benjaminogbonna/nigerian_common_voice_dataset

Dataset Card for Nigerian Common Voice Dataset

Dataset Summary

The Nigerian Common Voice Dataset is a comprehensive dataset consisting of 158 hours of audio recordings and corresponding transcription (sentence). This dataset includes metadata like accent, locale that can help improve the accuracy of speech recognition engines. This dataset is specifically curated to address the gap in speech and language datasets for African accents, making it a valuable resource for researchers and developers working on Automatic Speech Recognition (ASR), Speech-to-text (STT), Text-to-Speech (TTS), Accent recognition, and Natural language processing (NLP) systems.

The dataset currently consists of 158 hours of audio recordings in 4 languages, but more voices and languages are always added. Contributions are welcome.

Languages

English, Hausa, Igbo, Yoruba

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 Igbo config, simply specify the corresponding language config name (i.e., "igbo" for Igbo):

from datasets import load_dataset
dataset = load_dataset("benjaminogbonna/nigerian_common_voice_dataset", "igbo", split="train")

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.

from datasets import load_dataset
dataset = load_dataset("benjaminogbonna/nigerian_common_voice_dataset", "igbo", split="train", streaming=True)
print(next(iter(cv_17)))

Bonus: create a PyTorch dataloader directly with your own datasets (local/streamed).

Local

from datasets import load_dataset
from torch.utils.data.sampler import BatchSampler, RandomSampler
dataset = load_dataset("benjaminogbonna/nigerian_common_voice_dataset", "igbo", split="train")
batch_sampler = BatchSampler(RandomSampler(dataset), batch_size=32, drop_last=False)
dataloader = DataLoader(dataset, batch_sampler=batch_sampler)

Streaming

from datasets import load_dataset
from torch.utils.data import DataLoader
dataset = load_dataset("benjaminogbonna/nigerian_common_voice_dataset", "igbo", split="train")
dataloader = DataLoader(dataset, batch_size=32)

To find out more about loading and preparing audio datasets, head over to hf.co/blog/audio-datasets.

Example scripts

Train your own CTC or Seq2Seq Automatic Speech Recognition models on Common Voice 16 with transformers - here.

Dataset Structure

Data Instances

A typical data point comprises the path to the audio file and its sentence. Additional fields include accent, client_id and locale.

{
  'client_id': 'user_5256', 
  'path': 'clips/ng_voice_igbo_5257.mp3',
  'audio': {
    'path': 'clips/ng_voice_igbo_5257.mp3', 
    'array': array([-0.00048828, -0.00018311, -0.00137329, ...,  0.00079346, 0.00091553,  0.00085449], dtype=float32), 
    'sampling_rate': 48000
  },
  'sentence': 'n'ihu ọha mmadụ.', 
  'accent': 'nigerian', 
  'locale': 'igbo', 
}

Data Fields

client_id (string): An id for which client (voice) made the recording

path (string): The path to the audio file

audio (dict): A dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0].

sentence (string): The sentence the user was prompted to speak

accent (string): Accent of the speaker

locale (string): The locale of the speaker

Data Splits

The dataset has been subdivided into portions for dev, train and test.

Data Preprocessing Recommended by Hugging Face

The following are data preprocessing steps advised by the Hugging Face team. They are accompanied by an example code snippet that shows how to put them to practice.

Many examples in this dataset have trailing quotations marks, e.g “the cat sat on the mat.“. These trailing quotation marks do not change the actual meaning of the sentence, and it is near impossible to infer whether a sentence is a quotation or not a quotation from audio data alone. In these cases, it is advised to strip the quotation marks, leaving: the cat sat on the mat.

In addition, the majority of training sentences end in punctuation ( . or ? or ! ), whereas just a small proportion do not. In the dev set, almost all sentences end in punctuation. Thus, it is recommended to append a full-stop ( . ) to the end of the small number of training examples that do not end in punctuation.

from datasets import load_dataset
ds = load_dataset("benjaminogbonna/nigerian_common_voice_dataset", "igbo")
def prepare_dataset(batch):
  """Function to preprocess the dataset with the .map method"""
  transcription = batch["sentence"]
  
  if transcription.startswith('"') and transcription.endswith('"'):
    # we can remove trailing quotation marks as they do not affect the transcription
    transcription = transcription[1:-1]
  
  if transcription[-1] not in [".", "?", "!"]:
    # append a full-stop to sentences that do not end in punctuation
    transcription = transcription + "."
  
  batch["sentence"] = transcription
  
  return batch
ds = ds.map(prepare_dataset, desc="preprocess dataset")

Personal and Sensitive Information

The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in the Common Voice dataset.

Social Impact of Dataset

The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in the Common Voice dataset.

Reference/Disclaimer

Just to state it clearly, "the current languages and voices we have on the Nigerian Common Voice Dataset were not all collected from scratch". Infact, this wasn't the problem we set out to solve initially. We were working on a speech to speech (stt & tts) conversational model for Nigeria languages, but along the way we had a bottleneck:

1. The few data (audio) available were scattered and from different sources (Kaggle, Hugging Face, and many other websites).
2. The data weren't in the format required by the models.
3. Many of the audios had wrong or no corresponding transcriptions at all.

So while training our model, we had to gather them into one repository, structure them, clean them (remove/edit wrong transcriptions), and trim most of them to 30 seconds chunks.

We figured many people had the same issue, hence we uploaded it to Hugging Face and made it public.

Secondly, we haven't found any publicly available data (audios & transcriptions) for many Nigerian languages that we need (ex. Pidgin, etc). So the Nigerian Common Voice Dataset will be an ongoing project to collect as many languages & voices as possible.

Next, in order to add more languages and voices:

1. We will crowd-source from volunteers and contributors.
2. Take advantage of the hundreds of hours of Nigerian movies that are publicly available in different languages.

Our goal here is just to bring this data into one central repository and make it available to the public (researchers, developers, and all).

Contributions