File size: 3,554 Bytes
8964f08
 
 
 
 
d13fd65
f464d34
 
8964f08
 
f464d34
 
 
 
 
 
 
 
d13fd65
f464d34
 
 
 
d13fd65
 
 
8964f08
 
 
 
 
 
 
 
 
b1563e1
d13fd65
8964f08
d13fd65
8964f08
 
 
d13fd65
 
 
8964f08
 
 
3d9cf41
8964f08
 
 
 
 
 
b1563e1
8964f08
 
 
 
 
 
 
 
31357e9
8964f08
 
f464d34
 
 
 
b1563e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f464d34
 
8964f08
 
 
 
 
d13fd65
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
---
license: apache-2.0
tags:
- generated_from_trainer
datasets:
- mozilla-foundation/common_voice_12_0
metrics:
- wer
model-index:
- name: wav2vec2-large-xls-r-1b-frisian
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: common_voice_12_0
      type: common_voice_12_0
      config: fy-NL
      split: test
      args: fy-NL
    metrics:
    - name: Wer
      type: wer
      value: 0.15990775235054105
language:
- fy
---

<!-- 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

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_12_0 dataset.
It achieves the following results on the evaluation set:
- Loss: 0.2634
- WER: 0.1599

This model was developed together with [golesheed](https://huggingface.co/golesheed) for the course "Speech Recognition II" of the "MSc Voice Technology" program at Rijksuniversiteit Groningen - Campus Fryslân.

## Intended uses & limitations

Intended use is for recognizing Frisian speech.

Limitations include not enough hyperparameter tuning, no LM rescoring, and using v12 of Common Voice instead of v13.

## Training and evaluation data

Training and evaluation splits used are the ones available in the Common Voice dataset.

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 8e-05
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 50
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step | Validation Loss | Wer    |
|:-------------:|:-----:|:----:|:---------------:|:------:|
| 4.7284        | 2.1   | 250  | 2.9453          | 1.0    |
| 1.7496        | 4.2   | 500  | 0.5141          | 0.4771 |
| 0.8168        | 6.3   | 750  | 0.3220          | 0.3148 |
| 0.7403        | 8.4   | 1000 | 0.2988          | 0.2573 |
| 0.7298        | 10.5  | 1250 | 0.2794          | 0.2347 |
| 0.6303        | 12.61 | 1500 | 0.2577          | 0.2164 |
| 0.5201        | 14.71 | 1750 | 0.2746          | 0.2162 |
| 0.5189        | 16.81 | 2000 | 0.2543          | 0.2034 |
| 0.5054        | 18.91 | 2250 | 0.2847          | 0.2071 |
| 0.5112        | 21.01 | 2500 | 0.2772          | 0.1979 |
| 0.5105        | 23.11 | 2750 | 0.2633          | 0.1920 |
| 0.5032        | 25.21 | 3000 | 0.2667          | 0.1856 |
| 0.46          | 27.31 | 3250 | 0.2730          | 0.1852 |
| 0.4992        | 29.41 | 3500 | 0.2626          | 0.1782 |
| 0.4535        | 31.51 | 3750 | 0.2778          | 0.1749 |
| 0.4036        | 33.61 | 4000 | 0.2825          | 0.1747 |
| 0.3347        | 35.71 | 4250 | 0.2797          | 0.1708 |
| 0.2708        | 37.82 | 4500 | 0.2662          | 0.1712 |
| 0.1825        | 39.92 | 4750 | 0.2652          | 0.1648 |
| 0.1654        | 42.02 | 5000 | 0.2719          | 0.1628 |
| 0.1387        | 44.12 | 5250 | 0.2552          | 0.1607 |
| 0.1367        | 46.22 | 5500 | 0.2641          | 0.1591 |
| 0.1218        | 48.32 | 5750 | 0.2634          | 0.1598 |


### Framework versions

- Transformers 4.27.3
- Pytorch 2.0.0+cu117
- Datasets 2.10.1
- Tokenizers 0.13.2