TKU410410103
commited on
Commit
•
38b90d8
1
Parent(s):
3e2dc14
Update README.md
Browse files
README.md
CHANGED
@@ -1,199 +1,164 @@
|
|
1 |
---
|
2 |
-
|
3 |
-
tags:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
## Environmental Impact
|
142 |
-
|
143 |
-
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
|
144 |
-
|
145 |
-
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
|
146 |
-
|
147 |
-
- **Hardware Type:** [More Information Needed]
|
148 |
-
- **Hours used:** [More Information Needed]
|
149 |
-
- **Cloud Provider:** [More Information Needed]
|
150 |
-
- **Compute Region:** [More Information Needed]
|
151 |
-
- **Carbon Emitted:** [More Information Needed]
|
152 |
-
|
153 |
-
## Technical Specifications [optional]
|
154 |
-
|
155 |
-
### Model Architecture and Objective
|
156 |
-
|
157 |
-
[More Information Needed]
|
158 |
-
|
159 |
-
### Compute Infrastructure
|
160 |
-
|
161 |
-
[More Information Needed]
|
162 |
-
|
163 |
-
#### Hardware
|
164 |
-
|
165 |
-
[More Information Needed]
|
166 |
-
|
167 |
-
#### Software
|
168 |
-
|
169 |
-
[More Information Needed]
|
170 |
-
|
171 |
-
## Citation [optional]
|
172 |
-
|
173 |
-
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
174 |
-
|
175 |
-
**BibTeX:**
|
176 |
-
|
177 |
-
[More Information Needed]
|
178 |
-
|
179 |
-
**APA:**
|
180 |
-
|
181 |
-
[More Information Needed]
|
182 |
-
|
183 |
-
## Glossary [optional]
|
184 |
-
|
185 |
-
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
|
186 |
-
|
187 |
-
[More Information Needed]
|
188 |
-
|
189 |
-
## More Information [optional]
|
190 |
-
|
191 |
-
[More Information Needed]
|
192 |
-
|
193 |
-
## Model Card Authors [optional]
|
194 |
-
|
195 |
-
[More Information Needed]
|
196 |
-
|
197 |
-
## Model Card Contact
|
198 |
-
|
199 |
-
[More Information Needed]
|
|
|
1 |
---
|
2 |
+
license: apache-2.0
|
3 |
+
tags:
|
4 |
+
- generated_from_trainer
|
5 |
+
metrics:
|
6 |
+
- wer
|
7 |
+
- cer
|
8 |
+
model-index:
|
9 |
+
- name: uniTKU-hubert-japanese-asr
|
10 |
+
results:
|
11 |
+
- task:
|
12 |
+
name: Speech Recognition
|
13 |
+
type: automatic-speech-recognition
|
14 |
+
dataset:
|
15 |
+
name: common_voice_11_0
|
16 |
+
type: common_voice
|
17 |
+
args: ja
|
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:
|
28 |
+
- ja
|
29 |
---
|
30 |
|
31 |
+
# uniTKU-hubert-japanese-asr
|
32 |
+
|
33 |
+
This model was fine-tuned on a dataset provided by uniTKU, and it has maintained the original performance metrics on the [common_voice_11_0 dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/viewer/ja).
|
34 |
+
|
35 |
+
This model can only predict Hiragana.
|
36 |
+
|
37 |
+
## Training Procedure
|
38 |
+
|
39 |
+
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 |
+
|
56 |
+
The training hyperparameters remained consistent throughout the fine-tuning process:
|
57 |
+
|
58 |
+
- learning_rate: 1e-4
|
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
|
66 |
+
|
67 |
+
```python
|
68 |
+
from transformers import HubertForCTC, Wav2Vec2Processor
|
69 |
+
from datasets import load_dataset
|
70 |
+
import torch
|
71 |
+
import torchaudio
|
72 |
+
import librosa
|
73 |
+
import numpy as np
|
74 |
+
import re
|
75 |
+
import MeCab
|
76 |
+
import pykakasi
|
77 |
+
from evaluate import load
|
78 |
+
|
79 |
+
model = HubertForCTC.from_pretrained('TKU410410103/uniTKU-hubert-japanese-asr')
|
80 |
+
processor = Wav2Vec2Processor.from_pretrained("TKU410410103/uniTKU-hubert-japanese-asr")
|
81 |
+
|
82 |
+
# load dataset
|
83 |
+
test_dataset = load_dataset('mozilla-foundation/common_voice_11_0', 'ja', split='test')
|
84 |
+
remove_columns = [col for col in test_dataset.column_names if col not in ['audio', 'sentence']]
|
85 |
+
test_dataset = test_dataset.remove_columns(remove_columns)
|
86 |
+
|
87 |
+
# resample
|
88 |
+
def process_waveforms(batch):
|
89 |
+
speech_arrays = []
|
90 |
+
sampling_rates = []
|
91 |
+
|
92 |
+
for audio_path in batch['audio']:
|
93 |
+
speech_array, _ = torchaudio.load(audio_path['path'])
|
94 |
+
speech_array_resampled = librosa.resample(np.asarray(speech_array[0].numpy()), orig_sr=48000, target_sr=16000)
|
95 |
+
speech_arrays.append(speech_array_resampled)
|
96 |
+
sampling_rates.append(16000)
|
97 |
+
|
98 |
+
batch["array"] = speech_arrays
|
99 |
+
batch["sampling_rate"] = sampling_rates
|
100 |
+
|
101 |
+
return batch
|
102 |
+
|
103 |
+
# hiragana
|
104 |
+
CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
|
105 |
+
"؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
|
106 |
+
"{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
|
107 |
+
"、", "﹂", "﹁", "‧", "~", "﹏", ",", "{", "}", "(", ")", "[", "]", "【", "】", "‥", "〽",
|
108 |
+
"『", "』", "〝", "〟", "⟨", "⟩", "〜", ":", "!", "?", "♪", "؛", "/", "\\", "º", "−", "^", "'", "ʻ", "ˆ"]
|
109 |
+
chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
|
110 |
+
|
111 |
+
wakati = MeCab.Tagger("-Owakati")
|
112 |
+
kakasi = pykakasi.kakasi()
|
113 |
+
kakasi.setMode("J","H")
|
114 |
+
kakasi.setMode("K","H")
|
115 |
+
kakasi.setMode("r","Hepburn")
|
116 |
+
conv = kakasi.getConverter()
|
117 |
+
|
118 |
+
def prepare_char(batch):
|
119 |
+
batch["sentence"] = conv.do(wakati.parse(batch["sentence"]).strip())
|
120 |
+
batch["sentence"] = re.sub(chars_to_ignore_regex,'', batch["sentence"]).strip()
|
121 |
+
return batch
|
122 |
+
|
123 |
+
|
124 |
+
resampled_eval_dataset = test_dataset.map(process_waveforms, batched=True, batch_size=50, num_proc=4)
|
125 |
+
eval_dataset = resampled_eval_dataset.map(prepare_char, num_proc=4)
|
126 |
+
|
127 |
+
# begin the evaluation process
|
128 |
+
wer = load("wer")
|
129 |
+
cer = load("cer")
|
130 |
+
|
131 |
+
def evaluate(batch):
|
132 |
+
inputs = processor(batch["array"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
133 |
+
with torch.no_grad():
|
134 |
+
logits = model(inputs.input_values.to(device), attention_mask=inputs.attention_mask.to(device)).logits
|
135 |
+
pred_ids = torch.argmax(logits, dim=-1)
|
136 |
+
batch["pred_strings"] = processor.batch_decode(pred_ids)
|
137 |
+
return batch
|
138 |
+
|
139 |
+
columns_to_remove = [column for column in eval_dataset.column_names if column != "sentence"]
|
140 |
+
batch_size = 16
|
141 |
+
result = eval_dataset.map(evaluate, remove_columns=columns_to_remove, batched=True, batch_size=batch_size)
|
142 |
+
|
143 |
+
wer_result = wer.compute(predictions=result["pred_strings"], references=result["sentence"])
|
144 |
+
cer_result = cer.compute(predictions=result["pred_strings"], references=result["sentence"])
|
145 |
+
|
146 |
+
print("WER: {:2f}%".format(100 * wer_result))
|
147 |
+
print("CER: {:2f}%".format(100 * cer_result))
|
148 |
+
```
|
149 |
+
|
150 |
+
### Test results
|
151 |
+
The final model was evaluated as follows:
|
152 |
+
|
153 |
+
On uniTKU Dataset:
|
154 |
+
- WER: 20.5%
|
155 |
+
|
156 |
+
On common_voice_11_0:
|
157 |
+
- WER: 27.447168%
|
158 |
+
- CER: 11.607944%
|
159 |
+
|
160 |
+
### Framework versions
|
161 |
+
|
162 |
+
- Transformers 4.39.1
|
163 |
+
- Pytorch 2.2.1+cu118
|
164 |
+
- Datasets 2.17.1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|