# KanTV

An Android workbench for studying on-device AI in real scenarios such as online TV, combining llama.cpp, whisper.cpp and a customized FFmpeg to run offline LLMs and real-time transcription while watching or recording video.

KanTV is an open-source Android project for practicing on-device AI in real usage scenarios: playing and recording online TV with a customized FFmpeg and ExoPlayer, while simultaneously running real-time English subtitles powered by whisper.cpp. It packages several built-in AI models that run fully offline and can be downloaded directly in the app, including Whisper for speech-to-text, and Qwen, Gemma and SmolVLM2 text and multimodal LLMs.

It is intended as a turn-key workbench both for AI researchers validating algorithms, models and performance on Android, including ASR, LLM and multimodal benchmarks, and for beginners learning on-device AI, and includes an early open-source FastRPC-based ggml-hexagon backend for running ggml models on Qualcomm Hexagon NPUs. Java, MIT licensed.

Website: https://github.com/zhouwg/kantv