The Common Voice dataset consists of a unique MP3 and corresponding text file. Many of the 28117 recorded hours in the dataset also include demographic metadata like age, sex, and accent that can help improve the accuracy of speech recognition engines.

The dataset currently consists of 18651 validated hours in 112 languages, but more voices and languages are always added. Take a look at the Languages page to request a language or start contributing.

Supported Tasks and Leaderboards

The results for models trained on the Common Voice datasets are available via the 🤗 Speech Bench

Languages

Abkhaz, Albanian, Amharic, Arabic, Armenian, Assamese, Asturian, Azerbaijani, Basaa, Bashkir, Basque, Belarusian, Bengali, Breton, Bulgarian, Cantonese, Catalan, Central Kurdish, Chinese (China), Chinese (Hong Kong), Chinese (Taiwan), Chuvash, Czech, Danish, Dhivehi, Dioula, Dutch, English, Erzya, Esperanto, Estonian, Finnish, French, Frisian, Galician, Georgian, German, Greek, Guarani, Hakha Chin, Hausa, Hill Mari, Hindi, Hungarian, Icelandic, Igbo, Indonesian, Interlingua, Irish, Italian, Japanese, Kabyle, Kazakh, Kinyarwanda, Korean, Kurmanji Kurdish, Kyrgyz, Lao, Latvian, Lithuanian, Luganda, Macedonian, Malayalam, Maltese, Marathi, Meadow Mari, Moksha, Mongolian, Nepali, Norwegian Nynorsk, Occitan, Odia, Pashto, Persian, Polish, Portuguese, Punjabi, Quechua Chanka, Romanian, Romansh Sursilvan, Romansh Vallader, Russian, Sakha, Santali (Ol Chiki), Saraiki, Sardinian, Serbian, Slovak, Slovenian, Sorbian, Upper, Spanish, Swahili, Swedish, Taiwanese (Minnan), Tamazight, Tamil, Tatar, Thai, Tigre, Tigrinya, Toki Pona, Turkish, Turkmen, Twi, Ukrainian, Urdu, Uyghur, Uzbek, Vietnamese, Votic, Welsh, 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 Hindi config, simply specify the corresponding language config name (i.e., "hi" for Hindi):

from datasets import load_dataset

cv_14 = load_dataset("mozilla-foundation/common_voice_14_0", "hi", 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

cv_14 = load_dataset("mozilla-foundation/common_voice_14_0", "hi", split="train", streaming=True)

print(next(iter(cv_14)))

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

cv_14 = load_dataset("mozilla-foundation/common_voice_14_0", "hi", split="train")
batch_sampler = BatchSampler(RandomSampler(cv_14), batch_size=32, drop_last=False)
dataloader = DataLoader(cv_14, batch_sampler=batch_sampler)

Streaming

from datasets import load_dataset
from torch.utils.data import DataLoader

cv_14 = load_dataset("mozilla-foundation/common_voice_14_0", "hi", split="train")
dataloader = DataLoader(cv_14, 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 13 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, age, client_id, up_votes, down_votes, gender, locale and segment.

{
  'client_id': 'd59478fbc1ee646a28a3c652a119379939123784d99131b865a89f8b21c81f69276c48bd574b81267d9d1a77b83b43e6d475a6cfc79c232ddbca946ae9c7afc5', 
  'path': 'et/clips/common_voice_et_18318995.mp3', 
  'audio': {
    'path': 'et/clips/common_voice_et_18318995.mp3', 
    'array': array([-0.00048828, -0.00018311, -0.00137329, ...,  0.00079346, 0.00091553,  0.00085449], dtype=float32), 
    'sampling_rate': 48000
  }, 
  'sentence': 'Tasub kokku saada inimestega, keda tunned juba ammust ajast saati.', 
  'up_votes': 2, 
  'down_votes': 0, 
  'age': 'twenties', 
  'gender': 'male', 
  'accent': '', 
  'locale': 'et', 
  'segment': ''
}

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

up_votes (int64): How many upvotes the audio file has received from reviewers

down_votes (int64): How many downvotes the audio file has received from reviewers

age (string): The age of the speaker (e.g. teens, twenties, fifties)

gender (string): The gender of the speaker

accent (string): Accent of the speaker

locale (string): The locale of the speaker

segment (string): Usually an empty field

Data Splits

The speech material has been subdivided into portions for dev, train, test, validated, invalidated, reported and other.

The validated data is data that has been validated with reviewers and received upvotes that the data is of high quality.

The invalidated data is data has been invalidated by reviewers and received downvotes indicating that the data is of low quality.

The reported data is data that has been reported, for different reasons.

The other data is data that has not yet been reviewed.

The dev, test, train are all data that has been reviewed, deemed of high quality and split into dev, test and train.

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("mozilla-foundation/common_voice_14_0", "en", use_auth_token=True)

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")

Dataset Creation

Curation Rationale

[Needs More Information]

Source Data

Initial Data Collection and Normalization

[Needs More Information]

Who are the source language producers?

[Needs More Information]

Annotations

Annotation process

[Needs More Information]

Who are the annotators?

[Needs More Information]

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.

Considerations for Using the Data

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.

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

Public Domain, CC-0

Citation Information

@inproceedings{commonvoice:2020,
  author = {Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and Kohler, M. and Meyer, J. and Morais, R. and Saunders, L. and Tyers, F. M. and Weber, G.},
  title = {Common Voice: A Massively-Multilingual Speech Corpus},
  booktitle = {Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020)},
  pages = {4211--4215},
  year = 2020
}

Downloads last month: 245

Papers with Code

Homepage:

commonvoice.mozilla.org

Models trained or fine-tuned on mozilla-foundation/common_voice_14_0