | 0.406342 |
116
+ | 8600 | 0.170800 | 0.168050 | 0.405334 |
117
+ | 8800 | 0.157900 | 0.164290 | 0.396573 |
118
+ | 9000 | 0.159900 | 0.163188 | 0.397426 |
119
+ | 9200 | 0.151700 | 0.164370 | 0.390991 |
120
+ | 9400 | 0.146600 | 0.165053 | 0.392852 |
121
+ | 9600 | 0.142200 | 0.164939 | 0.391844 |
122
+ | 9800 | 0.148300 | 0.164422 | 0.385719 |
123
+ | 10000 | 0.136200 | 0.166569 | 0.385951 |
124
+ | 10200 | 0.140700 | 0.161377 | 0.379594 |
125
+ | 10400 | 0.133300 | 0.165194 | 0.378276 |
126
+ | 10600 | 0.131300 | 0.164328 | 0.369205 |
127
+ | 10800 | 0.135500 | 0.160254 | 0.373236 |
128
+ | 11000 | 0.121100 | 0.163522 | 0.372693 |
129
+
130
+
131
+ ### Framework versions
132
+ - Transformers 4.16.0.dev0
133
+ - Pytorch 1.10.0+cu102
134
+ - Datasets 1.17.1.dev0
135
+ - Tokenizers 0.10.3
136
+
137
+ #### Evaluation Commands
138
+
139
+ 1. To evaluate on `mozilla-foundation/common_voice_8_0` with split `test`
140
+
141
+ ```bash
142
+ python eval.py --model_id sammy786/wav2vec2-xlsr-dhivehi --dataset mozilla-foundation/common_voice_8_0 --config dv --split test
143
+ ```