Wav2Vec2-Large-XLSR-53-Arabic

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Arabic using the train splits of Common Voice and Arabic Speech Corpus. When using this model, make sure that your speech input is sampled at 16kHz.

Usage

The model can be used directly (without a language model) as follows:

%%capture
!pip install datasets
!pip install transformers==4.4.0
!pip install torchaudio
!pip install jiwer
!pip install tnkeeh

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

test_dataset = load_dataset("common_voice", "ar", split="test[:2%]")


processor = Wav2Vec2Processor.from_pretrained("mohammed/ar")
model = Wav2Vec2ForCTC.from_pretrained("mohammed/ar")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("The predicted sentence is: ", processor.batch_decode(predicted_ids))
print("The original sentence is:", test_dataset["sentence"][:2])

The output is:

The predicted sentence is : ['ألديك قلم', 'ليست نارك مكسافة على هذه الأرض أبعد من يوم أمس']
The original sentence is: ['ألديك قلم ؟', 'ليست هناك مسافة على هذه الأرض أبعد من يوم أمس.']

Evaluation

The model can be evaluated as follows on the Arabic test data of Common Voice:

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
# creating a dictionary with all diacritics
dict = {
'ِ': '',
'ُ': '', 
'ٓ': '', 
'ٰ': '', 
'ْ': '', 
'ٌ': '', 
'ٍ': '', 
'ً': '', 
'ّ': '', 
'َ': '',
'~': '',
',': '',
'ـ': '',
'—': '',
'.': '',
'!': '',
'-': '',
';': '',
':': '',
'\'': '',
'"': '',
'☭': '',
'«': '',
'»': '',
'؛': '',
'ـ': '',
'_': '',
'،': '',
'“': '',
'%': '',
'‘': '',
'”': '',
'�': '',
'_': '',
',': '',
'?': '',
'#': '',
'‘': '',
'.': '',
'؛': '',
'get': '',
'؟': '',
'  ': ' ',
'\'ۖ ': '',
'\'': '',
 '\'ۚ' : '',
 ' \'': '',
 '31': '', 
 '24': '',
 '39': ''
} 

# replacing multiple diacritics using dictionary (stackoverflow is amazing)
def remove_special_characters(batch):
  # Create a regular expression  from the dictionary keys
  regex = re.compile("(%s)" % "|".join(map(re.escape, dict.keys())))
  # For each match, look-up corresponding value in dictionary
  batch["sentence"] = regex.sub(lambda mo: dict[mo.string[mo.start():mo.end()]], batch["sentence"])
  return batch 
  

test_dataset = load_dataset("common_voice", "ar", split="test") 
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("mohammed/ar") 
model = Wav2Vec2ForCTC.from_pretrained("mohammed/ar")
model.to("cuda")


resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
test_dataset = test_dataset.map(remove_special_characters)
# Preprocessing the datasets.
# We need to read the audio files as arrays
def evaluate(batch):
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    with torch.no_grad():
        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits

    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch

result = test_dataset.map(evaluate, batched=True, batch_size=8)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

Test Result: 36.69%

Future Work

One can use data augmentation, transliteration, or attention_mask to increase the accuracy.