1
---
2
language: ta
3
datasets:
4
- common_voice
5
metrics:
6
- wer
7
tags:
8
- audio
9
- automatic-speech-recognition
10
- speech
11
- xlsr-fine-tuning-week
12
license: apache-2.0
13
model-index:
14
- name: thanish wav2vec2-large-xlsr-tamil
15
  results:
16
  - task: 
17
      name: Speech Recognition
18
      type: automatic-speech-recognition
19
    dataset:
20
      name: Common Voice ta
21
      type: common_voice
22
      args: ta
23
    metrics:
24
       - name: Test WER
25
         type: wer
26
         value: 100.00
27
---
28
29
# Wav2Vec2-Large-XLSR-53-Tamil
30
31
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Tamil using the [Common Voice](https://huggingface.co/datasets/common_voice) dataset.
32
When using this model, make sure that your speech input is sampled at 16kHz.
33
34
## Usage
35
36
The model can be used directly (without a language model) as follows:
37
38
```python
39
import torch
40
import torchaudio
41
from datasets import load_dataset
42
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
43
test_dataset = load_dataset("common_voice", "{lang_id}", split="test[:2%]") #TODO: replace {lang_id} in your language code here. Make sure the code is one of the *ISO codes* of [this](https://huggingface.co/languages) site.
44
processor = Wav2Vec2Processor.from_pretrained("{model_id}") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
45
model = Wav2Vec2ForCTC.from_pretrained("{model_id}") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
46
resampler = torchaudio.transforms.Resample(48_000, 16_000)
47
# Preprocessing the datasets.
48
# We need to read the aduio files as arrays
49
def speech_file_to_array_fn(batch):
50
\\tspeech_array, sampling_rate = torchaudio.load(batch["path"])
51
\\tbatch["speech"] = resampler(speech_array).squeeze().numpy()
52
\\treturn batch
53
test_dataset = test_dataset.map(speech_file_to_array_fn)
54
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
55
with torch.no_grad():
56
\\tlogits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
57
predicted_ids = torch.argmax(logits, dim=-1)
58
print("Prediction:", processor.batch_decode(predicted_ids))
59
print("Reference:", test_dataset["sentence"][:2])
60
```
61
62
63
## Evaluation
64
65
The model can be evaluated as follows on the Tamil test data of Common Voice.
66
67
68
```python
69
import torch
70
import torchaudio
71
from datasets import load_dataset, load_metric
72
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
73
import re
74
test_dataset = load_dataset("common_voice", "{lang_id}", split="test") #TODO: replace {lang_id} in your language code here. Make sure the code is one of the *ISO codes* of [this](https://huggingface.co/languages) site.
75
wer = load_metric("wer")
76
processor = Wav2Vec2Processor.from_pretrained("{model_id}") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
77
model = Wav2Vec2ForCTC.from_pretrained("{model_id}") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
78
model.to("cuda")
79
chars_to_ignore_regex = '[\\\\,\\\\?\\\\.\\\\!\\\\-\\\\;\\\\:\\\\"\\\\“]'  # TODO: adapt this list to include all special characters you removed from the data
80
resampler = torchaudio.transforms.Resample(48_000, 16_000)
81
# Preprocessing the datasets.
82
# We need to read the audio files as arrays
83
def speech_file_to_array_fn(batch):
84
\\tbatch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
85
\\tspeech_array, sampling_rate = torchaudio.load(batch["path"])
86
\\tbatch["speech"] = resampler(speech_array).squeeze().numpy()
87
\\treturn batch
88
test_dataset = test_dataset.map(speech_file_to_array_fn)
89
# Preprocessing the datasets.
90
# We need to read the aduio files as arrays
91
def evaluate(batch):
92
\\tinputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
93
\\twith torch.no_grad():
94
\\t\\tlogits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
95
\\tpred_ids = torch.argmax(logits, dim=-1)
96
\\tbatch["pred_strings"] = processor.batch_decode(pred_ids)
97
\\treturn batch
98
result = test_dataset.map(evaluate, batched=True, batch_size=8)
99
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
100
```
101
102
**Test Result**: 100.00 %  
103
104
105
## Training
106
107
The Common Voice `train`, `validation` were used for training 
108
109
The script used for training can be found [https://colab.research.google.com/drive/1PC2SjxpcWMQ2qmRw21NbP38wtQQUa5os#scrollTo=YKBZdqqJG9Tv](...)