A compact Arabic TTS model Nabra-7M-Distill has been released

Developers presented Nabra-7M-Distill — a specialized model for speech synthesis in the Arabic language. The system is built on an architecture with 7.48 million parameters and is designed to work in conditions of limited computational resources. The weight of the final model file is only 28.7 MB, which allows it to be run locally on a central processor (CPU). During testing, the model showed performance that exceeds the speed of the original "teacher" model by 4.2 times. At the same time, the speed of audio generation reaches 45x real-time. The model was obtained by distilling knowledge from a larger neural network, which made it possible to preserve the quality of synthesis while significantly reducing the amount of computation. The training process took 1 hour, as a result of which the model received a score of 8 points on the developers' internal efficiency scale. The release is oriented towards integration into mobile applications and embedded systems, where low latency and minimal RAM consumption are critical. The use of distilled models makes it possible to deploy high-quality voice interfaces on devices without dedicated graphics accelerators.