Common Voice

Common Voice
Name	Common Voice
Developer	Mozilla
Released	2017
Programming language	Python (programming language), JavaScript
Operating system	Linux, Windows, macOS
License	Creative Commons

Common Voice Common Voice is a multilingual voice dataset and open-source project initiated to improve speech recognition and synthesis technologies. Launched to broaden access to audio data, it engages volunteers, researchers, and corporations to crowdsource voice recordings and validation contributions for use by projects across the technology and research sectors. The project intersects with initiatives in natural language processing led by organizations, universities, and technology companies worldwide.

Overview

Common Voice aggregates crowdsourced speech recordings, transcriptions, and validation metadata to create large-scale corpora intended for training automatic speech recognition and text-to-speech systems. The dataset complements academic datasets produced by institutions such as Massachusetts Institute of Technology, Stanford University, Carnegie Mellon University, University of Cambridge, and University of Oxford while serving industry actors including Google, Microsoft, Amazon (company), Apple Inc., and IBM. Contributors include volunteers, nonprofit partners like Electronic Frontier Foundation, and regional organizations such as Wikimedia Foundation, European Language Resources Association, and language-specific groups like Basque Government, SIL International, and Institut National de la Langue Française.

History and Development

The project was announced by Mozilla in 2017 with engineering leadership drawing on prior speech efforts from teams influenced by research at Bell Labs, Xerox PARC, and academic labs at University of Toronto, University of Edinburgh, and Johns Hopkins University. Early milestones involved collaborations with language communities represented by institutions such as UNESCO, European Commission, and the African Academy of Languages. Funding, partnerships, and publicity engaged actors like Mozilla Foundation, philanthropic funders aligned with Knight Foundation, and conferences including NeurIPS, ACL (Association for Computational Linguistics), ICASSP, and EMNLP. Over successive releases the project incorporated advances referenced in work from groups including DeepMind, Facebook AI Research, OpenAI, and research consortia at Allen Institute for AI.

Data Collection and Dataset

Data collection uses web platforms and mobile applications to crowdsource recordings and validations from volunteers in language communities ranging from widely spoken languages like English language, Spanish language, Mandarin Chinese, Arabic language, Hindi language, and Portuguese language to under-resourced languages such as Welsh language, Basque language, Catalan language, Yoruba language, Hausa language, Amharic language, Kurdish language, Scottish Gaelic, Irish language, and Maori language. The dataset includes parallel efforts similar in scope to corpora curated by Linguistic Data Consortium, ELRA, and projects like Common Crawl and LibriSpeech. Contributions are validated via community review workflows comparable to annotation pipelines at Zooniverse and quality-assured using benchmarks inspired by evaluations at Speech Processing Conference events. Large-scale speech models trained on the dataset have been compared with benchmarks from LibriVox and datasets used in competitions such as those hosted by Kaggle and Google Research.

Technology and Usage

Researchers and developers apply Common Voice data to train models based on architectures popularized by groups like DeepMind (e.g., transformer variants), Google Research (e.g., wav2vec), Facebook AI Research (e.g., fairseq), and teams at OpenAI contributing end-to-end speech systems. Implementations leverage toolkits and frameworks from TensorFlow, PyTorch, Kaldi, and libraries maintained by organizations including Hugging Face. Use cases span speech recognition in devices by companies like Samsung Electronics, voice assistants from Alphabet Inc. subsidiaries, assistive technologies advocated by Human Rights Watch, and localization efforts at platforms such as Mozilla Firefox and Wikipedia. Integration patterns mirror deployment scenarios seen in products from NVIDIA, Intel Corporation, and cloud services from Amazon Web Services, Google Cloud Platform, and Microsoft Azure.

Licensing and Privacy

The dataset is distributed under permissive licenses inspired by open-content frameworks like Creative Commons to enable use by academic, nonprofit, and commercial actors including Universities and corporations. Governance and data handling practices reference privacy frameworks advanced by organizations such as Electronic Frontier Foundation, Privacy International, and regulatory contexts including legislation influenced by the European Union and directives shaped around data protection like General Data Protection Regulation. Consent and contributor controls align with community standards promoted at conferences like RightsCon and guidelines from bodies such as IEEE Standards Association.

Impact and Reception

Common Voice has influenced academic research published in venues like ACL, ICLR, NeurIPS, and ICASSP and informed product development at startups and companies showcased at events such as CES and Mobile World Congress. Reception among language preservation groups including UNESCO and regional culture institutions has been positive for enabling resources for endangered and minority languages. Critiques from researchers and advocacy organizations such as Electronic Frontier Foundation and Amnesty International have focused on representativeness, demographic balance, and consent practices, prompting ongoing improvements inspired by standards from ISO and community feedback channels including forums associated with Mozilla Foundation.

Category:Speech recognition datasets