File size: 10,993 Bytes
acdefb8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2f546f1
acdefb8
 
 
 
 
 
2f546f1
acdefb8
2f546f1
acdefb8
 
 
 
 
 
 
2f546f1
 
acdefb8
2f546f1
acdefb8
 
 
2f546f1
acdefb8
 
 
2f546f1
acdefb8
 
 
2f546f1
acdefb8
 
 
2f546f1
 
 
 
acdefb8
2f546f1
 
 
 
 
 
acdefb8
 
2f546f1
 
 
 
 
 
 
acdefb8
 
 
2f546f1
acdefb8
 
 
 
 
 
 
 
2f546f1
acdefb8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2f546f1
 
 
 
 
 
 
 
 
 
 
 
 
acdefb8
 
 
 
 
 
 
 
 
2f546f1
acdefb8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2f546f1
acdefb8
 
2f546f1
 
acdefb8
2f546f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
72cabd7
2f546f1
 
 
 
 
 
 
 
 
acdefb8
 
 
 
 
2f546f1
acdefb8
 
 
 
 
 
 
 
 
 
1
2
3
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
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
---
language: ka
datasets:
- common_voice
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
widget:
- label: Common Voice sample 566
  src: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-georgian/resolve/main/sample566.flac
- label: Common Voice sample 95
  src: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-georgian/resolve/main/sample95.flac
model-index:
- name: XLSR Wav2Vec2 Georgian by Mehrdad Farahani
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Common Voice ka
      type: common_voice
      args: ka
    metrics:
       - name: Test WER
         type: wer
         value: 54.00
        
---

# Wav2Vec2-Large-XLSR-53-Georgian

Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Georgian using [Common Voice](https://huggingface.co/datasets/common_voice). When using this model, make sure that your speech input is sampled at 16kHz.

## Usage
The model can be used directly (without a language model) as follows:

**Requirements**
```bash
# requirement packages
!pip install git+https://github.com/huggingface/datasets.git
!pip install git+https://github.com/huggingface/transformers.git
!pip install torchaudio
!pip install librosa
!pip install jiwer
```


**Prediction**
```python
import librosa
import torch
import torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from datasets import load_dataset

import numpy as np
import re
import string

import IPython.display as ipd

chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "๏ฟฝ",
    "#", "!", "?", "ยซ", "ยป", "(", ")", "ุ›", ",", "?", ".", "!", "-", ";", ":", '"', 
    "โ€œ", "%", "โ€˜", "๏ฟฝ", "โ€“", "โ€ฆ", "_", "โ€", 'โ€œ', 'โ€ž'
]
chars_to_mapping = {
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
}

def multiple_replace(text, chars_to_mapping):
    pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
    return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))

def remove_special_characters(text, chars_to_ignore_regex):
    text = re.sub(chars_to_ignore_regex, '', text).lower() + " "
    return text

def normalizer(batch, chars_to_ignore, chars_to_mapping):
    chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
    text = batch["sentence"].lower().strip()
    
    text = multiple_replace(text, chars_to_mapping)
    text = remove_special_characters(text, chars_to_ignore_regex)

    batch["sentence"] = text
    return batch


def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array = speech_array.squeeze().numpy()
    speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)

    batch["speech"] = speech_array
    return batch


def predict(batch):
    features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    input_values = features.input_values.to(device)
    attention_mask = features.attention_mask.to(device)

    with torch.no_grad():
        logits = model(input_values, attention_mask=attention_mask).logits 
        
    pred_ids = torch.argmax(logits, dim=-1)

    batch["predicted"] = processor.batch_decode(pred_ids)[0]
    return batch


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian").to(device)

dataset = load_dataset("common_voice", "ka", split="test[:1%]")
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
    remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict)

max_items = np.random.randint(0, len(result), 20).tolist()
for i in max_items:
    reference, predicted =  result["sentence"][i], result["predicted"][i]
    print("reference:", reference)
    print("predicted:", predicted)
    print('---')
```

**Output:**
```text
reference: แƒแƒ“แƒ›แƒ˜แƒœแƒ˜แƒกแƒขแƒ แƒแƒชแƒ˜แƒฃแƒšแƒ˜ แƒชแƒ”แƒœแƒขแƒ แƒ˜ แƒฅแƒแƒšแƒแƒฅแƒ˜ แƒ˜แƒ›แƒ˜แƒจแƒšแƒ˜ 
predicted: แƒแƒ“แƒ›แƒ˜แƒœแƒ˜แƒกแƒขแƒ แƒแƒชแƒ˜แƒฃแƒšแƒ˜ แƒชแƒ”แƒœแƒขแƒ แƒ˜ แƒฅแƒแƒšแƒแƒฅแƒ˜ แƒ˜แƒ›แƒ˜แƒจแƒšแƒ˜
---
reference: แƒ“แƒแƒ˜แƒ‘แƒแƒ“แƒ แƒแƒ“แƒ•แƒแƒ™แƒแƒขแƒ˜แƒก แƒแƒฏแƒแƒฎแƒจแƒ˜ 
predicted: แƒแƒ˜แƒ‘แƒแƒ“แƒ แƒแƒ“แƒ›แƒแƒ™แƒแƒขแƒ˜แƒก แƒแƒฏแƒแƒฎแƒจแƒ˜
---
reference: แƒแƒฆแƒกแƒแƒœแƒ˜แƒจแƒœแƒแƒ•แƒ˜แƒ แƒ แƒแƒ› แƒกแƒ˜แƒ›แƒฆแƒ”แƒ แƒ แƒฌแƒแƒ แƒ›แƒแƒแƒ“แƒ’แƒ”แƒœแƒก แƒžแƒแƒš แƒ›แƒแƒ™แƒ™แƒแƒ แƒขแƒœแƒ˜แƒกแƒ แƒ“แƒ แƒฏแƒแƒ แƒฏ แƒฐแƒแƒ แƒ˜แƒกแƒแƒœแƒ˜แƒก แƒ˜แƒจแƒ•แƒ˜แƒแƒ— แƒ•แƒแƒ™แƒแƒšแƒฃแƒ  แƒ“แƒฃแƒ”แƒขแƒก 
predicted: แƒแƒฆแƒกแƒ”แƒœแƒ˜แƒจแƒœแƒแƒ•แƒ˜แƒแƒ แƒ แƒกแƒ˜แƒ›แƒฆแƒ” แƒ แƒแƒฌแƒแƒ แƒ›แƒแƒแƒ“แƒ’แƒ”แƒ›แƒก แƒ‘แƒแƒš แƒ›แƒแƒ™แƒแƒ แƒ“แƒœแƒ˜แƒก แƒ“แƒ แƒฏแƒแƒ แƒฉแƒฎแƒแƒ แƒ˜แƒกแƒแƒœแƒ˜แƒก แƒ˜แƒจแƒ•แƒ˜แƒแƒ“ แƒ•แƒแƒ™แƒแƒšแƒฃแƒ  แƒ“แƒฃแƒ”แƒ—แƒก
---
reference: แƒ˜แƒ™แƒ แƒซแƒแƒšแƒ”แƒ‘แƒแƒ“แƒ แƒฌแƒ˜แƒ แƒ•แƒแƒšแƒแƒชแƒ•แƒ แƒฅแƒแƒ แƒ—แƒฃแƒš แƒ”แƒœแƒแƒ–แƒ” 
predicted: แƒ˜แƒ™แƒ แƒซแƒแƒšแƒ”แƒ‘แƒแƒ“แƒ” แƒฌแƒ˜แƒ แƒ•แƒ แƒšแƒแƒชแƒ•แƒ แƒฅแƒแƒ แƒ—แƒฃแƒš แƒ”แƒœแƒแƒ–แƒ”
---
reference: แƒแƒฆแƒ›แƒแƒ แƒ—แƒฃแƒšแƒ˜แƒ แƒ•แƒแƒšแƒ”แƒกแƒ แƒ“แƒ แƒ‘แƒ”แƒ แƒœแƒ˜แƒก แƒ™แƒแƒœแƒขแƒแƒœแƒ”แƒ‘แƒ˜แƒก แƒกแƒแƒ–แƒฆแƒ•แƒแƒ แƒ–แƒ” 
predicted: แƒแƒฆแƒ›แƒแƒ แƒ—แƒฃแƒšแƒ˜แƒ แƒ•แƒแƒšแƒ”แƒกแƒ แƒ“แƒ แƒ‘แƒ”แƒ แƒœแƒ˜แƒก แƒ™แƒแƒœแƒ—แƒแƒœแƒ”แƒ‘แƒ˜แƒก แƒกแƒแƒ–แƒฆแƒ•แƒแƒ แƒ–แƒ”
---
reference: แƒแƒฅ แƒ˜แƒ’แƒ˜ แƒ›แƒ˜แƒ˜แƒฌแƒ•แƒ˜แƒ”แƒก แƒกแƒแƒ›แƒฎแƒแƒขแƒ•แƒ แƒ แƒแƒ™แƒแƒ“แƒ”แƒ›แƒ˜แƒแƒจแƒ˜ แƒกแƒแƒ“แƒแƒช แƒกแƒ˜แƒชแƒแƒชแƒฎแƒšแƒ˜แƒก แƒ‘แƒแƒšแƒแƒ›แƒ“แƒ” แƒ”แƒฌแƒ”แƒแƒ“แƒ แƒžแƒ”แƒ“แƒแƒ’แƒแƒ’แƒ˜แƒฃแƒ  แƒ›แƒแƒฆแƒ•แƒแƒฌแƒ”แƒแƒ‘แƒแƒก 
predicted: แƒแƒฅ แƒ˜แƒ’แƒ˜ แƒ›แƒ˜แƒ˜แƒกแƒฌแƒ แƒ•แƒ˜แƒ”แƒก แƒกแƒแƒ›แƒฎแƒแƒขแƒ แƒ แƒแƒ™แƒแƒ“แƒ”แƒ›แƒ˜ แƒแƒจแƒ˜แƒกแƒ แƒ“แƒ แƒชแƒ˜แƒชแƒแƒชแƒฎแƒšแƒ˜แƒก แƒ‘แƒแƒšแƒแƒ›แƒ“แƒ” แƒ”แƒฌแƒงแƒ”แƒ‘แƒแƒ‘ แƒ“แƒ แƒžแƒ”แƒ“แƒแƒ’แƒฃแƒ“แƒ˜แƒ•แƒ˜แƒ  แƒ›แƒแƒงแƒ•แƒแƒฌแƒ”แƒ•แƒ”แƒ‘แƒแƒก
---
reference: แƒ™แƒšแƒแƒ แƒ˜แƒกแƒ แƒ—แƒแƒœแƒฎแƒ›แƒ“แƒ”แƒ‘แƒ แƒจแƒ”แƒ›แƒแƒ—แƒแƒ•แƒแƒ–แƒ”แƒ‘แƒแƒ–แƒ” แƒ“แƒ แƒšแƒ”แƒฅแƒขแƒ”แƒ แƒ˜แƒก แƒ“แƒแƒฎแƒ›แƒแƒ แƒ”แƒ‘แƒ˜แƒ— แƒกแƒ”แƒ แƒ˜แƒฃแƒšแƒ˜ แƒ›แƒ™แƒ•แƒšแƒ”แƒšแƒ˜แƒก แƒ™แƒ•แƒแƒšแƒก แƒ“แƒแƒแƒ“แƒ’แƒ”แƒ‘แƒ 
predicted: แƒ™แƒšแƒแƒ แƒ˜แƒก แƒ—แƒแƒœ แƒฎแƒ•แƒ“แƒ”แƒ‘แƒ แƒจแƒ”แƒ›แƒฃแƒ—แƒแƒ•แƒแƒ–แƒ” แƒ‘แƒแƒ–แƒ” แƒ“แƒ แƒšแƒ”แƒฅแƒขแƒ”แƒ แƒ˜แƒก แƒ“แƒแƒฎแƒ›แƒแƒ แƒ”แƒ‘แƒ˜แƒช แƒกแƒ”แƒ แƒ˜แƒฃแƒ แƒ˜ แƒ›แƒ™แƒ•แƒšแƒ”แƒšแƒ˜แƒก แƒ™แƒ•แƒ”แƒšแƒก แƒ“แƒแƒแƒ“แƒ’แƒ”แƒ‘แƒแƒ
---
reference: แƒ˜แƒ‘แƒ แƒซแƒแƒ“แƒ แƒขแƒงแƒ•แƒ”แƒ”แƒ‘แƒ˜แƒ— แƒ•แƒแƒญแƒ แƒแƒ‘แƒ˜แƒก แƒฌแƒ˜แƒœแƒแƒแƒฆแƒ›แƒ“แƒ”แƒ’ 
predicted: แƒ“แƒ˜แƒ‘แƒ แƒซแƒแƒขแƒ แƒขแƒงแƒ•แƒ”แƒ”แƒ‘แƒ˜แƒ— แƒ•แƒแƒญแƒ แƒแƒ‘แƒ˜แƒก แƒฌแƒ˜แƒœแƒแƒแƒฆแƒ“แƒ”แƒ’
---
reference: แƒกแƒแƒ—แƒแƒ•แƒกแƒก แƒแƒฆแƒ›แƒแƒกแƒแƒ•แƒšแƒ”แƒ—แƒ˜แƒ— แƒ“แƒ แƒ“แƒแƒกแƒแƒ•แƒšแƒ”แƒ—แƒ˜แƒ— แƒ—แƒ˜แƒ—แƒ แƒกแƒแƒ แƒ™แƒ›แƒ”แƒšแƒ˜ แƒแƒฅแƒ•แƒก 
predicted: แƒกแƒแƒ—แƒแƒ•แƒก แƒแƒฆแƒ›แƒแƒกแƒแƒ•แƒ”แƒšแƒ”แƒ—แƒ˜ แƒ“แƒ แƒ“แƒแƒกแƒแƒ•แƒšแƒ”แƒ— แƒ›แƒ—แƒ˜แƒ“แƒ แƒกแƒแƒ แƒ™แƒ›แƒ”แƒšแƒ˜ แƒแƒฅแƒ•แƒก
---
reference: แƒ˜แƒ’แƒ˜ แƒ›แƒ“แƒ”แƒ‘แƒแƒ แƒ”แƒแƒ‘แƒก แƒฅแƒแƒšแƒแƒฅแƒ˜แƒก แƒฉแƒ แƒ“แƒ˜แƒšแƒแƒแƒฆแƒ›แƒแƒกแƒแƒ•แƒšแƒ”แƒ— แƒœแƒแƒฌแƒ˜แƒšแƒจแƒ˜ 
predicted: แƒ˜แƒ’แƒ˜ แƒ›แƒ“แƒ”แƒ‘แƒแƒ แƒ”แƒแƒ‘แƒก แƒฅแƒแƒšแƒแƒฅแƒ˜แƒก แƒฉแƒ แƒ“แƒ˜แƒšแƒ แƒแƒฆแƒ›แƒแƒกแƒแƒ•แƒšแƒ”แƒ— แƒœแƒแƒฌแƒ˜แƒšแƒจแƒ˜
---
```


## Evaluation

The model can be evaluated as follows on the Georgian test data of Common Voice.

```python
import librosa
import torch
import torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from datasets import load_dataset, load_metric

import numpy as np
import re
import string


chars_to_ignore = [
    ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "๏ฟฝ",
    "#", "!", "?", "ยซ", "ยป", "(", ")", "ุ›", ",", "?", ".", "!", "-", ";", ":", '"', 
    "โ€œ", "%", "โ€˜", "๏ฟฝ", "โ€“", "โ€ฆ", "_", "โ€", 'โ€œ', 'โ€ž'
]
chars_to_mapping = {
    "\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
}

def multiple_replace(text, chars_to_mapping):
    pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
    return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))

def remove_special_characters(text, chars_to_ignore_regex):
    text = re.sub(chars_to_ignore_regex, '', text).lower() + " "
    return text

def normalizer(batch, chars_to_ignore, chars_to_mapping):
    chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
    text = batch["sentence"].lower().strip()
    
    text = multiple_replace(text, chars_to_mapping)
    text = remove_special_characters(text, chars_to_ignore_regex)

    batch["sentence"] = text
    return batch


def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    speech_array = speech_array.squeeze().numpy()
    speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)

    batch["speech"] = speech_array
    return batch


def predict(batch):
    features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    input_values = features.input_values.to(device)
    attention_mask = features.attention_mask.to(device)

    with torch.no_grad():
        logits = model(input_values, attention_mask=attention_mask).logits 
        
    pred_ids = torch.argmax(logits, dim=-1)

    batch["predicted"] = processor.batch_decode(pred_ids)[0]
    return batch


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian")
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian").to(device)

dataset = load_dataset("common_voice", "ka", split="test")
dataset = dataset.map(
    normalizer, 
    fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
    remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
)

dataset = dataset.map(speech_file_to_array_fn)
result = dataset.map(predict)

wer = load_metric("wer")

print("WER: {:.2f}".format(100 * wer.compute(predictions=result["predicted"], references=result["sentence"])))
```


**Test Result**: 
- WER: 54.00%


## Training & Report
The Common Voice `train`, `validation` datasets were used for training.

You can see the training states [here](https://wandb.ai/m3hrdadfi/finetuned_wav2vec_xlsr_georgian/reports/Fine-Tuning-for-Wav2Vec2-Large-XLSR-53-Georgian--Vmlldzo1NTg5MDQ?accessToken=rsmd0p83iln13yq23b9kzj8bim6nco21w8cqn2tb19v51okakqk92c71h6hbxmfj)

The script used for training can be found [here](https://colab.research.google.com/github/m3hrdadfi/notebooks/blob/main/Fine_Tune_XLSR_Wav2Vec2_on_Georgian_ASR_with_%F0%9F%A4%97_Transformers_ipynb.ipynb)