Update README.md
Browse files
README.md
CHANGED
@@ -30,7 +30,7 @@ model-index:
|
|
30 |
|
31 |
# Wav2Vec2-Large-XLSR-53-khmer
|
32 |
|
33 |
-
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Khmer using the [Common Voice](https://huggingface.co/datasets/common_voice), and [OpenSLR Kh](http://www.openslr.org/
|
34 |
|
35 |
When using this model, make sure that your speech input is sampled at 16kHz.
|
36 |
|
@@ -44,12 +44,12 @@ import torchaudio
|
|
44 |
from datasets import load_dataset
|
45 |
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
46 |
|
47 |
-
!wget https://www.openslr.org/resources/
|
48 |
!unzip km_kh_male.zip
|
49 |
!ls km_kh_male
|
50 |
|
51 |
colnames=['path','sentence']
|
52 |
-
df = pd.read_csv('/content/km_kh_male/line_index.tsv',sep='
|
53 |
df['path'] = '/content/km_kh_male/wavs/'+df['path'] +'.wav'
|
54 |
|
55 |
train, test = train_test_split(df, test_size=0.1)
|
@@ -66,15 +66,15 @@ resampler = torchaudio.transforms.Resample(48_000, 16_000)
|
|
66 |
# Preprocessing the datasets.
|
67 |
# We need to read the aduio files as arrays
|
68 |
def speech_file_to_array_fn(batch):
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
|
73 |
test_dataset = test_dataset.map(speech_file_to_array_fn)
|
74 |
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
|
75 |
|
76 |
with torch.no_grad():
|
77 |
-
|
78 |
|
79 |
predicted_ids = torch.argmax(logits, dim=-1)
|
80 |
|
@@ -103,13 +103,17 @@ from sklearn.model_selection import train_test_split
|
|
103 |
import pandas as pd
|
104 |
from datasets import load_dataset
|
105 |
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
!mkdir cer
|
110 |
-
!wget -O cer/cer.py https://huggingface.co/ctl/wav2vec2-large-xlsr-cantonese/raw/main/cer.py
|
111 |
|
|
|
|
|
|
|
112 |
|
|
|
|
|
|
|
113 |
|
114 |
test_dataset = load_dataset('csv', data_files='/content/km_kh_male/line_index_test.csv',split = 'train')
|
115 |
wer = load_metric("wer")
|
@@ -120,30 +124,30 @@ processor = Wav2Vec2Processor.from_pretrained("gagan3012/wav2vec2-xlsr-khmer")
|
|
120 |
model = Wav2Vec2ForCTC.from_pretrained("gagan3012/wav2vec2-xlsr-khmer")
|
121 |
model.to("cuda")
|
122 |
|
123 |
-
chars_to_ignore_regex = '[
|
124 |
resampler = torchaudio.transforms.Resample(48_000, 16_000)
|
125 |
|
126 |
# Preprocessing the datasets.
|
127 |
# We need to read the aduio files as arrays
|
128 |
def speech_file_to_array_fn(batch):
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
|
134 |
test_dataset = test_dataset.map(speech_file_to_array_fn)
|
135 |
|
136 |
# Preprocessing the datasets.
|
137 |
# We need to read the aduio files as arrays
|
138 |
def evaluate(batch):
|
139 |
-
|
140 |
|
141 |
-
|
142 |
-
|
143 |
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
|
148 |
cer = load_metric("cer")
|
149 |
|
30 |
|
31 |
# Wav2Vec2-Large-XLSR-53-khmer
|
32 |
|
33 |
+
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Khmer using the [Common Voice](https://huggingface.co/datasets/common_voice), and [OpenSLR Kh](http://www.openslr.org/42/).
|
34 |
|
35 |
When using this model, make sure that your speech input is sampled at 16kHz.
|
36 |
|
44 |
from datasets import load_dataset
|
45 |
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
46 |
|
47 |
+
!wget https://www.openslr.org/resources/42/km_kh_male.zip
|
48 |
!unzip km_kh_male.zip
|
49 |
!ls km_kh_male
|
50 |
|
51 |
colnames=['path','sentence']
|
52 |
+
df = pd.read_csv('/content/km_kh_male/line_index.tsv',sep='\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\t',header=None,names = colnames)
|
53 |
df['path'] = '/content/km_kh_male/wavs/'+df['path'] +'.wav'
|
54 |
|
55 |
train, test = train_test_split(df, test_size=0.1)
|
66 |
# Preprocessing the datasets.
|
67 |
# We need to read the aduio files as arrays
|
68 |
def speech_file_to_array_fn(batch):
|
69 |
+
\\\\\\\\\\\\\\\\tspeech_array, sampling_rate = torchaudio.load(batch["path"])
|
70 |
+
\\\\\\\\\\\\\\\\tbatch["speech"] = resampler(speech_array).squeeze().numpy()
|
71 |
+
\\\\\\\\\\\\\\\\treturn batch
|
72 |
|
73 |
test_dataset = test_dataset.map(speech_file_to_array_fn)
|
74 |
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
|
75 |
|
76 |
with torch.no_grad():
|
77 |
+
\\\\\\\\\\\\\\\\tlogits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
|
78 |
|
79 |
predicted_ids = torch.argmax(logits, dim=-1)
|
80 |
|
103 |
import pandas as pd
|
104 |
from datasets import load_dataset
|
105 |
|
106 |
+
!wget https://www.openslr.org/resources/42/km_kh_male.zip
|
107 |
+
!unzip km_kh_male.zip
|
108 |
+
!ls km_kh_male
|
|
|
|
|
109 |
|
110 |
+
colnames=['path','sentence']
|
111 |
+
df = pd.read_csv('/content/km_kh_male/line_index.tsv',sep='\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\t',header=None,names = colnames)
|
112 |
+
df['path'] = '/content/km_kh_male/wavs/'+df['path'] +'.wav'
|
113 |
|
114 |
+
train, test = train_test_split(df, test_size=0.1)
|
115 |
+
|
116 |
+
test.to_csv('/content/km_kh_male/line_index_test.csv')
|
117 |
|
118 |
test_dataset = load_dataset('csv', data_files='/content/km_kh_male/line_index_test.csv',split = 'train')
|
119 |
wer = load_metric("wer")
|
124 |
model = Wav2Vec2ForCTC.from_pretrained("gagan3012/wav2vec2-xlsr-khmer")
|
125 |
model.to("cuda")
|
126 |
|
127 |
+
chars_to_ignore_regex = '[\\\\,\\\\?\\\\.\\\\!\\\\-\\\\;\\\\:\\\\"\\\\“]'
|
128 |
resampler = torchaudio.transforms.Resample(48_000, 16_000)
|
129 |
|
130 |
# Preprocessing the datasets.
|
131 |
# We need to read the aduio files as arrays
|
132 |
def speech_file_to_array_fn(batch):
|
133 |
+
\\tbatch["text"] = re.sub(chars_to_ignore_regex, '', batch["text"]).lower()
|
134 |
+
\\tspeech_array, sampling_rate = torchaudio.load(batch["path"])
|
135 |
+
\\tbatch["speech"] = resampler(speech_array).squeeze().numpy()
|
136 |
+
\\treturn batch
|
137 |
|
138 |
test_dataset = test_dataset.map(speech_file_to_array_fn)
|
139 |
|
140 |
# Preprocessing the datasets.
|
141 |
# We need to read the aduio files as arrays
|
142 |
def evaluate(batch):
|
143 |
+
\\tinputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
144 |
|
145 |
+
\\twith torch.no_grad():
|
146 |
+
\\t\\tlogits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
|
147 |
|
148 |
+
\\tpred_ids = torch.argmax(logits, dim=-1)
|
149 |
+
\\tbatch["pred_strings"] = processor.batch_decode(pred_ids)
|
150 |
+
\\treturn batch
|
151 |
|
152 |
cer = load_metric("cer")
|
153 |
|