---
language: ary
datasets:
- mgb5 
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Moroccan Arabic dialect by Othmane Rifki
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: MGB5 from ELDA and https://arabicspeech.org/
      type: ELDA and https://arabicspeech.org/
      args: ary
    metrics:
       - name: Test WER
         type: wer
         value: 44.51
---

# Wav2Vec2-Large-XLSR-53-Moroccan

Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on [MGB5 Moroccan Arabic](http://www.islrn.org/resources/938-639-614-524-5/) kindly provided by [ELDA](http://www.elra.info/en/about/elda/) and [ArabicSpeech](https://arabicspeech.org/mgb5/).

In order to have access to MGB5, please request it from ELDA.

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:

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

test_dataset = load_dataset("ma_speech_corpus", split="test")

processor = Wav2Vec2Processor.from_pretrained("othrif/wav2vec2-large-xlsr-arabic")
model = Wav2Vec2ForCTC.from_pretrained("othrif/wav2vec2-large-xlsr-arabic")

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("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])
```


## Evaluation

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


```python
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re

test_dataset = load_dataset("ma_speech_corpus", split="test") 
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("othrif/wav2vec2-large-xlsr-moroccan") 
model = Wav2Vec2ForCTC.from_pretrained("othrif/wav2vec2-large-xlsr-moroccan")
model.to("cuda")

chars_to_ignore_regex = '[0\,\?\.\!\-\;\:\"\“\%\‘\”\�\n\@\ـ\؟\*\ \#\'\ \…\\u2003]'
#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):
    batch["text"] = re.sub(chars_to_ignore_regex, '', batch["text"]).lower()
    batch["text"] = re.sub('[a-zA-z]', '', batch["text"]).lower() + " "
        batch["text"] = re.sub('[ًٌٍَُِ~]', '', batch["text"]).lower() + " "

        # batch["text"] = re.sub('\\n','', batch["text"])
        batch["text"] = re.sub("[إأٱآا]", "ا", batch["text"])
        batch["text"] = re.sub("ڸ", "ل", batch["text"])
        noise = re.compile(""" ّ    | # Tashdid
                             َ    | # Fatha
                             ً    | # Tanwin Fath
                             ُ    | # Damma
                             ٌ    | # Tanwin Damm
                             ِ    | # Kasra
                             ٍ    | # Tanwin Kasr
                             ْ    | # Sukun
                             ـ     # Tatwil/Kashida
                         """, re.VERBOSE)
        batch["text"] = re.sub(noise, '', batch["text"])
        batch["text"] = re.sub('ٖ', '', batch["text"]).lower() + " "
    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)

# 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["text"])))
```

**Test Result**: 44.51


## Training

The  [MGB5](http://www.islrn.org/resources/938-639-614-524-5/) `train`, `validation` datasets were used for training.

The script used for training can be found [here](https://github.com/othrif/xlsr-wav2vec2)