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: 49.68
---
# Wav2Vec2-Large-XLSR-53-Moroccan-Darija-V1

[othrif/wav2vec2-large-xlsr-moroccan](https://huggingface.co/othrif/wav2vec2-large-xlsr-moroccan) fine-tuned on 6 hours of labeled Darija Audios

I have also added 3 phonetic units to this model ڭ, ڤ and پ represented by k, v and p. For example: ڭال , ڤيديو , پودكاست

## Usage

The model can be used directly (without a language model) 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-v1', tokenizer=tokenizer)
model=Wav2Vec2ForCTC.from_pretrained('boumehdi/wav2vec2-large-xlsr-moroccan-darija-v1')


# 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.replace('k', 'ڭ').replace('v', 'ڤ').replace('p', 'پ'))
```

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

## Evaluation

**Wer**: 49.68

**Training Loss**: 9.88

**Validation Loss**: 45.24


## Future Work

Currently working on **wav2vec2-large-xlsr-moroccan-darija-v2** which will be available very soon by adding more data (from 6hours to 12hours).

I am also working on audio data augmentation techniques (pitch shift, reberbation, additive augmentation.. ) to see if it is going to improve the **WER**.