v0.17.0

Added

• Fono recognises who is speaking (on-device speaker recognition). Enrol
  your voice once and Fono can tell it's you when you dictate or talk to the
  assistant, tagging each entry in your history with the recognised person.
  Enrol from the browser settings page (the new Speakers page) or the
  fono speaker commands, then use "test my voice" — on the page or via
  fono speaker test — to see how reliably Fono tells you apart from other
  people: it reads a few short clips, charts your scores against a shipped
  comparison set, recommends a sensitivity, and works out the automatic
  match cut-off for you (leaning toward keeping other people out). You can
  review each recorded sample, see its length, loudness, noise, and how well
  it matches the others, and drop weak clips in one click — Fono always keeps
  enough good audio behind and never deletes your last sample without asking.
  Recognition runs at the same time as transcription so it adds no delay, and
  the model loads only once. When the assistant recognises you it's quietly
  told who it's talking to, so it can answer more personally.

  Your voiceprint never leaves the machine: it is computed locally, stored in
  an on-device voiceprint database, and is never attached to any cloud
  transcription or cleanup request — only the recognised name is ever saved,
  alongside your history. The full walkthrough is in the new Speakers guide
  (docs/speakers.md), linked from the README and docs index, and the
  [speaker] config block is documented in the configuration reference.

• Let other apps on your network use Fono's voices and transcription.
  Building on the local LLM server from 0.13.0, Fono can now also serve its
  speech-to-text and text-to-speech to other apps and machines on your LAN,
  so one Fono box can voice a whole network. Access is gated by inbound API
  keys you manage yourself, with per-key usage tracking, and fono doctor
  (and the keys check) now confirm a configured key actually works. As always
  the server binds to loopback by default and only reaches the LAN when you
  opt in.

• Settings-page improvements. Pick and try your local voice engine right
  from the settings page with a built-in voice tester (which now plays the
  engine you actually selected), see system health inline with a one-click
  fono doctor view, and configure Network options for the cleanup and
  assistant stages. The "watch it think" activity overlay was rebuilt into a
  clearer, livelier readout, and the assistant now shows its replies on
  screen even when no voice output is set up.

Changed

• Fono is now described as what it has become: a complete, open-source
  voice-AI stack in one small binary — speech-to-text, natural voices, a
  local LLM, wake word, and speaker recognition, local by default with
  per-stage cloud switching. This release retells that story across the
  README, the roadmap, and the packaging descriptions (Debian, AUR,
  SlackBuild, Nix, desktop entries). Copy and metadata only — no change to
  how Fono works.

Fixed

• Recording your "test my voice" clips no longer fails with an "invalid or
  oversized JSON body" error. The settings server capped every upload at
  1 MB, which a handful of short voice clips could exceed; voice enrolment and
  testing now accept a larger local-only upload while ordinary settings
  changes keep the tight limit.

• Voice enrolment no longer fails silently, and now shows you a review
  step before it saves, so you can confirm the captured audio first.

• The Node.js 20 deprecation warning in the container build is gone — the
  GHCR login step moved to docker/login-action@v4, which runs on Node.js 24.

Internal

• Release artefacts are simpler: one consolidated SHA256SUMS instead of a
  .sha256 file per download. fono update now verifies each download
  against the matching row in SHA256SUMS (and still accepts the old
  per-asset sidecar for historical releases), so nothing changes for the
  self-updater — there are just fewer files on each release.

Full Changelog: v0.16.0...v0.17.0