README.md

---
language: ary
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 Boumehdi
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    metrics:
       - name: Test WER
         type: wer
         value: 23.44
---
# Wav2Vec2-Large-XLSR-53-Moroccan-Darija

**wav2vec2-large-xlsr-53** fine-tuned on 10 hours (~70 speakers) of labeled Darija Audios

The vocabulary contains 3 additional phonetic units ڭ, ڤ and پ. For example: ڭال , ڤيديو , پودكاست

## Usage

The model can be used directly as follows:

```python
import librosa
import torch
from transformers import Wav2Vec2CTCTokenizer, Wav2Vec2ForCTC, Wav2Vec2Processor, TrainingArguments, Wav2Vec2FeatureExtractor, Trainer

tokenizer = Wav2Vec2CTCTokenizer("./vocab.json", unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")
processor = Wav2Vec2Processor.from_pretrained('boumehdi/wav2vec2-large-xlsr-moroccan-darija', tokenizer=tokenizer)
model=Wav2Vec2ForCTC.from_pretrained('boumehdi/wav2vec2-large-xlsr-moroccan-darija')


# load the audio data (use your own wav file here!)
input_audio, sr = librosa.load('file.wav', sr=16000)

# tokenize
input_values = processor(input_audio, return_tensors="pt", padding=True).input_values

# retrieve logits
logits = model(input_values).logits

tokens=torch.argmax(logits, axis=-1)

# decode using n-gram
transcription = tokenizer.batch_decode(tokens)

# print the output
print(transcription)
```

Here's the output: ڭالت ليا هاد السيد هادا ما كاينش بحالو 

## Evaluation & Previous works

====================================

-v3 (fine-tuned on 10 hours of audio + changed hyperparameters + discovered a huge mistake when using the letter 'ا' that has improved the WER dramatically)

**Wer**: 23.44

**Training Loss**: 15.96

**Validation Loss**: 33.92

The validation loss goes down as we add more data for training.

Further training to decrease the training Loss makes this model overfit a little bit.

====================================

-v2 (fine-tuned on 9 hours of audio + replaced أ and ى and إ with ا as it creates a lot of problems + tried to standardize the Moroccan Darija)

**Wer**: 44.30

**Training Loss**: 12.99

**Validation Loss**: 36.93

Validation Loss has decreased on this version which means that the model can more generalize for unknown data compared to the previous version.

The validation loss is still high also because the validation data contains words that have never been trained before. The solution is to add more data and more hours of training.

Further training to decrease the training Loss makes this model overfit a little bit.

====================================

-v1 (fine-tuned on 6 hours of audio)

**Wer**: 49.68

**Training Loss**: 9.88

**Validation Loss**: 45.24

====================================

## Future Work

I am currently working on improving this model.

email: [email protected]