Update readme for more info
Browse files
README.md
CHANGED
@@ -30,14 +30,16 @@ model-index:
|
|
30 |
|
31 |
---
|
32 |
|
33 |
-
# Wav2Vec2-Large-XLSR-53
|
34 |
|
35 |
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.
|
36 |
|
37 |
## Usage
|
38 |
The model can be used directly (without a language model) as follows:
|
39 |
|
|
|
40 |
```bash
|
|
|
41 |
!pip install git+https://github.com/huggingface/datasets.git
|
42 |
!pip install git+https://github.com/huggingface/transformers.git
|
43 |
!pip install torchaudio
|
@@ -45,52 +47,49 @@ The model can be used directly (without a language model) as follows:
|
|
45 |
!pip install jiwer
|
46 |
```
|
47 |
|
|
|
|
|
48 |
```python
|
|
|
49 |
import torch
|
50 |
import torchaudio
|
51 |
-
from datasets import load_dataset, load_metric
|
52 |
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
|
|
53 |
|
54 |
-
import librosa
|
55 |
-
|
56 |
-
import pandas as pd
|
57 |
import numpy as np
|
58 |
-
|
59 |
-
import random
|
60 |
-
import os
|
61 |
-
import string
|
62 |
-
import six
|
63 |
import re
|
|
|
64 |
|
65 |
import IPython.display as ipd
|
66 |
|
67 |
-
|
68 |
-
dataset = load_dataset("common_voice", "ka", split="test")
|
69 |
-
print(dataset)
|
70 |
-
|
71 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
72 |
-
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian")
|
73 |
-
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian").to(device)
|
74 |
-
|
75 |
-
|
76 |
-
# Preprocessing the datasets.
|
77 |
-
chars_to_ignore_regex = f"""[{"".join([
|
78 |
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "οΏ½",
|
79 |
"#", "!", "?", "Β«", "Β»", "(", ")", "Ψ", ",", "?", ".", "!", "-", ";", ":", '"',
|
80 |
"β", "%", "β", "οΏ½", "β", "β¦", "_", "β", 'β', 'β'
|
81 |
-
]
|
|
|
|
|
|
|
82 |
|
83 |
-
def
|
84 |
-
|
|
|
|
|
|
|
|
|
85 |
return text
|
86 |
|
87 |
-
def normalizer(batch, chars_to_ignore):
|
88 |
-
|
89 |
-
text =
|
|
|
|
|
|
|
|
|
90 |
batch["sentence"] = text
|
91 |
return batch
|
92 |
|
93 |
-
|
94 |
def speech_file_to_array_fn(batch):
|
95 |
speech_array, sampling_rate = torchaudio.load(batch["path"])
|
96 |
speech_array = speech_array.squeeze().numpy()
|
@@ -99,6 +98,7 @@ def speech_file_to_array_fn(batch):
|
|
99 |
batch["speech"] = speech_array
|
100 |
return batch
|
101 |
|
|
|
102 |
def predict(batch):
|
103 |
features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
104 |
|
@@ -113,14 +113,21 @@ def predict(batch):
|
|
113 |
batch["predicted"] = processor.batch_decode(pred_ids)[0]
|
114 |
return batch
|
115 |
|
116 |
-
dataset = dataset.map(normalizer, fn_kwargs={"chars_to_ignore": chars_to_ignore_regex})
|
117 |
-
dataset = dataset.map(speech_file_to_array_fn, remove_columns=list(set(dataset.column_names) - set(['sentence', 'path'])))
|
118 |
-
result = dataset.map(predict)
|
119 |
-
```
|
120 |
|
121 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
122 |
|
123 |
-
```python
|
124 |
max_items = np.random.randint(0, len(result), 20).tolist()
|
125 |
for i in max_items:
|
126 |
reference, predicted = result["sentence"][i], result["predicted"][i]
|
@@ -129,6 +136,7 @@ for i in max_items:
|
|
129 |
print('---')
|
130 |
```
|
131 |
|
|
|
132 |
```text
|
133 |
reference: ααααααα‘α’α ααͺαα£αα αͺααα’α α α₯αααα₯α αααα¨αα
|
134 |
predicted: ααααααα‘α’α ααͺαα£αα αͺααα’α α α₯αααα₯α αααα¨αα
|
@@ -162,14 +170,95 @@ predicted: ααα αααααα αααα‘ α₯αααα₯αα‘ α©α
|
|
162 |
---
|
163 |
```
|
164 |
|
|
|
165 |
## Evaluation
|
166 |
|
|
|
|
|
167 |
```python
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
168 |
wer = load_metric("wer")
|
169 |
|
170 |
print("WER: {:.2f}".format(100 * wer.compute(predictions=result["predicted"], references=result["sentence"])))
|
171 |
```
|
172 |
|
|
|
173 |
**Test Result**:
|
174 |
- WER: 54.00%
|
175 |
|
30 |
|
31 |
---
|
32 |
|
33 |
+
# Wav2Vec2-Large-XLSR-53-Georgian
|
34 |
|
35 |
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.
|
36 |
|
37 |
## Usage
|
38 |
The model can be used directly (without a language model) as follows:
|
39 |
|
40 |
+
**Requirements**
|
41 |
```bash
|
42 |
+
# requirement packages
|
43 |
!pip install git+https://github.com/huggingface/datasets.git
|
44 |
!pip install git+https://github.com/huggingface/transformers.git
|
45 |
!pip install torchaudio
|
47 |
!pip install jiwer
|
48 |
```
|
49 |
|
50 |
+
|
51 |
+
**Prediction**
|
52 |
```python
|
53 |
+
import librosa
|
54 |
import torch
|
55 |
import torchaudio
|
|
|
56 |
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
57 |
+
from datasets import load_dataset
|
58 |
|
|
|
|
|
|
|
59 |
import numpy as np
|
|
|
|
|
|
|
|
|
|
|
60 |
import re
|
61 |
+
import string
|
62 |
|
63 |
import IPython.display as ipd
|
64 |
|
65 |
+
chars_to_ignore = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "οΏ½",
|
67 |
"#", "!", "?", "Β«", "Β»", "(", ")", "Ψ", ",", "?", ".", "!", "-", ";", ":", '"',
|
68 |
"β", "%", "β", "οΏ½", "β", "β¦", "_", "β", 'β', 'β'
|
69 |
+
]
|
70 |
+
chars_to_mapping = {
|
71 |
+
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
|
72 |
+
}
|
73 |
|
74 |
+
def multiple_replace(text, chars_to_mapping):
|
75 |
+
pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
|
76 |
+
return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))
|
77 |
+
|
78 |
+
def remove_special_characters(text, chars_to_ignore_regex):
|
79 |
+
text = re.sub(chars_to_ignore_regex, '', text).lower() + " "
|
80 |
return text
|
81 |
|
82 |
+
def normalizer(batch, chars_to_ignore, chars_to_mapping):
|
83 |
+
chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
|
84 |
+
text = batch["sentence"].lower().strip()
|
85 |
+
|
86 |
+
text = multiple_replace(text, chars_to_mapping)
|
87 |
+
text = remove_special_characters(text, chars_to_ignore_regex)
|
88 |
+
|
89 |
batch["sentence"] = text
|
90 |
return batch
|
91 |
|
92 |
+
|
93 |
def speech_file_to_array_fn(batch):
|
94 |
speech_array, sampling_rate = torchaudio.load(batch["path"])
|
95 |
speech_array = speech_array.squeeze().numpy()
|
98 |
batch["speech"] = speech_array
|
99 |
return batch
|
100 |
|
101 |
+
|
102 |
def predict(batch):
|
103 |
features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
104 |
|
113 |
batch["predicted"] = processor.batch_decode(pred_ids)[0]
|
114 |
return batch
|
115 |
|
|
|
|
|
|
|
|
|
116 |
|
117 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
118 |
+
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian")
|
119 |
+
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian").to(device)
|
120 |
+
|
121 |
+
dataset = load_dataset("common_voice", "ka", split="test[:1%]")
|
122 |
+
dataset = dataset.map(
|
123 |
+
normalizer,
|
124 |
+
fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
|
125 |
+
remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
|
126 |
+
)
|
127 |
+
|
128 |
+
dataset = dataset.map(speech_file_to_array_fn)
|
129 |
+
result = dataset.map(predict)
|
130 |
|
|
|
131 |
max_items = np.random.randint(0, len(result), 20).tolist()
|
132 |
for i in max_items:
|
133 |
reference, predicted = result["sentence"][i], result["predicted"][i]
|
136 |
print('---')
|
137 |
```
|
138 |
|
139 |
+
**Output:**
|
140 |
```text
|
141 |
reference: ααααααα‘α’α ααͺαα£αα αͺααα’α α α₯αααα₯α αααα¨αα
|
142 |
predicted: ααααααα‘α’α ααͺαα£αα αͺααα’α α α₯αααα₯α αααα¨αα
|
170 |
---
|
171 |
```
|
172 |
|
173 |
+
|
174 |
## Evaluation
|
175 |
|
176 |
+
The model can be evaluated as follows on the Georgian test data of Common Voice.
|
177 |
+
|
178 |
```python
|
179 |
+
import librosa
|
180 |
+
import torch
|
181 |
+
import torchaudio
|
182 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
183 |
+
from datasets import load_dataset, load_metric
|
184 |
+
|
185 |
+
import numpy as np
|
186 |
+
import re
|
187 |
+
import string
|
188 |
+
|
189 |
+
|
190 |
+
chars_to_ignore = [
|
191 |
+
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "οΏ½",
|
192 |
+
"#", "!", "?", "Β«", "Β»", "(", ")", "Ψ", ",", "?", ".", "!", "-", ";", ":", '"',
|
193 |
+
"β", "%", "β", "οΏ½", "β", "β¦", "_", "β", 'β', 'β'
|
194 |
+
]
|
195 |
+
chars_to_mapping = {
|
196 |
+
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ",
|
197 |
+
}
|
198 |
+
|
199 |
+
def multiple_replace(text, chars_to_mapping):
|
200 |
+
pattern = "|".join(map(re.escape, chars_to_mapping.keys()))
|
201 |
+
return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text))
|
202 |
+
|
203 |
+
def remove_special_characters(text, chars_to_ignore_regex):
|
204 |
+
text = re.sub(chars_to_ignore_regex, '', text).lower() + " "
|
205 |
+
return text
|
206 |
+
|
207 |
+
def normalizer(batch, chars_to_ignore, chars_to_mapping):
|
208 |
+
chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]"""
|
209 |
+
text = batch["sentence"].lower().strip()
|
210 |
+
|
211 |
+
text = multiple_replace(text, chars_to_mapping)
|
212 |
+
text = remove_special_characters(text, chars_to_ignore_regex)
|
213 |
+
|
214 |
+
batch["sentence"] = text
|
215 |
+
return batch
|
216 |
+
|
217 |
+
|
218 |
+
def speech_file_to_array_fn(batch):
|
219 |
+
speech_array, sampling_rate = torchaudio.load(batch["path"])
|
220 |
+
speech_array = speech_array.squeeze().numpy()
|
221 |
+
speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000)
|
222 |
+
|
223 |
+
batch["speech"] = speech_array
|
224 |
+
return batch
|
225 |
+
|
226 |
+
|
227 |
+
def predict(batch):
|
228 |
+
features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
229 |
+
|
230 |
+
input_values = features.input_values.to(device)
|
231 |
+
attention_mask = features.attention_mask.to(device)
|
232 |
+
|
233 |
+
with torch.no_grad():
|
234 |
+
logits = model(input_values, attention_mask=attention_mask).logits
|
235 |
+
|
236 |
+
pred_ids = torch.argmax(logits, dim=-1)
|
237 |
+
|
238 |
+
batch["predicted"] = processor.batch_decode(pred_ids)[0]
|
239 |
+
return batch
|
240 |
+
|
241 |
+
|
242 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
243 |
+
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian")
|
244 |
+
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian").to(device)
|
245 |
+
|
246 |
+
dataset = load_dataset("common_voice", "ka", split="test[:1%]")
|
247 |
+
dataset = dataset.map(
|
248 |
+
normalizer,
|
249 |
+
fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping},
|
250 |
+
remove_columns=list(set(dataset.column_names) - set(['sentence', 'path']))
|
251 |
+
)
|
252 |
+
|
253 |
+
dataset = dataset.map(speech_file_to_array_fn)
|
254 |
+
result = dataset.map(predict)
|
255 |
+
|
256 |
wer = load_metric("wer")
|
257 |
|
258 |
print("WER: {:.2f}".format(100 * wer.compute(predictions=result["predicted"], references=result["sentence"])))
|
259 |
```
|
260 |
|
261 |
+
|
262 |
**Test Result**:
|
263 |
- WER: 54.00%
|
264 |
|