First Version
Browse files
README.md
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: mn
|
3 |
+
datasets:
|
4 |
+
- common_voice
|
5 |
+
tags:
|
6 |
+
- audio
|
7 |
+
- automatic-speech-recognition
|
8 |
+
- speech
|
9 |
+
- xlsr-fine-tuning-week
|
10 |
+
license: apache-2.0
|
11 |
+
model-index:
|
12 |
+
- name: XLSR Wav2Vec2 Mongolian by Salim Shaikh
|
13 |
+
results:
|
14 |
+
- task:
|
15 |
+
name: Speech Recognition
|
16 |
+
type: automatic-speech-recognition
|
17 |
+
dataset:
|
18 |
+
name: Common Voice mn
|
19 |
+
type: common_voice
|
20 |
+
args: {mn}
|
21 |
+
metrics:
|
22 |
+
- name: Test WER
|
23 |
+
type: wer
|
24 |
+
value: 41.92
|
25 |
+
---
|
26 |
+
# Wav2Vec2-Large-XLSR-53-Mongolian
|
27 |
+
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Mongolian using the [Common Voice](https://huggingface.co/datasets/common_voice)
|
28 |
+
When using this model, make sure that your speech input is sampled at 16kHz.
|
29 |
+
## Usage
|
30 |
+
The model can be used directly (without a language model) as follows:
|
31 |
+
|
32 |
+
```python
|
33 |
+
|
34 |
+
import torchaudio
|
35 |
+
from datasets import load_dataset, load_metric
|
36 |
+
from transformers import (
|
37 |
+
Wav2Vec2ForCTC,
|
38 |
+
Wav2Vec2Processor,
|
39 |
+
)
|
40 |
+
import torch
|
41 |
+
import re
|
42 |
+
import sys
|
43 |
+
|
44 |
+
model_name = "sammy786/wav2vec2-large-xlsr-mongolian"
|
45 |
+
device = "cuda"
|
46 |
+
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�]'
|
47 |
+
|
48 |
+
model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
|
49 |
+
processor = Wav2Vec2Processor.from_pretrained(model_name)
|
50 |
+
|
51 |
+
ds = load_dataset("common_voice", "mn", split="test", data_dir="./cv-corpus-6.1-2020-12-11")
|
52 |
+
|
53 |
+
resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
|
54 |
+
|
55 |
+
def map_to_array(batch):
|
56 |
+
speech, _ = torchaudio.load(batch["path"])
|
57 |
+
batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
|
58 |
+
batch["sampling_rate"] = resampler.new_freq
|
59 |
+
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + " "
|
60 |
+
return batch
|
61 |
+
|
62 |
+
ds = ds.map(map_to_array)
|
63 |
+
|
64 |
+
def map_to_pred(batch):
|
65 |
+
features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
|
66 |
+
input_values = features.input_values.to(device)
|
67 |
+
attention_mask = features.attention_mask.to(device)
|
68 |
+
with torch.no_grad():
|
69 |
+
logits = model(input_values, attention_mask=attention_mask).logits
|
70 |
+
pred_ids = torch.argmax(logits, dim=-1)
|
71 |
+
batch["predicted"] = processor.batch_decode(pred_ids)
|
72 |
+
batch["target"] = batch["sentence"]
|
73 |
+
return batch
|
74 |
+
|
75 |
+
result = ds.map(map_to_pred, batched=True, batch_size=20, remove_columns=list(ds.features.keys()))
|
76 |
+
|
77 |
+
wer = load_metric("wer")
|
78 |
+
print(wer.compute(predictions=result["predicted"], references=result["target"]))
|
79 |
+
|
80 |
+
```
|
81 |
+
**Test Result**: 41.92 %
|