RASMUS commited on
Commit
ff76274
1 Parent(s): ee1a8bf

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +12 -10
README.md CHANGED
@@ -5,6 +5,7 @@ license: apache-2.0
5
  tags:
6
  - whisper-event
7
  - finnish
 
8
  datasets:
9
  - mozilla-foundation/common_voice_11_0
10
  - google/fleurs
@@ -46,18 +47,19 @@ model-index:
46
  - name: Cer
47
  type: cer
48
  value: 3.23
 
 
49
  ---
50
 
51
- <h3>This is our improved Whisper model that is now finetuned from OpenAI Whisper Large V3 </h3>
52
- <p>We improve from our previously finetuned V2 model <a>https://huggingface.co/Finnish-NLP/whisper-large-v2-finnish</a> </p>
53
- <p>CV11 WER 10.42 --> 8.23</p>
54
- <p>Fleurs WER 10.20 --> 8.21</p>
55
- <p>Model was trained on RTX4080 for 32k steps with batch size 8, gradient accumulation 2</p>
56
-
57
 
58
  <br></br>
59
 
60
- Original Whisper Large V3
61
  - CV11
62
  - WER: 14.81
63
  - WER NORMALIZED: 10.82
@@ -71,9 +73,9 @@ Original Whisper Large V3
71
  - CER NORMALIZED: 3.64
72
 
73
 
74
- After Finetuning V3:
75
 
76
- - @14000 steps
77
  - CV11
78
  - WER: 11.36
79
  - WER NORMALIZED: 8.31
@@ -86,7 +88,7 @@ After Finetuning V3:
86
  - CER: 2.26
87
  - CER NORMALIZED: 3.54
88
 
89
- - @32000 steps
90
  - CV11
91
  - WER: 11.47
92
  - WER NORMALIZED: 8.23
 
5
  tags:
6
  - whisper-event
7
  - finnish
8
+ - speech-recognition
9
  datasets:
10
  - mozilla-foundation/common_voice_11_0
11
  - google/fleurs
 
47
  - name: Cer
48
  type: cer
49
  value: 3.23
50
+ library_name: transformers
51
+ pipeline_tag: automatic-speech-recognition
52
  ---
53
 
54
+ <h3>This is our improved Whisper v3 model that is now finetuned from OpenAI Whisper Large V3 </h3>
55
+ <p>We improve from our previously finetuned Whisper V2 model in the following manner<a>https://huggingface.co/Finnish-NLP/whisper-large-v2-finnish</a> </p>
56
+ <p>CV11 (Common Voice 11 test set) WER (Word error rate) 10.42 --> 8.23</p>
57
+ <p>Fleurs (A speech recognition test set by Google) WER (Word error rate) 10.20 --> 8.21</p>
58
+ <p>Model was trained on Nvidia RTX4080 for 32k steps with batch size 8, gradient accumulation 2</p>
 
59
 
60
  <br></br>
61
 
62
+ Original OpenAI Whisper Large V3
63
  - CV11
64
  - WER: 14.81
65
  - WER NORMALIZED: 10.82
 
73
  - CER NORMALIZED: 3.64
74
 
75
 
76
+ After Finetuning with Finnish data our V3 got these scores on the test set:
77
 
78
+ - @14000 finetuning steps
79
  - CV11
80
  - WER: 11.36
81
  - WER NORMALIZED: 8.31
 
88
  - CER: 2.26
89
  - CER NORMALIZED: 3.54
90
 
91
+ - @32000 finetuning steps
92
  - CV11
93
  - WER: 11.47
94
  - WER NORMALIZED: 8.23