Common Voice 22.0: Multilingual Speech Corpus for ASR in 20+ Languages

Mozilla Common Voice Corpus 22.0 is a multilingual speech dataset featuring audio recordings and text transcriptions across a wide array of global languages. This version is an unofficial conversion of the Mozilla project data provided by fsicoli and updated in August 2025. It includes data for dozens of languages including Arabic, Bengali, and Chinese.

Use Cases

• Training Automatic Speech Recognition (ASR) models using the audio and text pairs associated with the 'Task Categoriesautomatic Speech Recognition' tag
• Language identification for low-resource tongues using specific language tags like 'Languagebas' (Basaa) and 'Languagecv' (Chuvash)
• Dialectal variation research comparing samples from 'Languagecnh' (Hakha Chin) and 'Languageas' (Assamese)

Strengths

• Includes rare and low-resource languages such as Abkhaz, Basaa, and Chuvash
• Covers over 20 distinct language tags including Amharic and Bashkir
• Derived from the Mozilla Common Voice project, a standard benchmark for open-source speech data

Limitations

• Unofficial distribution which may lack the standard metadata or validation found in the official Mozilla release
• Unknown record count and file size make it difficult to assess storage requirements before download

Provenance

Source

Mozilla Common Voice (https://commonvoice.mozilla.org/)

Collection Method

Downloaded and converted from the Mozilla Common Voice project website

Freshness

Updated as of August 11, 2025.

Geography

Global