I built an Android audiobook reader that runs Kokoro TTS fully offline on-device

A developer has created VoiceShelf, an Android application that performs text-to-speech conversion entirely on-device, turning EPUB and PDF files into audiobooks without any cloud processing. The app uses the Kokoro neural speech model running locally through a custom-built pipeline that includes EPUB parsing, sentence chunking, grapheme-to-phoneme conversion, and real-time audio streaming. On the developer's Samsung Galaxy Z Fold 7 with a Snapdragon 8 Elite chipset, the system achieves a real-time factor of 2.8, meaning it generates audio nearly three times faster than playback speed. The current APK is approximately 1GB in size, as it bundles the complete Kokoro model and specialized libraries to maintain quality on Android hardware.

VoiceShelf represents a significant step toward privacy-preserving and accessible AI tools, operating fully offline with features like screen-off narration and a sleep timer. The developer is actively seeking beta testers with recent Android flagships (2023 or newer) to gather performance data across different chipsets, including Snapdragon 8 Gen 2/3, Tensor, and Dimensity processors. Key testing goals include measuring the real-time factor on various devices and observing how thermal throttling affects performance during extended listening sessions of an hour or more. This community-driven testing phase aims to set realistic expectations for users at launch and demonstrate the feasibility of high-quality, on-device AI inference for everyday applications.