Hugging Face's logo

TKU410410103
/
uniTKU-hubert-japanese-asr

Automatic Speech Recognition
Transformers
Safetensors
Japanese
hubert
Generated from Trainer
Eval Results
Inference Endpoints
Model card Files Community
TKU410410103 commited on
Commit
13ed52a
•
1 Parent(s): b39408d

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +15 -17
README.md CHANGED
@@ -18,10 +18,10 @@ model-index:
18
  metrics:
19
  - name: Test WER
20
  type: wer
21
- value: 27.447168
22
  - name: Test CER
23
  type: cer
24
- value: 11.607944
25
  datasets:
26
  - mozilla-foundation/common_voice_11_0
27
  language:
@@ -40,16 +40,14 @@ Fine-tuning on the uniTKU dataset led to the following results:
40
 
41
  | Step | Training Loss | Validation Loss | WER |
42
  |-------|---------------|-----------------|--------|
43
- | 100 | 0.910100 | 1.051628 | 0.669118|
44
- | 200 | 0.747700 | 0.691642 | 0.551471|
45
- | 300 | 0.718000 | 0.705763 | 0.544118|
46
- | 400 | 0.663700 | 0.532831 | 0.397059|
47
- | 500 | 0.667700 | 0.491024 | 0.352941|
48
- | 600 | 0.546800 | 0.365637 | 0.330882|
49
- | 700 | 0.569000 | 0.274410 | 0.279412|
50
- | 800 | 0.591800 | 0.274801 | 0.235294|
51
- | 900 | 0.575400 | 0.257891 | 0.220588|
52
- | 1000 | 0.579100 | 0.280559 | 0.205882|
53
 
54
  ### Training hyperparameters
55
 
@@ -59,7 +57,7 @@ The training hyperparameters remained consistent throughout the fine-tuning proc
59
  - train_batch_size: 16
60
  - eval_batch_size: 16
61
  - gradient_accumulation_steps: 2
62
- - num_train_epochs: 15
63
  - lr_scheduler_type: linear
64
 
65
  ### How to evaluate the model
@@ -151,12 +149,12 @@ print("CER: {:2f}%".format(100 * cer_result))
151
  The final model was evaluated as follows:
152
 
153
  On uniTKU Dataset:
154
- - WER: 20.588235%
155
- - CER: 13.027523%
156
 
157
  On common_voice_11_0:
158
- - WER: 27.447168%
159
- - CER: 11.607944%
160
 
161
  ### Framework versions
162
 
 
18
  metrics:
19
  - name: Test WER
20
  type: wer
21
+ value: 27.511982
22
  - name: Test CER
23
  type: cer
24
+ value: 11.563649
25
  datasets:
26
  - mozilla-foundation/common_voice_11_0
27
  language:
 
40
 
41
  | Step | Training Loss | Validation Loss | WER |
42
  |-------|---------------|-----------------|--------|
43
+ | 100 | 1.127100 | 1.089644 | 0.668508|
44
+ | 200 | 0.873500 | 0.682353 | 0.508287|
45
+ | 300 | 0.786200 | 0.482965 | 0.397790|
46
+ | 400 | 0.670400 | 0.345377 | 0.381215|
47
+ | 500 | 0.719500 | 0.387554 | 0.337017|
48
+ | 600 | 0.707700 | 0.371083 | 0.292818|
49
+ | 700 | 0.658300 | 0.236447 | 0.243094|
50
+ | 800 | 0.611100 | 0.207679 | 0.193370|
 
 
51
 
52
  ### Training hyperparameters
53
 
 
57
  - train_batch_size: 16
58
  - eval_batch_size: 16
59
  - gradient_accumulation_steps: 2
60
+ - max_steps: 800
61
  - lr_scheduler_type: linear
62
 
63
  ### How to evaluate the model
 
149
  The final model was evaluated as follows:
150
 
151
  On uniTKU Dataset:
152
+ - WER: 19.003370%
153
+ - CER: 11.027523%
154
 
155
  On common_voice_11_0:
156
+ - WER: 27.511982%
157
+ - CER: 11.563649%
158
 
159
  ### Framework versions
160