VoiceShelf brings Kokoro TTS offline to Android

This is a small but meaningful proof point for edge AI: open-weight speech models are getting light enough to power real consumer apps locally on phones, not just demos on desktops.

• –The interesting part is not the reader itself but the systems work behind it: EPUB parsing, chunking, Misaki G2P, Kokoro inference, and buffered streaming playback all running on Android
• –Reported 2.8x real-time on a Galaxy Z Fold 7 suggests modern flagship silicon can finally support smooth offline neural narration for at least compact English models
• –The current tradeoffs are obvious: a roughly 1 GB APK, likely steep device variance, and open questions around thermal throttling during long listening sessions
• –If broader chipset testing holds up, this could point toward a wider class of offline voice apps built on small open TTS models rather than recurring cloud subscriptions