Wikispeech/Speech data collection

Speech Data Collection

Wikispeech develops tools and infrastructure for collecting speech data. This is a key component in making speech technologies more accessible, inclusive, and available under free licences.

Why collect speech data

High-quality speech data is essential for building text-to-speech (TTS) systems in new languages. Our goal is to make Wikispeech available in as many languages as possible. Beyond this, we aim to create a diverse and openly available speech data repository that includes:

Words, short phrases, and longer sentences
Different dialects and language variations
Speech from people with speech impairments
Recordings with varying sound quality, reflecting real-world conditions

Each recording is associated with metadata, such as:

Language and dialect
Type of speech (word, sentence, etc.)
Audio quality
Age and other relevant attributes for research

Open data for research and development

We aim to make both recordings and metadata available under free licences. This allows universities, researchers, and organisations to:

Improve text-to-speech systems
Develop better speech recognition technologies
Build more inclusive and accessible digital tools

Enabling voice-based access to knowledge

As speech recognition becomes more common, voice is increasingly used to access digital content. While Wikispeech enables users to listen to Wikipedia, navigation still often depends on screens, keyboards, or touch interfaces.

We see a future where users can interact with knowledge through natural voice interaction — asking questions and receiving answers conversationally.

With sufficient speech data, it is also possible to:

Improve the naturalness of synthetic speech (intonation, rhythm, emphasis)
Support more languages and dialects
Contribute to documenting language development and variation over time

Challenges

Collecting and sharing speech data comes with important challenges.

Privacy and sensitive data

Some metadata — such as information about speech impairments — can be highly sensitive. We must ensure that all data collection and processing fully respects users’ privacy and rights.

Storage and accessibility

We need sustainable and cost-effective solutions for storing and distributing large amounts of speech data, while maintaining integrity and accessibility.