Hugging Face's logo

MaxLSB
/
Splitformer

Automatic Speech Recognition
English
Model card Files
xet
Community
MaxLSB commited on
Commit
88b950f
·
verified ·
1 Parent(s): f46c8cc

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +96 -4
README.md CHANGED
@@ -23,10 +23,102 @@ pipeline_tag: automatic-speech-recognition
23
  </a>
24
  </div>
25
 
26
- ## Overview
27
 
28
- **Splitformer** is a 36.7M parameters Conformer-based ASR model trained from scratch on 1000 hours of the LibriSpeech dataset with an early‐exit objective.
29
 
30
- This architecture introduces parallel downsampling layers before the first and last exits to improve performance with minimal extra overhead, while retaining inference speed.
31
 
32
- Our code for training and inference is available on our [GitHub](https://github.com/augustgw/early-exit-transformer) repository.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23
  </a>
24
  </div>
25
 
26
+ ## 1. Overview
27
 
28
+ **Splitformer** is a 36.7M parameters Conformer-based ASR model trained from scratch on 1000 hours of the **LibriSpeech dataset** with an **early‐exit objective**.
29
 
30
+ This architecture introduces **parallel downsampling layers** before the first and last exits to improve performance with minimal extra overhead, while retaining inference speed.
31
 
32
+ Our code for training and inference is available on our [GitHub](https://github.com/augustgw/early-exit-transformer) repository.
33
+
34
+ ### 2. Results on LibriSpeech
35
+
36
+ <table>
37
+ <thead>
38
+ <tr>
39
+ <th rowspan="2">Layer</th>
40
+ <th colspan="2">EE-baseline (31.5M)</th>
41
+ <th colspan="2">Splitformer (36.7M)</th>
42
+ <th colspan="2">Wav2Vec2 (94.0M)</th>
43
+ <th colspan="2">WavLM (94.7M)</th>
44
+ </tr>
45
+ <tr>
46
+ <th>test-clean</th>
47
+ <th>test-other</th>
48
+ <th>test-clean</th>
49
+ <th>test-other</th>
50
+ <th>test-clean</th>
51
+ <th>test-other</th>
52
+ <th>test-clean</th>
53
+ <th>test-other</th>
54
+ </tr>
55
+ </thead>
56
+ <tbody>
57
+ <tr>
58
+ <td>2</td>
59
+ <td>31.0</td>
60
+ <td>51.0</td>
61
+ <td>28.1</td>
62
+ <td>48.3</td>
63
+ <td>33.7</td>
64
+ <td>56.0</td>
65
+ <td>28.0</td>
66
+ <td>48.5</td>
67
+ </tr>
68
+ <tr>
69
+ <td>4</td>
70
+ <td>11.7</td>
71
+ <td>27.8</td>
72
+ <td>10.8</td>
73
+ <td>26.4</td>
74
+ <td>17.4</td>
75
+ <td>36.7</td>
76
+ <td>13.9</td>
77
+ <td>27.3</td>
78
+ </tr>
79
+ <tr>
80
+ <td>6</td>
81
+ <td>7.1</td>
82
+ <td>19.8</td>
83
+ <td>6.7</td>
84
+ <td>19.2</td>
85
+ <td>9.6</td>
86
+ <td>23.7</td>
87
+ <td>8.7</td>
88
+ <td>18.4</td>
89
+ </tr>
90
+ <tr>
91
+ <td>8</td>
92
+ <td>5.8</td>
93
+ <td>16.6</td>
94
+ <td>5.5</td>
95
+ <td>16.3</td>
96
+ <td>5.8</td>
97
+ <td>15.9</td>
98
+ <td>4.8</td>
99
+ <td>12.4</td>
100
+ </tr>
101
+ <tr>
102
+ <td>10</td>
103
+ <td>5.3</td>
104
+ <td>15.3</td>
105
+ <td>5.1</td>
106
+ <td>15.1</td>
107
+ <td>4.5</td>
108
+ <td>12.6</td>
109
+ <td>4.0</td>
110
+ <td>9.5</td>
111
+ </tr>
112
+ <tr>
113
+ <td>12</td>
114
+ <td>5.1</td>
115
+ <td>14.8</td>
116
+ <td>4.8</td>
117
+ <td>14.7</td>
118
+ <td>4.3</td>
119
+ <td>12.2</td>
120
+ <td>3.6</td>
121
+ <td>8.8</td>
122
+ </tr>
123
+ </tbody>
124
+ </table>