v0.17.1

Changed

• The built-in text-to-speech now uses the multilingual Supertonic voice by
  default. It's one compact download that speaks many languages — including
  Romanian — in a single natural voice, and it stays comfortably faster than
  real time even on older laptops. This replaces the previous setup that
  automatically switched between English-only and per-language voices behind
  the scenes. The older Piper and Kokoro voices are still available if you
  prefer them — just pick one in Settings under the on-device voice engine —
  but the automatic voice-switching option has been removed, since the new
  default covers every language on its own. Existing setups keep working;
  nothing needs to be reconfigured.

Added

• Run large local assistant models that don't fit in your RAM. Fono can now
  use big on-device chat models — such as the 9.6 GB Gemma and 11.7 GB Qwen
  builds — by streaming their weights from disk as they are needed instead of
  loading the whole model into memory. That means a model can be larger than the
  free RAM on your machine and still run, loading in about a second rather than
  stalling or running you out of memory. Pick one as your assistant and Fono
  downloads it for you, or point it at a GGUF you downloaded yourself; a name
  copied straight from Hugging Face works either way. These models trade some
  speed for the ability to run big models locally, so replies come a little
  slower than the small default — the small default is still what ships out of
  the box.

• Two new controls for the built-in voice in Settings. A Speed control
  (slower / normal / faster) and an optional extra passes toggle that
  trades a little processing time for a small quality boost (off by default,
  which already sounds great for most people).

Fixed

• The diagnostic log now shows the exact words speech-to-text produced
  before your personal-vocabulary corrections are applied. Previously the
  stt.raw log line already reflected the corrected text, so there was no way
  to see the original wording; it now prints the original and, when a
  correction fired, the corrected form alongside it.

• Switching to a local assistant now downloads its model automatically, the
  same way switching the speech, cleanup, and voice engines already does.
  Before this it could fail with "model not found" until you fetched the file
  by hand.

• A local model name copied from Hugging Face — whatever the capitalisation,
  and with or without a trailing "-GGUF" — now resolves to the right file, so
  the automatic download and the loader always agree on where it lives.

Full Changelog: v0.17.0...v0.17.1