Piper - Neural TTS seems to be ready to be daily driver for narration

Piper Neural TTS has released version 1.0.3 few days ago on the App Store: https://apps.apple.com/us/app/piper-neural-tts/id6759636010 .

Previous version 1.0.2 was interesting, but had at least two bugs that would prevent it from being a daily driver:
- it reversed speed command directions
- apostrophes were not properly parsed.

Those bugs are now fixed and after some experiments I think I can say that the app is now ready to be used for narration as a daily driver, but only with a proper setup.

Most notably voice quality may significantly vary among voices and some voices may leave poor impression. Based on some community sources I currently recommend two voices: Lessac and LJSpeech in their mid variants. If your device performs poorly try low variants.

I have tested voices with my app Speech Central on the iPhone 15 Pro. Most of the time it worked flawlessly except in rare occasions of very long sentences where latency/pause was somewhat noticeable. I would rate that as "manageable" but definitely mid-annoying. That leads to conclusion that Piper may have latency problem with longer text inputs, which may be much more noticeable in some other reader apps as their reading strategy may be different. Further it certainly works even better on more recent iPhones. I would expect a noticeable performance degradation on any older model (possibly even iPhone 15), but as said low quality voices may help on some of those devices.

I wouldn't recommend this as a daily driver for VoiceOver voice, though for some people and in some use cases it may work, just be warn not to expect too much and not to be disappointed. Especially that sensitivity to reading strategy will come on top with VoiceOver in some cases, and I would expect that more robotic voices that have extremely low latency and better prosodies at high speeds are better fits for well trained blind users.