I spent a few weeks figuring out how to run a real neural TTS model entirely on-device — no server, no API key, nothing leaving your phone.
It uses Kokoro running on MLX, Apple’s on-device ML framework.
The tricky part wasn’t running the 82M-parameter model, but making it work with large documents like full books and long PDFs. A naive approach either runs out of memory or makes you wait a long time before hearing the first word. It took a lot of iteration to get it streaming smoothly from the first sentence.
You can tap any sentence to jump straight to it. The app re-synthesizes instantly from that point — no scrubbing and no waiting for the whole chapter to reload.
Because everything runs locally on your phone, there’s no signup required and no usage limits or “fair use” caps. You can generate as much audio as your device can handle.
One quirk worth knowing: iOS suspends GPU access when an app goes to the background, so synthesis stops if the screen locks. The workaround is keeping the screen on while the app is open — similar to how navigation apps keep the display awake. Not ideal, but it’s the trade-off for running a real neural model entirely on-device.
Features:
English and Spanish voices
PDF and EPUB, MD, TXT, Website Article support
Export to MP3
Requires *iPhone 15 Pro or newer.
jantheman•2h ago
It uses Kokoro running on MLX, Apple’s on-device ML framework.
The tricky part wasn’t running the 82M-parameter model, but making it work with large documents like full books and long PDFs. A naive approach either runs out of memory or makes you wait a long time before hearing the first word. It took a lot of iteration to get it streaming smoothly from the first sentence.
You can tap any sentence to jump straight to it. The app re-synthesizes instantly from that point — no scrubbing and no waiting for the whole chapter to reload.
Because everything runs locally on your phone, there’s no signup required and no usage limits or “fair use” caps. You can generate as much audio as your device can handle.
One quirk worth knowing: iOS suspends GPU access when an app goes to the background, so synthesis stops if the screen locks. The workaround is keeping the screen on while the app is open — similar to how navigation apps keep the display awake. Not ideal, but it’s the trade-off for running a real neural model entirely on-device.
Features:
English and Spanish voices PDF and EPUB, MD, TXT, Website Article support Export to MP3 Requires *iPhone 15 Pro or newer.
Free for 30 days.*
https://apps.apple.com/us/app/ghost-reader-ai/id6759826819