A language programme targeting more than 100 languages, including 50 African languages.
The numbers describe the planned language programme. Individual languages will only be presented as production-ready after speech, translation, voice and video quality have been evaluated.
Supporting a language is more than translating text.
Video localisation requires a language to work across several AI tasks, not just one.
For each target language, development can involve speech recognition, translation, speech generation, pronunciation testing, duration control, code-switching, accent and dialect handling, and evaluation by people who understand the language naturally.
Speech data
Audio and transcripts that represent real speakers and recording conditions.
Translation
Meaning, terminology and context across source and target languages.
Pronunciation
Natural target-language speech, including names and local terms.
Timing
Speech duration that can fit naturally into the source video.
Human evaluation
Native-speaker review of accuracy, naturalness and known failure cases.
The African-language focus is intended to address both major and underrepresented language communities.
Examples under consideration span West, East, Central, Southern and North Africa.
The final supported list will depend on data availability, model performance and evaluation. The examples below are programme targets and research candidates, not a claim of current production support.
Regional examples in the current programme direction.
Use the filters to view examples. This directory is not a production-availability list.
Each language will move through defined stages before production use.
This prevents a language from being presented as supported simply because a model can generate a sample.
Data, baseline models and language-specific requirements are investigated.
Speech, translation and voice components are tested on controlled examples.
Native-language review and audiovisual quality tests are performed on broader material.
The language is released only when the full workflow reaches an acceptable quality level.