1
---
2
language: ta
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: Rajaram1996/wav2vec2-large-xlsr-tamil
13
  results:
14
  - task: 
15
      name: Speech Recognition
16
      type: automatic-speech-recognition
17
    dataset:
18
      name: Common Voice ta
19
      type: common_voice
20
      args: ta 
21
    metrics:
22
       - name: Test WER
23
         type: wer
24
         value: 69.76
25
---
26
# Wav2Vec2-Large-XLSR-53-tamil
27
28
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Tamil using the [Common Voice](https://huggingface.co/datasets/common_voice)
29
30
When using this model, make sure that your speech input is sampled at 16kHz.
31
32
## Usage
33
The model can be used directly (without a language model) as follows:
34
35
```python
36
import torch
37
import torchaudio
38
from datasets import load_dataset
39
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
40
41
test_dataset = load_dataset("common_voice", "ta", split="test[:2%]")
42
43
processor = Wav2Vec2Processor.from_pretrained("Rajaram1996/wav2vec2-large-xlsr-53-tamil")
44
model = Wav2Vec2ForCTC.from_pretrained("Rajaram1996/wav2vec2-large-xlsr-53-tamil")
45
46
resampler = torchaudio.transforms.Resample(48_000, 16_000)
47
48
# Preprocessing the datasets.
49
# We need to read the aduio files as arrays
50
51
def speech_file_to_array_fn(batch):
52
    speech_array, sampling_rate = torchaudio.load(batch["path"])
53
    batch["speech"] = resampler(speech_array).squeeze().numpy()
54
    return batch
55
56
test_dataset = test_dataset.map(speech_file_to_array_fn)
57
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
58
59
with torch.no_grad():
60
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
61
62
predicted_ids = torch.argmax(logits, dim=-1)
63
print("Prediction:", processor.batch_decode(predicted_ids))
64
print("Reference:", test_dataset["sentence"][:2])
65
```
66
67
## Evaluation
68
69
The model can be evaluated as follows on the {language} test data of Common Voice.
70
71
```python
72
import torch
73
import torchaudio
74
from datasets import load_dataset, load_metric
75
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
76
import re
77
78
test_dataset = load_dataset("common_voice", "ta", split="test")
79
80
wer = load_metric("wer")
81
82
processor = Wav2Vec2Processor.from_pretrained("Rajaram1996/wav2vec2-large-xlsr-53-tamil")
83
model = Wav2Vec2ForCTC.from_pretrained("Rajaram1996/wav2vec2-large-xlsr-53-tamil")
84
85
model.to("cuda")
86
chars_to_ignore_regex = '[\\\\,\\\\?\\\\.\\\\!\\\\-\\\\;\\\\:\\\\"\\\\“]'
87
88
resampler = torchaudio.transforms.Resample(48_000, 16_000)
89
90
# Preprocessing the datasets.
91
# We need to read the aduio files as arrays
92
def speech_file_to_array_fn(batch):
93
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
94
    speech_array, sampling_rate = torchaudio.load(batch["path"])
95
    batch["speech"] = resampler(speech_array).squeeze().numpy()
96
    return batch
97
98
test_dataset = test_dataset.map(speech_file_to_array_fn)
99
100
# Preprocessing the datasets.
101
# We need to read the aduio files as arrays
102
def evaluate(batch):
103
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
104
    with torch.no_grad():
105
        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
106
    pred_ids = torch.argmax(logits, dim=-1)
107
    batch["pred_strings"] = processor.batch_decode(pred_ids)
108
    return batch
109
110
result = test_dataset.map(evaluate, batched=True, batch_size=8)
111
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
112
```
113
114
**Test Result**: 69.76 %