File size: 5,099 Bytes
7c0de89
 
 
 
 
 
 
 
 
c736a1b
7c0de89
 
 
 
 
 
 
 
 
 
 
 
 
 
885c200
 
 
 
 
 
 
 
 
 
 
 
 
7c0de89
 
 
 
 
c736a1b
7c0de89
 
 
 
 
 
885c200
 
 
7c0de89
 
4009363
e64a221
7c0de89
 
 
4009363
 
 
7c0de89
 
 
4009363
7c0de89
 
 
4009363
 
7c0de89
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
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
132
133
134
135
---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- common_voice_8_0
metrics:
- wer
model-index:
- name: wav2vec2-large-xls-r-1b-frisian-cv-8-10h
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: common_voice_8_0
      type: common_voice_8_0
      config: fy-NL
      split: validation
      args: fy-NL
    metrics:
    - name: Wer
      type: wer
      value: 0.09612912441079846
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: common_voice_8_0
      type: common_voice_8_0
      config: fy-NL
      split: test
      args: fy-NL
    metrics:
    - name: Wer
      type: wer
      value: 0.08830755889579418
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# wav2vec2-large-xls-r-1b-frisian-cv-8-10h

This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the common_voice_8_0 dataset.
It achieves the following results on the evaluation set:
- Loss: 0.1207
- Wer: 0.0961

And on the test set:
- Wer: 0.0883

## Model description

This model has been developed for my Master's thesis in "Voice Technology" at Rijksuniversiteit Groningen - Campus Fryslân. It corresponds to experiment 3 where 
I use as training set 10 hours of Frisian speech randomly selected from all validated data except the test and evaluation sets.

## Intended uses & limitations

The intended use is for recognizing Frisian speech.

Limitations include no LM rescoring and using version 8.0 of Common Voice instead of 13.0.

## Training and evaluation data

The evaluation split used is the one available in the Common Voice 8.0 Frisian subset. The train split is 10 hours of Frisian randomly selected from validated data except for the recordings from test and evaluation splits.

## Training procedure

The script used for training this model can be found in this GitHub repository: [link](https://github.com/greenw0lf/MSc-VT-Thesis/).

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 32
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 50
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step  | Validation Loss | Wer    |
|:-------------:|:-----:|:-----:|:---------------:|:------:|
| 5.6342        | 1.32  | 300   | 2.9760          | 1.0    |
| 2.2716        | 2.63  | 600   | 0.6877          | 0.6024 |
| 1.1303        | 3.95  | 900   | 0.3522          | 0.3450 |
| 0.9038        | 5.26  | 1200  | 0.2714          | 0.2603 |
| 0.846         | 6.58  | 1500  | 0.2143          | 0.2036 |
| 0.8044        | 7.89  | 1800  | 0.1829          | 0.1788 |
| 0.7069        | 9.21  | 2100  | 0.1751          | 0.1667 |
| 0.6995        | 10.53 | 2400  | 0.1741          | 0.1727 |
| 0.7115        | 11.84 | 2700  | 0.1591          | 0.1486 |
| 0.677         | 13.16 | 3000  | 0.1636          | 0.1459 |
| 0.6032        | 14.47 | 3300  | 0.1535          | 0.1439 |
| 0.6218        | 15.79 | 3600  | 0.1427          | 0.1406 |
| 0.6519        | 17.11 | 3900  | 0.1498          | 0.1488 |
| 0.5739        | 18.42 | 4200  | 0.1438          | 0.1319 |
| 0.567         | 19.74 | 4500  | 0.1379          | 0.1322 |
| 0.4982        | 21.05 | 4800  | 0.1315          | 0.1237 |
| 0.5825        | 22.37 | 5100  | 0.1349          | 0.1252 |
| 0.5085        | 23.68 | 5400  | 0.1297          | 0.1233 |
| 0.4946        | 25.0  | 5700  | 0.1343          | 0.1127 |
| 0.5677        | 26.32 | 6000  | 0.1323          | 0.1228 |
| 0.4858        | 27.63 | 6300  | 0.1292          | 0.1098 |
| 0.4709        | 28.95 | 6600  | 0.1267          | 0.1204 |
| 0.3241        | 30.26 | 6900  | 0.1315          | 0.1274 |
| 0.2796        | 31.58 | 7200  | 0.1315          | 0.1202 |
| 0.3171        | 32.89 | 7500  | 0.1315          | 0.1200 |
| 0.2591        | 34.21 | 7800  | 0.1322          | 0.1106 |
| 0.2716        | 35.53 | 8100  | 0.1233          | 0.1030 |
| 0.2446        | 36.84 | 8400  | 0.1273          | 0.1087 |
| 0.2377        | 38.16 | 8700  | 0.1243          | 0.1101 |
| 0.2183        | 39.47 | 9000  | 0.1230          | 0.1116 |
| 0.2059        | 40.79 | 9300  | 0.1240          | 0.1001 |
| 0.1916        | 42.11 | 9600  | 0.1223          | 0.1003 |
| 0.196         | 43.42 | 9900  | 0.1246          | 0.0965 |
| 0.1969        | 44.74 | 10200 | 0.1222          | 0.1038 |
| 0.1951        | 46.05 | 10500 | 0.1208          | 0.1003 |
| 0.1809        | 47.37 | 10800 | 0.1213          | 0.1003 |
| 0.1793        | 48.68 | 11100 | 0.1202          | 0.0959 |
| 0.1837        | 50.0  | 11400 | 0.1207          | 0.0961 |


### Framework versions

- Transformers 4.28.1
- Pytorch 2.0.0+cu117
- Datasets 2.11.0
- Tokenizers 0.13.3