Why this page exists: the lone δ½ in the app still doesn't sound right to you β and you're the judge, not the numbers. Under each word: what the app plays today, then the best take each candidate source could produce. Play δ½ first, side by side with Pleco's δ½ as your reference.
The percentages are the app's own tone-engine "shape match" against the canonical tone template β advisory only. A take can measure well and still sound wrong (that's how we got here). Trust your ears; the measurements are just there so you can see what the machine thinks.
| Source | Mean shape match | Right tone (of 8) | Full-library coverage | Licensing |
|---|---|---|---|---|
| App now (Piper, as shipped mq-v10) | 83% | 8/8 | already in app | β |
| Piper best-of-net (regenerated) | 84% | 8/8 | all 148 units generatable | open-source TTS (current engine) |
| Kokoro TTS | 57% | 4/8 | all units generatable in principle | Apache-2.0 model |
| Google Translate TTS | 59% | 7/8 | any text, via Google endpoint | β οΈ ToS-grey β unofficial endpoint; not cleared to ship |
| Human recordings (Commons) | 55% | 6/8 | 86 of 92 library characters found (Lingua Libre + Zh series) | CC BY-SA 4.0 mostly; some CC0 / CC BY (attribution needed) |
Licensing notes. Human recordings come from Wikimedia Commons (Lingua Libre + the older Zh series): mostly CC BY-SA 4.0 β free to use with attribution and share-alike; a few are CC0/public domain. Google Translate's voice comes from an unofficial endpoint β it works today, but its terms are a grey zone for shipping inside an app; it's here so you can hear the quality bar, flagged not recommended without your explicit call. Kokoro is an Apache-2.0 open model. Piper is the app's current open-source engine.
Pleco: its audio is licensed content locked inside the Pleco app β it can't be reused as a source. It's your reference standard only: the sound these candidates are trying to match.