File size: 1,722 Bytes
f825e4c
 
aa2e5a4
 
f825e4c
 
aa2e5a4
 
f825e4c
 
aa2e5a4
f825e4c
 
 
 
 
 
 
 
 
 
fc46585
f825e4c
fc46585
 
f825e4c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fc46585
f825e4c
 
 
 
 
 
fc46585
 
 
 
f825e4c
 
 
 
 
 
 
 
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
---
license: apache-2.0
language:
- gn
tags:
- generated_from_trainer
- mozilla-foundation/common_voice_8_0
- robust-speech-event
datasets:
- common_voice
- mozilla-foundation/common_voice_8_0
model-index:
- name: wav2vec2-base-gn-demo
  results: []
---

<!-- 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-base-gn-demo

This model is a fine-tuned version of [azuur/wav2vec2-base-gn-demo](https://huggingface.co/azuur/wav2vec2-base-gn-demo) on the common_voice dataset.
It achieves the following results on the evaluation set:
- Loss: 1.1139
- Wer: 0.7835

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 300
- num_epochs: 55
- mixed_precision_training: Native AMP

### Training results

| Training Loss | Epoch | Step | Validation Loss | Wer    |
|:-------------:|:-----:|:----:|:---------------:|:------:|
| 0.1803        | 13.16 | 500  | 1.2114          | 0.8415 |
| 0.1343        | 26.32 | 1000 | 1.1728          | 0.8155 |
| 0.0986        | 39.47 | 1500 | 1.1220          | 0.8293 |
| 0.0581        | 52.63 | 2000 | 1.1139          | 0.7835 |


### Framework versions

- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.18.3
- Tokenizers 0.10.3