patrickvonplaten
commited on
Commit
•
036108d
1
Parent(s):
b038889
correct readme
Browse files
README.md
CHANGED
@@ -17,3 +17,53 @@ Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/210
|
|
17 |
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
|
18 |
|
19 |
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
|
18 |
|
19 |
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
|
20 |
+
|
21 |
+
|
22 |
+
# Usage for inference
|
23 |
+
|
24 |
+
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
|
25 |
+
|
26 |
+
```python
|
27 |
+
#!/usr/bin/env python3
|
28 |
+
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
|
29 |
+
from datasets import load_dataset
|
30 |
+
import torchaudio
|
31 |
+
import torch
|
32 |
+
|
33 |
+
# resample audio
|
34 |
+
|
35 |
+
# load model & processor
|
36 |
+
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-nl")
|
37 |
+
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-nl")
|
38 |
+
|
39 |
+
# load dataset
|
40 |
+
ds = load_dataset("common_voice", "nl", split="validation[:1%]")
|
41 |
+
|
42 |
+
# common voice does not match target sampling rate
|
43 |
+
common_voice_sample_rate = 48000
|
44 |
+
target_sample_rate = 16000
|
45 |
+
|
46 |
+
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
|
47 |
+
|
48 |
+
|
49 |
+
# define mapping fn to read in sound file and resample
|
50 |
+
def map_to_array(batch):
|
51 |
+
speech, _ = torchaudio.load(batch["path"])
|
52 |
+
speech = resampler(speech)
|
53 |
+
batch["speech"] = speech[0]
|
54 |
+
return batch
|
55 |
+
|
56 |
+
|
57 |
+
# load all audio files
|
58 |
+
ds = ds.map(map_to_array)
|
59 |
+
|
60 |
+
# run inference on the first 5 data samples
|
61 |
+
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
|
62 |
+
|
63 |
+
# inference
|
64 |
+
logits = model(**inputs).logits
|
65 |
+
predicted_ids = torch.argmax(logits, axis=-1)
|
66 |
+
|
67 |
+
print(processor.batch_decode(predicted_ids))
|
68 |
+
```
|
69 |
+
|