Mozilla has updated its Common Voice dataset, which includes pronunciation samples from over 200,000 individuals. The data is published as public domain (CC0). The proposed datasets can be used in machine learning systems to build models for speech recognition and synthesis. Compared to the previous update, the volume of speech material in the collection increased from 32.6 to 33.1 thousand hours of speech, of which 22.1 thousand hours have been validated. The number of supported languages has increased from 129 to 133 — adding Aragonese, Isindebele, Southern Sotho, and Tupi.
In preparing materials in English, 94.9 thousand people contributed 3,631 hours of speech (up from 93.9 thousand participants and 3,587 hours). The Belarusian language dataset includes 8,521 participants and 1,860 hours of speech material (up from 8,444 participants and 1,846 hours), the Russian dataset includes 3,365 participants and 281 hours (up from 3,296 participants and 278 hours), the Uzbek dataset includes 2,211 participants and 265 hours (the same number of hours as 2,200 participants), and the Ukrainian dataset has 1,120 participants and 114 hours (up from 1,104 participants and 114 hours).
The Common Voice project facilitates collaborative efforts to build a database of voice samples that account for the diversity of voices and speech styles. Users are invited to record displayed phrases or evaluate the quality of data added by other users. The accumulated database of recordings of various pronunciations of standard phrases in human speech can be used without restrictions in machine learning systems and research projects.
Source: opennet.ru
