---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: dev
    path: data/dev-*
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: transcription
    dtype: string
  - name: dialect
    dtype: string
  splits:
  - name: train
    num_bytes: 30053008388.047405
    num_examples: 187816
  - name: test
    num_bytes: 3026464537.665628
    num_examples: 18801
  - name: dev
    num_bytes: 758023819.869451
    num_examples: 4709
  download_size: 32588024982
  dataset_size: 33837496745.58248

task_categories:
- automatic-speech-recognition
- text-to-speech
- audio-classification
language:
- ff
pretty_name: Fula Multidialectal Speech Corpora Augmented
size_categories:
- 100K<n<1M
tags:
- speech
- low-ressource
- audio
---

# Sources of the data

This dataset is created from [FulaSpeechCorpora dataset](https://huggingface.co./datasets/cawoylel/FulaSpeechCorpora) and splitted into **train**, **test** and **dev** sets. It has been augmented using `audiomentations` library [https://iver56.github.io/audiomentations/](https://iver56.github.io/audiomentations/).

The augmentation process includes the following tranformations:
- Add background noise  to simulate an environment where background noise is present
- Add gaussian noise to the samples
- Multiply the audio by a random amplitude factor to reduce or increase the volume. This technique can help a model become somewhat invariant to the overall gain of the input audio.
- Change the speed or duration of the signal without changing the pitch.

# How to augment

**Download noise and background sounds**

```python
wget https://www.openslr.org/resources/17/musan.tar.gz
tar -zxvf musan.tar.gz
rm -r ./musan/speech
```

**Augment**

```python
from audiomentations import *

def get_transformations():
  """
  Define audio augmentation transformations
  """
  musan_dir = "./musan"
  augment = Compose([
      Gain(min_gain_in_db=-6, max_gain_in_db=6, p=0.1),
      PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
      TimeStretch(min_rate=0.9, max_rate=1.1, p=0.4, leave_length_unchanged=False),
      OneOf([
      AddBackgroundNoise(
                      sounds_path=musan_dir, min_snr_in_db=1.0, max_snr_in_db=5.0, noise_transform=PolarityInversion(), p=1.
                  ),
      AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=1.),],  p=1.0)

      ])

  return augment

def augment_dataset(batch):
  sample = batch['audio']

  augmentation = get_transformations()
    # apply augmentation
  augmented_waveform = augmentation(sample["array"], sample_rate=sample["sampling_rate"])
  batch['audio']["array"] = augmented_waveform
  return batch
```

# Fula varieties in this dataset

This dataset contains 8 varrieties:
- __Pulaar__: spoken in Senegal, Mauritania and West-Mali.
- __Pular__: spoken in Guinea.
- __Maacina__: spoken in the Center and East of Mali.
- __Liptako__: spoken in Burkina Faso and Niger.
- __Caka__: spoken in the Central Nigeria.
- __Bororro__: a very nomad group living in Cameroon, Central African Republic and Tchad.
- __Borgu__: spoken in Togo and Benin.
- __Adamawa__: spoken in Cameroon and South-East Nigeria.


# How to use

The `datasets` library allows you to load and pre-process your dataset in pure Python, at scale. The dataset can be downloaded and prepared in one call to your local drive by using the `load_dataset` function. 

For example, to download the training set, simply specify the corresponding split name:

```python
from datasets import load_dataset
train = load_dataset("cawoylel/FulaSpeechCorpora-splited-noise_augmented", split = "train")
```

You can also load all the dataset:

```python
from datasets import load_dataset
data = load_dataset("cawoylel/FulaSpeechCorpora-splited-noise_augmented")
```


Using the datasets library, you can also stream the dataset on-the-fly by adding a `streaming=True` argument to the `load_dataset` function call. Loading a dataset in streaming mode loads individual samples of the dataset at a time, rather than downloading the entire dataset to disk.
```python
from datasets import load_dataset
data = load_dataset("cawoylel/FulaSpeechCorpora-splited-noise_augmented", streaming=True)
print(next(iter(data)))
```


# Data Fields

The data fields are the same among all splits.
- **dialect** (str): The name of the dialect
- **audio** (dict): Audio object including loaded audio array, sampling rate and path to audio
- **transcription** (str): Transcription of the audio file

# Social Impact of Dataset

Under-representation, dialectal and cultural diversity.

# Discussion of Biases

Gender biases, noise robustness


# Additional Information

All datasets are licensed under the [Creative Commons license (CC-BY)](https://creativecommons.org/licenses/).

# Citation Information

Please cite us when using the FulaSpeechCorpora:

```
@article{fleurs2022arxiv,
  title = {FulaSpeechCorpora-splited-noise_augmented: A multidialectal speech dataset for Fula.},
  author = {Sy, Yaya and Doucouré, Dioula},
  url = {https://huggingface.co./datasets/cawoylel/FulaSpeechCorpora-splited-noise_augmented},
  year = {2023},
```