Data and models for African langauges
Over the course of several months, researchers from Coqui collaborated with a global team of academics, language activists, and technologists in order to create high quality Text-to-Speech for six African languages. This blog post covers who this excellent team was, what we did, and how you can use these new voices for yourself. All the synthetic voices discussed here are available under a Creative Commons BY-SA 4.0 License — a free, open, and commercial friendly license.
There are literally thousands of languages spoken in Africa, and as such this current work is only the tip of the iceberg. Nevertheless, we hope our work inspires others to create new open, synthetic voices for as many of Africa’s languages as possible. Coqui’s TTS can be fine-tuned to any new language, even with tiny amounts of data, regardless of the alphabet or grammar or linguistic attributes. The more data the better, as you will see (and hear) here. Data is almost always the bottleneck in deep learning, and in this blogpost we’ll discuss how we found raw data that wasn’t ready for TTS, and massaged it into a place where beautiful, high-fidelity synthetic voices could be built. Once the data was ready, training the models was a piece of cake.
This project wouldn’t have been possible without our collaborators. Specifically, the excellent Masakhane NLP community is what brought us all together in the first place. We eagerly look forward to more Coqui + Masakhane collaborations in the future! If you want to see the future of natural language technology (especially for African languages), Masakhane is the place to be.
Without further ado, here’s the team of individuals that brought these voices into reality (in alphabetical order):
- Alp Öktem, Apelete Agbolo, Bernard Opoku, Chris Emezue, Colin Leong, Daniel Whitenack, David Ifeoluwa Adelani, Edresson Casanova, Elizabeth Salesky, Iroro Orife, Jesujoba Alabi, Jonathan Mukiibi, Josh Meyer, Julian Weber, Perez Ogayo, Salomey Osei, Salomon Kabongo, Samuel Olanrewaju, Shamsuddeen Muhammad, Victor Akinode
Featuring activists and technologists from:
- CLEAR Global, Col·lectivaT, Ewegbe Akademi, Masakhane, Niger-Volta LTI, SIL International
Featuring academic researchers from:
- Carnegie Mellon University, Johns Hopkins University, Kwame Nkrumah University of Science and Technology, Leibniz Universität, Makerere University, Saarland University, Technical University of Munich, University of São Paulo
The six new languages added to TTS are:
|Number of Speakers
|Niger-Congo / Kwa
|Afro-Asiatic / Chadic
|Niger-Congo / Bantu
|Niger-Congo / Akan
|Niger-Congo / Akan
|Niger-Congo / Volta-Niger
Both the “Number of Speakers” and “Classification” columns come from Ethnologue. These six languages are all tonal, come from two of the largest language families in Africa (Niger-Congo and Afro-Asiatic), and are spoken primarily in Central and West Africa. Needless to say, there are a lot of people speaking these languages in a huge geographic area. By releasing these models under an open Creative Commons license, we hope they will be immediately useful to speakers of these languages.
As with all machine learning projects, data is the starting point. This entire collaboration spawned from a short URL posted into a chatroom: open.bible. A researcher from Coqui was hanging out with the folks from Masakhane in their slack group when someone posted the link saying something like “looks like some cool data!“. In no time at all, a lively discussion ensued. The data was absolutely beautiful. All the data was explicitly licensed under CC-BY-SA, made of hours and hours of high-quality recordings from professional voice actors. This was without exaggeration the highest-quality voice data for speech synthesis Coqui had ever found in the open — for any language.
There was only one problem — the original audio files are too long for training TTS models. The audio was saved as chapters (from the Bible), which were several minutes long each. It’s best to train synthetic voices with audio clips under 30 seconds long, so we couldn’t use the data out of the box. The intuitively simple task of breaking chapters into verses is not so simple in practice, and it requires significant compute power. Nevertheless, over a couple months and more than a couple cups of coffee, we aligned the recordings to the verse-level, and then we extracted only the best data. The resulting datasets will be released under the same CC-BY-SA 4.0 license, as well as our research paper detailing how we made it possible. Both the dataset release and publication of our methods are slated for INTERSPEECH 2022.
All models discussed here can be used from:
Keep an eye our for our INTERSPEECH paper for all the technical details on how we created the datasets and trained the models. Until then, take the models and do something great! We want to thank again all the folks who helped make this possible, especially the Masakhane community for bringing us all together. We want to also acknowledge the individuals who spent hours and hours narrarating the Bible in these languages for the Open.Bible project, and for releasing the recordings under a Creative Commons license in the first place. We also want to thank the team of folks at Biblica for taking such care to record, organize, and release the raw data. They did not participate in this research, but it would not have been possible without them. On a last, but important note, anyone using these synthetic voices should be using them to create more good in the world, and no more harm. Out of respect to the original voice actors and the nature of the original recordings, we want the voices to be used to help people, and we’re sure they can find great use in places like education and accessibility. For example, these voices can easily be used to create audiobooks for people who can’t see or read well, and make reading more fun for students. If you use these models, let us know what great things you’re creating in the world!