Create README.md
Browse files
README.md
CHANGED
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## Evaluation on Zeroth-Korean ASR corpus
|
2 |
+
|
3 |
+
(Google colab notebook(Korean))[https://colab.research.google.com/github/indra622/tutorials/blob/master/wav2vec2_korean_tutorial.ipynb]
|
4 |
+
|
5 |
+
```
|
6 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
7 |
+
from datasets import load_dataset
|
8 |
+
import soundfile as sf
|
9 |
+
import torch
|
10 |
+
from jiwer import wer
|
11 |
+
|
12 |
+
processor = Wav2Vec2Processor.from_pretrained("kresnik/wav2vec2-large-xlsr-korean")
|
13 |
+
|
14 |
+
model = Wav2Vec2ForCTC.from_pretrained("kresnik/wav2vec2-large-xlsr-korean").to('cuda')
|
15 |
+
|
16 |
+
ds = load_dataset("kresnik/zeroth_korean", "clean")
|
17 |
+
|
18 |
+
test_ds = ds['test']
|
19 |
+
|
20 |
+
def map_to_array(batch):
|
21 |
+
speech, _ = sf.read(batch["file"])
|
22 |
+
batch["speech"] = speech
|
23 |
+
return batch
|
24 |
+
|
25 |
+
test_ds = test_ds.map(map_to_array)
|
26 |
+
|
27 |
+
def map_to_pred(batch):
|
28 |
+
inputs = processor(batch["speech"], sampling_rate=16000, return_tensors="pt", padding="longest")
|
29 |
+
input_values = inputs.input_values.to("cuda")
|
30 |
+
#attention_mask = inputs.attention_mask.to("cuda")
|
31 |
+
|
32 |
+
with torch.no_grad():
|
33 |
+
#logits = model(input_values, attention_mask=attention_mask).logits
|
34 |
+
logits = model(input_values).logits
|
35 |
+
|
36 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
37 |
+
transcription = processor.batch_decode(predicted_ids)
|
38 |
+
batch["transcription"] = transcription
|
39 |
+
return batch
|
40 |
+
|
41 |
+
result = test_ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=["speech"])
|
42 |
+
|
43 |
+
print("WER:", wer(result["text"], result["transcription"]))
|
44 |
+
|
45 |
+
```
|
46 |
+
|
47 |
+
### Expected WER: 7.43%
|