1
---
2
language: th
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 Large Thai by Sakares
13
  results:
14
  - task: 
15
      name: Speech Recognition
16
      type: automatic-speech-recognition
17
    dataset:
18
      name: Common Voice th
19
      type: common_voice
20
      args: th  
21
    metrics:
22
       - name: Test WER
23
         type: wer
24
         value: 44.46
25
---
26
27
# Wav2Vec2-Large-XLSR-53-Thai
28
29
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Thai using the [Common Voice](https://huggingface.co/datasets/common_voice)
30
When using this model, make sure that your speech input is sampled at 16kHz.
31
32
## Usage
33
34
The model can be used directly (without a language model) as follows:
35
36
```python
37
import torch
38
import torchaudio
39
from datasets import load_dataset
40
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
41
from pythainlp.tokenize import word_tokenize
42
43
test_dataset = load_dataset("common_voice", "th", split="test[:2%]")
44
45
processor = Wav2Vec2Processor.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")
46
model = Wav2Vec2ForCTC.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")
47
48
resampler = torchaudio.transforms.Resample(48_000, 16_000)
49
50
## For Thai NLP Library, please feel free to check https://pythainlp.github.io/docs/2.2/api/tokenize.html
51
def th_tokenize(batch):
52
    batch["sentence"] = " ".join(word_tokenize(batch["sentence"], engine="newmm"))
53
    return batch
54
55
# Preprocessing the datasets.
56
# We need to read the aduio files as arrays
57
def speech_file_to_array_fn(batch):
58
	speech_array, sampling_rate = torchaudio.load(batch["path"])
59
	batch["speech"] = resampler(speech_array).squeeze().numpy()
60
	return batch
61
62
test_dataset = test_dataset.map(speech_file_to_array_fn).map(th_tokenize)
63
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
64
65
with torch.no_grad():
66
	logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
67
68
predicted_ids = torch.argmax(logits, dim=-1)
69
70
print("Prediction:", processor.batch_decode(predicted_ids))
71
print("Reference:", test_dataset["sentence"][:2])
72
```
73
Usage script [here](https://colab.research.google.com/drive/1w0VywsBtjrO2pHHPmiPugYI9yeF8nUKg?usp=sharing)
74
75
## Evaluation
76
77
The model can be evaluated as follows on the {language} test data of Common Voice.
78
79
80
```python
81
import torch
82
import torchaudio
83
from datasets import load_dataset, load_metric
84
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
85
from pythainlp.tokenize import word_tokenize
86
import re
87
88
test_dataset = load_dataset("common_voice", "th", split="test")
89
wer = load_metric("wer")
90
91
processor = Wav2Vec2Processor.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")
92
model = Wav2Vec2ForCTC.from_pretrained("sakares/wav2vec2-large-xlsr-thai-demo")
93
model.to("cuda")
94
95
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“]'
96
resampler = torchaudio.transforms.Resample(48_000, 16_000)
97
98
## For Thai NLP Library, please feel free to check https://pythainlp.github.io/docs/2.2/api/tokenize.html
99
def th_tokenize(batch):
100
    batch["sentence"] = " ".join(word_tokenize(batch["sentence"], engine="newmm"))
101
    return batch
102
103
# Preprocessing the datasets.
104
# We need to read the aduio files as arrays
105
def speech_file_to_array_fn(batch):
106
	batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
107
	speech_array, sampling_rate = torchaudio.load(batch["path"])
108
	batch["speech"] = resampler(speech_array).squeeze().numpy()
109
	return batch
110
111
test_dataset = test_dataset.map(speech_file_to_array_fn).map(th_tokenize)
112
113
# Preprocessing the datasets.
114
# We need to read the aduio files as arrays
115
def evaluate(batch):
116
	inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
117
118
	with torch.no_grad():
119
		logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
120
121
	pred_ids = torch.argmax(logits, dim=-1)
122
	batch["pred_strings"] = processor.batch_decode(pred_ids)
123
	return batch
124
125
result = test_dataset.map(evaluate, batched=True, batch_size=8)
126
127
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
128
```
129
130
**Test Result**: 44.46 %
131
Evaluate script [here](https://colab.research.google.com/drive/1WZGtHKWXBztRsuXHIdebf6uoAsp7rTnK?usp=sharing)
132
133
## Training
134
135
The Common Voice `train`, `validation` datasets were used for training.
136
137
The script used for training can be found [here](https://colab.research.google.com/drive/18oUbeZgBGSkz16zC_WOa154QZOdmvjyt?usp=sharing)
138