File size: 4,082 Bytes
f189410
 
 
3f92893
f189410
 
 
 
 
3f92893
 
 
b42e604
 
f189410
 
 
 
 
 
 
c3f8d10
f189410
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c3f8d10
f189410
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c3f8d10
 
 
f189410
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
tags:
- generated_from_trainer
- asr
metrics:
- wer
model-index:
- name: libri-alpha-0.85-Temp-1-processor-change
  results: []
#   - task:
#     type: automatic-speech-recognition
#     name: Speech Recognition
datasets:
- librispeech_asr
---

<!-- 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. -->

# libri-alpha-0.85-Temp-1-processor-change

This model is a distilled version of [Wav2vec2](https://huggingface.co/) on the 30% of the Librispeech-clean.100 dataset.
It achieves the following results on the evaluation set:
- Loss: 78.4467
- Wer: 0.1153

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure
Knowledge distillation of Wav2vec2-base-960h teacher model with 6 attention layers for student model.

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 30
- mixed_precision_training: Native AMP
- alpha: 0.75(ignore name of repo)
- temperature: 1
  

### Training results

| Training Loss | Epoch | Step | Validation Loss | Wer    |
|:-------------:|:-----:|:----:|:---------------:|:------:|
| 493.9213      | 0.75  | 100  | 145.7981        | 0.1515 |
| 410.8468      | 1.49  | 200  | 119.1579        | 0.1498 |
| 368.5187      | 2.24  | 300  | 109.7572        | 0.1505 |
| 329.7762      | 2.99  | 400  | 99.2350         | 0.1439 |
| 323.7352      | 3.73  | 500  | 92.1173         | 0.1356 |
| 305.1129      | 4.48  | 600  | 89.3685         | 0.1314 |
| 294.2529      | 5.22  | 700  | 88.3937         | 0.1287 |
| 284.5355      | 5.97  | 800  | 87.0589         | 0.1292 |
| 284.2181      | 6.72  | 900  | 86.4474         | 0.1298 |
| 273.915       | 7.46  | 1000 | 84.6149         | 0.1265 |
| 267.7668      | 8.21  | 1100 | 84.1840         | 0.1264 |
| 262.1592      | 8.96  | 1200 | 83.8678         | 0.1253 |
| 262.5562      | 9.7   | 1300 | 83.2756         | 0.1207 |
| 262.9982      | 10.45 | 1400 | 81.8095         | 0.1218 |
| 256.2891      | 11.19 | 1500 | 82.1241         | 0.1204 |
| 251.4134      | 11.94 | 1600 | 80.8432         | 0.1207 |
| 250.0854      | 12.69 | 1700 | 81.1467         | 0.1203 |
| 250.0077      | 13.43 | 1800 | 80.9370         | 0.1196 |
| 239.0915      | 14.18 | 1900 | 80.5060         | 0.1201 |
| 240.9192      | 14.93 | 2000 | 80.4557         | 0.1190 |
| 241.1668      | 15.67 | 2100 | 80.6453         | 0.1203 |
| 244.9744      | 16.42 | 2200 | 80.0101         | 0.1192 |
| 232.4748      | 17.16 | 2300 | 79.4798         | 0.1170 |
| 237.3503      | 17.91 | 2400 | 79.5743         | 0.1175 |
| 237.9698      | 18.66 | 2500 | 79.3368         | 0.1178 |
| 235.8808      | 19.4  | 2600 | 79.5519         | 0.1174 |
| 230.8314      | 20.15 | 2700 | 79.0367         | 0.1166 |
| 229.5856      | 20.9  | 2800 | 79.1809         | 0.1172 |
| 233.1034      | 21.64 | 2900 | 78.9896         | 0.1167 |
| 231.6986      | 22.39 | 3000 | 78.7184         | 0.1154 |
| 222.0106      | 23.13 | 3100 | 78.7308         | 0.1160 |
| 225.1484      | 23.88 | 3200 | 78.6649         | 0.1159 |
| 232.4254      | 24.63 | 3300 | 78.5096         | 0.1154 |
| 230.9492      | 25.37 | 3400 | 78.4873         | 0.1153 |
| 228.3062      | 26.12 | 3500 | 78.5155         | 0.1147 |
| 225.5572      | 26.87 | 3600 | 78.5693         | 0.1148 |
| 227.7358      | 27.61 | 3700 | 78.5487         | 0.1149 |
| 221.2486      | 28.36 | 3800 | 78.4307         | 0.1151 |
| 231.5915      | 29.1  | 3900 | 78.4270         | 0.1153 |
| 231.7214      | 29.85 | 4000 | 78.4467         | 0.1153 |


### Framework versions

- Transformers 4.25.1
- Pytorch 1.12.1
- Datasets 2.7.1
- Tokenizers 0.11.0