A compact TTS model Kokoro-7M-Distill has been released

Developers introduced Kokoro-7M-Distill — a lightweight speech synthesis model (TTS) designed to work on mobile devices. The model contains only 7.48 million parameters and takes up 28.7 MB of disk space. The architecture of the novelty is based on the distillation of knowledge from a larger Kokoro-82M model. During tests, the solution demonstrated a generation speed that exceeds the original's performance by 4.1 times. The processing time is 22 ms per second of audio stream. The model is focused on working in a fully autonomous mode (offline) without the need to access cloud servers. The low requirements for RAM allow the system to be integrated into smartphone applications with limited hardware resources. The release of Kokoro-7M-Distill expands the capabilities of local speech synthesis, providing high response speed with minimal resource consumption. The use of distillation made it possible to preserve the intelligibility of speech while significantly reducing the volume of weights, making the model suitable for embedded systems and mobile interfaces.