README.md

language: fon
datasets:
- [Fon Dataset](https://github.com/laleye/pyFongbe/tree/master/data)
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: Fon XLSR Wav2Vec2 Large 53
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: fon
      type: fon_dataset
      args: fon
    metrics:
       - name: Test WER
         type: wer
         value: 14.97
---

# Wav2Vec2-Large-XLSR-53-Fon

Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on [Fon](https://en.wikipedia.org/wiki/Fon_language) using the [Fon Dataset](https://github.com/laleye/pyFongbe/tree/master/data).

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

#This will download the files from Layele's Github to the directory FonAudio
if not os.path.isdir("./FonAudio"):
  !wget https://github.com/laleye/pyFongbe/archive/master/data.zip
  with zipfile.ZipFile("data.zip","r") as zip_ref:
    zip_ref.extractall("./FonAudio")
    
with open('./FonAudio/pyFongbe-master/data/train.csv', newline='',encoding='UTF-8') as f:
      reader = csv.reader(f)
      data = list(reader)
      train_data = [data[i] for i in range(len(data)) if i!=0]

with open('./FonAudio/pyFongbe-master/data/test.csv', newline='',encoding='UTF-8') as f:
      reader = csv.reader(f)
      data = list(reader)
      t_data = [data[i] for i in range(len(data)) if i!=0]
      
      
#Get valid indices
random.seed(42) #this seed was used specifically to compare with Okwugbe model


v = 1500 #200 samples for valid. Change as you want
test_list = [i for i in range(len(t_data))]
valid_indices = random.choices(test_list, k=v)

test_data = [t_data[i] for i in range(len(t_data)) if i not in valid_indices] 
valid_data = [t_data[i] for i in range(len(t_data)) if i in valid_indices]

#Length of validation_dataset -> 1107
#Length of test_dataset -> 1061

#Please note, the final validation size is is smaller than the expected (1500) because we used random.choices which could contain duplicates.

#Create JSON files 
def create_json_file(d):
  utterance = d[2]
  wav_path =d[0]
  wav_path = wav_path.replace("/home/frejus/Projects/Fongbe_ASR/pyFongbe","./FonAudio/pyFongbe-master")
  return {
      "path": wav_path,
      "sentence": utterance
  }

train_json = [create_json_file(i) for i in train_data]
test_json = [create_json_file(i) for i in test_data]
valid_json = [create_json_file(i) for i in valid_data]

#Save JSON files to your Google Drive folders
#Make folder in GDrive to store files
train_path = '/content/drive/MyDrive/fon_xlsr/train'
test_path = '/content/drive/MyDrive/fon_xlsr/test'
valid_path = '/content/drive/MyDrive/fon_xlsr/valid'

if not os.path.isdir(train_path):
  print("Creating paths")
  os.makedirs(train_path)
  os.makedirs(test_path) #this is where we save the test files
  os.makedirs(valid_path)
  

#for train
for i, sample in enumerate(train_json):
  file_path = os.path.join(train_path,'train_fon_{}.json'.format(i))
  with open(file_path, 'w') as outfile:
    json.dump(sample, outfile)

#for test
for i, sample in enumerate(test_json):
  file_path = os.path.join(test_path,'test_fon_{}.json'.format(i))
  with open(file_path, 'w') as outfile:
    json.dump(sample, outfile)

#for valid
for i, sample in enumerate(valid_json):
  file_path = os.path.join(valid_path,'valid_fon_{}.json'.format(i))
  with open(file_path, 'w') as outfile:
    json.dump(sample, outfile)
  

#Load test_dataset from saved files in folder
from datasets import load_dataset, load_metric

#for test
for root, dirs, files in os.walk(test_path):
  test_dataset= load_dataset("json", data_files=[os.path.join(root,i) for i in files],split="train")

#Remove unnecessary chars
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�]'
def remove_special_characters(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + " "
    return batch

test_dataset = test_dataset.map(remove_special_characters)

processor = Wav2Vec2Processor.from_pretrained("chrisjay/wav2vec2-large-xlsr-53-fon") 
model = Wav2Vec2ForCTC.from_pretrained("chrisjay/wav2vec2-large-xlsr-53-fon") 

#No need for resampling because audio dataset already at 16kHz
#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"]=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 our unique Fon test data. 

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

for root, dirs, files in os.walk(test_path):
  test_dataset = load_dataset("json", data_files=[os.path.join(root,i) for i in files],split="train")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�]'
def remove_special_characters(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower() + " "
    return batch

test_dataset = test_dataset.map(remove_special_characters)
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("chrisjay/wav2vec2-large-xlsr-53-fon")
model = Wav2Vec2ForCTC.from_pretrained("chrisjay/wav2vec2-large-xlsr-53-fon") #use checkpoint-12400 to get our WER test results
model.to("cuda")

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = speech_array[0].numpy()
    batch["sampling_rate"] = sampling_rate
    batch["target_text"] = batch["sentence"]
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

#Evaluation on test dataset
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**: 14.97 % 

## Training

The [Fon dataset](https://github.com/laleye/pyFongbe/tree/master/data) was split into `train`(8235 samples), `validation`(1107 samples), and `test`(1061 samples).

The script used for training can be found [here](https://colab.research.google.com/drive/11l6qhJCYnPTG1TQZ8f3EvKB9z12TQi4g?usp=sharing)