Update readme for more info
Browse files
README.md
CHANGED
@@ -26,21 +26,21 @@ model-index:
|
|
26 |
metrics:
|
27 |
- name: Test WER
|
28 |
type: wer
|
29 |
-
value: 32.
|
30 |
- name: Test CER
|
31 |
type: cer
|
32 |
-
value: 8.
|
33 |
|
34 |
---
|
35 |
|
36 |
-
# Wav2Vec2-Large-XLSR-53
|
37 |
|
38 |
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Persian (Farsi) using [Common Voice](https://huggingface.co/datasets/common_voice). When using this model, make sure that your speech input is sampled at 16kHz.
|
39 |
|
40 |
-
##
|
41 |
The model can be used directly (without a language model) as follows:
|
42 |
|
43 |
-
|
44 |
```bash
|
45 |
# requirement packages
|
46 |
!pip install git+https://github.com/huggingface/datasets.git
|
@@ -51,23 +51,28 @@ The model can be used directly (without a language model) as follows:
|
|
51 |
!pip install hazm
|
52 |
```
|
53 |
|
54 |
-
### Preprocessing
|
55 |
|
|
|
56 |
```python
|
57 |
-
|
58 |
-
|
|
|
|
|
|
|
59 |
|
|
|
60 |
import hazm
|
61 |
import re
|
62 |
import string
|
63 |
|
|
|
64 |
|
65 |
_normalizer = hazm.Normalizer()
|
66 |
-
chars_to_ignore =
|
67 |
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
|
68 |
"#", "!", "؟", "?", "«", "»", "ء", "،", "(", ")", "؛", "'ٔ", "٬",'ٔ', ",", "?",
|
69 |
".", "!", "-", ";", ":",'"',"“", "%", "‘", "”", "�", "–", "…", "_", "”", '“', '„'
|
70 |
-
]
|
71 |
|
72 |
# In case of farsi
|
73 |
chars_to_ignore = chars_to_ignore + list(string.ascii_lowercase + string.digits)
|
@@ -100,37 +105,6 @@ def normalizer(batch, chars_to_ignore, chars_to_mapping):
|
|
100 |
|
101 |
batch["sentence"] = text
|
102 |
return batch
|
103 |
-
```
|
104 |
-
|
105 |
-
### Loading The Data
|
106 |
-
|
107 |
-
```python
|
108 |
-
from datasets import load_dataset
|
109 |
-
|
110 |
-
dataset = load_dataset("common_voice", "fa", split="test[:1%]")
|
111 |
-
print(dataset)
|
112 |
-
```
|
113 |
-
|
114 |
-
**Output:**
|
115 |
-
```text
|
116 |
-
>>>
|
117 |
-
Dataset({
|
118 |
-
features: ['client_id', 'path', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],
|
119 |
-
num_rows: 52
|
120 |
-
})
|
121 |
-
```
|
122 |
-
|
123 |
-
### Model
|
124 |
-
|
125 |
-
```python
|
126 |
-
import librosa
|
127 |
-
import torch
|
128 |
-
import torchaudio
|
129 |
-
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
130 |
-
|
131 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
132 |
-
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian")
|
133 |
-
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian").to(device)
|
134 |
|
135 |
|
136 |
def speech_file_to_array_fn(batch):
|
@@ -155,13 +129,11 @@ def predict(batch):
|
|
155 |
|
156 |
batch["predicted"] = processor.batch_decode(pred_ids)[0]
|
157 |
return batch
|
158 |
-
```
|
159 |
|
160 |
-
## Prediction
|
161 |
-
|
162 |
-
```python
|
163 |
-
import IPython.display as ipd
|
164 |
|
|
|
|
|
|
|
165 |
|
166 |
dataset = load_dataset("common_voice", "fa", split="test[:1%]")
|
167 |
dataset = dataset.map(
|
@@ -246,13 +218,91 @@ predicted: من سفر کردم را دوست دارم
|
|
246 |
|
247 |
## Evaluation
|
248 |
|
|
|
|
|
249 |
```bash
|
250 |
-
mkdir cer
|
251 |
-
wget -O cer/cer.py https://huggingface.co/ctl/wav2vec2-large-xlsr-cantonese/raw/main/cer.py
|
252 |
```
|
253 |
|
254 |
```python
|
255 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
256 |
|
257 |
dataset = load_dataset("common_voice", "fa", split="test")
|
258 |
dataset = dataset.map(
|
@@ -270,11 +320,9 @@ print("WER: {:.2f}".format(100 * wer.compute(predictions=result["predicted"], re
|
|
270 |
print("CER: {:.2f}".format(100 * cer.compute(predictions=result["predicted"], references=result["sentence"])))
|
271 |
```
|
272 |
|
273 |
-
**
|
274 |
-
|
275 |
-
|
276 |
-
CER: 8.23%
|
277 |
-
```
|
278 |
|
279 |
|
280 |
## Training
|
|
|
26 |
metrics:
|
27 |
- name: Test WER
|
28 |
type: wer
|
29 |
+
value: 32.18
|
30 |
- name: Test CER
|
31 |
type: cer
|
32 |
+
value: 8.27
|
33 |
|
34 |
---
|
35 |
|
36 |
+
# Wav2Vec2-Large-XLSR-53-Persian
|
37 |
|
38 |
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Persian (Farsi) using [Common Voice](https://huggingface.co/datasets/common_voice). When using this model, make sure that your speech input is sampled at 16kHz.
|
39 |
|
40 |
+
## Usage
|
41 |
The model can be used directly (without a language model) as follows:
|
42 |
|
43 |
+
**Requirements**
|
44 |
```bash
|
45 |
# requirement packages
|
46 |
!pip install git+https://github.com/huggingface/datasets.git
|
|
|
51 |
!pip install hazm
|
52 |
```
|
53 |
|
|
|
54 |
|
55 |
+
**Prediction**
|
56 |
```python
|
57 |
+
import librosa
|
58 |
+
import torch
|
59 |
+
import torchaudio
|
60 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
61 |
+
from datasets import load_dataset
|
62 |
|
63 |
+
import numpy as np
|
64 |
import hazm
|
65 |
import re
|
66 |
import string
|
67 |
|
68 |
+
import IPython.display as ipd
|
69 |
|
70 |
_normalizer = hazm.Normalizer()
|
71 |
+
chars_to_ignore = [
|
72 |
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
|
73 |
"#", "!", "؟", "?", "«", "»", "ء", "،", "(", ")", "؛", "'ٔ", "٬",'ٔ', ",", "?",
|
74 |
".", "!", "-", ";", ":",'"',"“", "%", "‘", "”", "�", "–", "…", "_", "”", '“', '„'
|
75 |
+
]
|
76 |
|
77 |
# In case of farsi
|
78 |
chars_to_ignore = chars_to_ignore + list(string.ascii_lowercase + string.digits)
|
|
|
105 |
|
106 |
batch["sentence"] = text
|
107 |
return batch
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
|
109 |
|
110 |
def speech_file_to_array_fn(batch):
|
|
|
129 |
|
130 |
batch["predicted"] = processor.batch_decode(pred_ids)[0]
|
131 |
return batch
|
|
|
132 |
|
|
|
|
|
|
|
|
|
133 |
|
134 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
135 |
+
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian")
|
136 |
+
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian").to(device)
|
137 |
|
138 |
dataset = load_dataset("common_voice", "fa", split="test[:1%]")
|
139 |
dataset = dataset.map(
|
|
|
218 |
|
219 |
## Evaluation
|
220 |
|
221 |
+
The model can be evaluated as follows on the Persian (Farsi) test data of Common Voice.
|
222 |
+
|
223 |
```bash
|
224 |
+
!mkdir cer
|
225 |
+
!wget -O cer/cer.py https://huggingface.co/ctl/wav2vec2-large-xlsr-cantonese/raw/main/cer.py
|
226 |
```
|
227 |
|
228 |
```python
|
229 |
+
import librosa
|
230 |
+
import torch
|
231 |
+
import torchaudio
|
232 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
233 |
+
from datasets import load_dataset, load_metric
|
234 |
+
|
235 |
+
import hazm
|
236 |
+
import re
|
237 |
+
import string
|
238 |
+
|
239 |
+
_normalizer = hazm.Normalizer()
|
240 |
+
chars_to_ignore = [
|
241 |
+
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�",
|
242 |
+
"#", "!", "؟", "?", "«", "»", "ء", "،", "(", ")", "؛", "'ٔ", "٬",'ٔ', ",", "?",
|
243 |
+
".", "!", "-", ";", ":",'"',"“", "%", "‘", "”", "�", "–", "…", "_", "”", '“', '„'
|
244 |
+
]
|
245 |
+
|
246 |
+
# In case of farsi
|
247 |
+
# chars_to_ignore = chars_to_ignore + list(string.ascii_lowercase + string.digits)
|
248 |
+
|
249 |
+
chars_to_mapping = {
|
250 |
+
'ك': 'ک', 'دِ': 'د', 'بِ': 'ب', 'زِ': 'ز', 'ذِ': 'ذ', 'شِ': 'ش', 'سِ': 'س', 'ى': 'ی',
|
251 |
+
'ي': 'ی', 'أ': 'ا', 'ؤ': 'و', "ے": "ی", "ۀ": "ه", "ﭘ": "پ", "ﮐ": "ک", "ﯽ": "ی",
|
252 |
+
"ﺎ": "ا", "ﺑ": "ب", "ﺘ": "ت", "ﺧ": "خ", "ﺩ": "د", "ﺱ": "س", "ﻀ": "ض", "ﻌ": "ع",
|
253 |
+
"ﻟ": "ل", "ﻡ": "م", "ﻢ": "م", "ﻪ": "ه", "ﻮ": "و", "ئ": "ی", 'ﺍ': "ا", 'ة': "ه",
|
254 |
+
'ﯾ': "ی", 'ﯿ': "ی", 'ﺒ': "ب", 'ﺖ': "ت", 'ﺪ': "د", 'ﺮ': "ر", 'ﺴ': "س", 'ﺷ': "ش",
|
255 |
+
'ﺸ': "ش", 'ﻋ': "ع", 'ﻤ': "م", 'ﻥ': "ن", 'ﻧ': "ن", 'ﻭ': "و", 'ﺭ': "ر", "ﮔ": "گ",
|
256 |
+
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
|
257 |
+
}
|
258 |
+
|
259 |
+
def multiple_replace(text, chars_to_mapping):
|
260 |
+
pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
|
261 |
+
return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))
|
262 |
+
|
263 |
+
def remove_special_characters(text, chars_to_ignore_regex):
|
264 |
+
text = re.sub(chars_to_ignore_regex, '', text).lower() + " "
|
265 |
+
return text
|
266 |
+
|
267 |
+
def normalizer(batch, chars_to_ignore, chars_to_mapping):
|
268 |
+
chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
|
269 |
+
text = batch["sentence"].lower().strip()
|
270 |
+
|
271 |
+
text = _normalizer.normalize(text)
|
272 |
+
text = multiple_replace(text, chars_to_mapping)
|
273 |
+
text = remove_special_characters(text, chars_to_ignore_regex)
|
274 |
+
|
275 |
+
batch["sentence"] = text
|
276 |
+
return batch
|
277 |
+
|
278 |
+
|
279 |
+
def speech_file_to_array_fn(batch):
|
280 |
+
speech_array, sampling_rate = torchaudio.load(batch["path"])
|
281 |
+
speech_array = speech_array.squeeze().numpy()
|
282 |
+
speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)
|
283 |
+
|
284 |
+
batch["speech"] = speech_array
|
285 |
+
return batch
|
286 |
+
|
287 |
+
|
288 |
+
def predict(batch):
|
289 |
+
features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
290 |
+
|
291 |
+
input_values = features.input_values.to(device)
|
292 |
+
attention_mask = features.attention_mask.to(device)
|
293 |
+
|
294 |
+
with torch.no_grad():
|
295 |
+
logits = model(input_values, attention_mask=attention_mask).logits
|
296 |
+
|
297 |
+
pred_ids = torch.argmax(logits, dim=-1)
|
298 |
+
|
299 |
+
batch["predicted"] = processor.batch_decode(pred_ids)[0]
|
300 |
+
return batch
|
301 |
+
|
302 |
+
|
303 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
304 |
+
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian")
|
305 |
+
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-persian").to(device)
|
306 |
|
307 |
dataset = load_dataset("common_voice", "fa", split="test")
|
308 |
dataset = dataset.map(
|
|
|
320 |
print("CER: {:.2f}".format(100 * cer.compute(predictions=result["predicted"], references=result["sentence"])))
|
321 |
```
|
322 |
|
323 |
+
**Test Result:**
|
324 |
+
- WER: 32.18%
|
325 |
+
- CER: 8.27%
|
|
|
|
|
326 |
|
327 |
|
328 |
## Training
|