Asia-Plus Media Group and the startup Re-skill have announced the beginning of a partnership aimed at training a neural network for Tajik speech recognition and synthesis.
As part of the collaboration, Asia-Plus will provide an audio archive of news programs and shows in Tajik to support the development of an open neural network for text-to-speech (TTS) and automatic speech recognition (ASR) in the Tajik language, currently being developed by the Re-skill team.
This partnership will use hundreds of hours of live Tajik speech—such as news, interviews, talk shows, and analytical programs—as a foundation for training the neural network model.
In return, Asia-Plus will receive official partner status, access to all versions of the tool, and the opportunity to participate in the project’s further development. Additionally, Asia-Plus will report on the project’s progress, use the tool for content production, and share the results across its media platforms.
Re-skill TTS is the first open system for Tajik speech recognition and synthesis. The early version of the model is already available on the HuggingFace platform. Despite having only about 35 hours of training data, the neural network can already produce Tajik speech with natural intonation and rhythm.
For Asia-Plus, participation in this project is not just a technological contribution, but also a reflection of its cultural and social mission.
“This is not just a technological partnership, but a contribution to promoting our national language in the digital world,” emphasized Asia-Plus Director-General, Umed Babakhanov. “Today, it is especially important to create such solutions in Tajik so that our language is heard in modern services and is available in voice assistants and educational applications. We are pleased to support the Re-skill team in this endeavor and are ready to continue supporting similar initiatives in the future.”
Farrukh Umarov, the founder of Re-skill, believes that this partnership with Asia-Plus will be a key step in the development of Tajik speech technology.
“We believe that Tajik deserves the same quality technological representation as any major world language. Our goal is to make a tool that will allow anyone to integrate Tajik into their services and applications,” he said. “We have made this model available to the community to foster collaborative development and the collection of more data. Cooperation with Asia-Plus provides us with a unique and valuable source of audio materials, helping lay the technological foundation for the next generations of voice assistants and educational products in Tajik.”
How the system works
The neural network is based on the architecture of large language models, integrated with modern audio codecs. The system recognizes speech, generates a response using large language models (LLM), and then synthesizes it in Tajik. This cascading approach enables the production of not just synthesized sound, but expressive, dialogic speech.
The model can be run locally, for example, directly on your MacBook.
Re-skill has also launched its own Whisper speech recognition model, which offers high transcription accuracy and supports multiple languages.
Currently, Tajik is nearly absent from global digital platforms. This project is laying the foundation for its comprehensive presence in voice interfaces, educational products, media, and services.
Руcский