Skip to content

omamobile-voice 0.1.0, omamobile-shell 0.2.82: voice works out of the box - #355

Merged
derekross merged 5 commits into
mainfrom
voice-out-of-the-box
Nov 2, 2026
Merged

derekross merged 5 commits into
mainfrom
voice-out-of-the-box

Conversation

@derekross

@derekross derekross commented Nov 2, 2026 •

Copy link
Copy Markdown
Owner

A fresh release install had no Voxtype: the voice bar said "Voxtype isn't running. Install it with Install > AI > Dictation", and that install downloads the model, so dictation needed a network and a setup step. This makes a fresh image ship voice ready, offline on first use.

What Omarchy's Install > AI > Dictation does (omarchy-voxtype-install, omarchy 4.0.6)

omarchy-pkg-add wtype voxtype-bin, copy $OMARCHY_PATH/default/voxtype/config.toml to ~/.config/voxtype/, voxtype setup --download --no-post-install (base.en into ~/.local/share/voxtype/models), voxtype setup gpu --enable if Vulkan (aarch64's voxtype-bin has one CPU binary, so not here), voxtype setup systemd (a user unit, enable, start).

Changes

  • packages/omamobile-voice 0.1.0 (new): fetches ggml-base.en.bin at build time from a pinned commit of huggingface.co/ggerganov/whisper.cpp (5359861c…) with a pinned sha256 (a03779c8…), and installs it in /usr/share/omamobile/voice/models. Nothing binary is committed. MIT (OpenAI's Whisper, converted by whisper.cpp); the license ships with it. voxtype-bin is an optional dependency only, so Omarchy's Remove > Dictation (pacman -Rns voxtype-bin) still works. test/all checks the pin, the directory against setup-user, and the root image lists; OMAMOBILE_VOICE_FETCH=1 also checks the server's file against the sha256 with a HEAD request.

  • Root image: voxtype-bin (in [omarchy] aarch64 stable and edge, 1.1.1-2; it ships /usr/lib/systemd/user/voxtype.service) goes in rootfs/packages.base, because packages.omamobile accepts only [omamobile] packages. omamobile-voice goes in packages.omamobile. wtype is already in Omarchy's list, and packages.omarchy-skip doesn't drop it. tests/rootfs-smoke checks for voxtype, its unit and the model when the MANIFEST lists omamobile-voice.

  • omamobile-shell 0.2.82 (0.2.80 and 0.2.81 are taken by omamobile-shell 0.2.98: the user setup makes Omarchy's default keyring #343 and omamobile-shell 0.2.81: Omawrite instead of btop on a fresh home screen #353): omamobile-setup-user does what Omarchy's install does, minus the download, once per user and only when voxtype and the shipped model are both installed. It copies Omarchy's Voxtype config if the user has none, links the model into ~/.local/share/voxtype/models (Voxtype looks nowhere else) unless a file of that name is already there, runs systemctl --user enable voxtype.service (no --now: firstboot runs this with no session, and at login it runs before graphical-session-pre; the enable has its own marker, set only on success, and is retried at each login while voxtype and the model are there, so a firstboot without a user manager recovers), and marks Omarchy's voxtype-install-invitation done, so the "Install Dictation" toast doesn't show after the first update. The existing step then adds the context cut to the new config. The voice bar's error leads with starting voxtype.service. Optional dependencies, a README Voice input section, and tests in test/setup-user.sh: a new user, a second login, a user's own config and model, a failed enable retried, no voxtype, no model.

  • Docs: the release notes draft and VOICE-PLAN say voice works out of the box.

  • Round 40 review: origin/main merged in (shell stays 0.2.82 over omamobile-shell 0.2.81: Omawrite instead of btop on a fresh home screen #353's 0.2.81), the enable retry above, the README counts the model twice and says how to remove dictation fully (sudo pacman -R omamobile-voice after Remove > Dictation), LICENSE.whisper carries whisper.cpp's notice next to OpenAI's, and the PKGBUILD and README note the first build needs huggingface.co (makepkg caches it in SRCDEST; it has no fallback for one source, so a copy on our server would replace the URL).

Model decision

Ship base.en only: 147,964,211 bytes (141 MiB), and it barely compresses. It counts twice in the image, because the image also carries its own copy of the [omamobile] repository (/var/lib/omamobile/repo) with the package archive, so with voxtype-bin (28 MB download, 71 MB installed) the root image grows by about 370 MB unpacked and about 325 MB compressed, on top of ~2.8 GB zst. base.en is the daemon's default (Omarchy's and the shell's), and with the context cut it turns 12 s of speech into text in about 2.2 s on the 7 Pro. tiny.en (77,704,715 bytes, 74 MiB) is left out. Without it the live preview falls back to base.en, with a pass every 2.5 s instead of 1.2 s (a fallback omamobile-voice-preview already has). Fast (tiny.en) is still one tap in the Voice settings. Shipping tiny.en too would mean adding a second source line.

Phones installed before this keep their setup. The shell only optionally depends on omamobile-voice, so an update doesn't pull 141 MB.

Tests

tests/all, tests/packages omamobile-voice omamobile-shell, tests/build-rootfs, tools/pr-check origin/main: all pass. No root image was built and nothing ran on the phone.

Check on the phone after a fresh install

  1. pacman -Q voxtype-bin omamobile-voice; ls -l ~/.local/share/voxtype/models/ shows ggml-base.en.bin -> /usr/share/omamobile/voice/models/ggml-base.en.bin.
  2. systemctl --user is-enabled voxtype.service reads enabled, and is-active reads active after login; voxtype status reads idle.
  3. grep -A4 '^\[whisper\]' ~/.config/voxtype/config.toml shows model = "base.en" and context_window_optimization = true.
  4. Hold the pill in an app with the phone offline (Wi-Fi off), dictate a sentence: live words appear (base.en preview) and the final text is typed.
  5. After the first omarchy update, no "Install Dictation with Voxtype" toast appears (ls ~/.local/state/omarchy/done/voxtype-install-invitation). Install > AI > Dictation is hidden, Remove > Dictation is shown.
  6. /usr/bin/voxtype resolves to /usr/lib/voxtype/voxtype-cpu (voxtype-bin's install scriptlet ran under pacstrap).

🤖 Generated with Claude Code

https://claude.ai/code/session_01BeEcdVeSfyLCbijF8XQF1d

derekross and others added 5 commits November 2, 2026 17:07
A fresh install had no Voxtype, so the voice bar said "Voxtype isn't
running. Install it with Install > AI > Dictation", and that install
downloads the model, so voice needed a network and a setup step.

The new package fetches ggml-base.en.bin (147964211 bytes, 141 MiB), the
model Omarchy's installer downloads and the shell's default, at build time
from a pinned commit of whisper.cpp's model repository, checked against its
sha256, and installs it in /usr/share/omamobile/voice/models. Nothing binary
is committed. Voxtype looks for models only in the user's
~/.local/share/voxtype/models, so omamobile-setup-user links it there
(omamobile-shell 0.2.82). The weights are MIT (OpenAI's Whisper,
converted by whisper.cpp); the license ships with the package.

voxtype-bin (in [omarchy] for aarch64, 1.1.1-2, stable and edge) goes in
rootfs/packages.base: packages.omamobile takes [omamobile] packages only.
It stays an optional dependency, so Omarchy's Remove > Dictation
(pacman -Rns voxtype-bin) still works. wtype is already in Omarchy's list
and not skipped. tests/rootfs-smoke checks the image has voxtype, its user
unit and the model when the MANIFEST lists omamobile-voice.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BeEcdVeSfyLCbijF8XQF1d
…hipped model

What Omarchy's omarchy-voxtype-install does, without its download, once
per user and only when voxtype and omamobile-voice's base.en are installed:
copy Omarchy's Voxtype config (unless the user has one), link the shipped
model into ~/.local/share/voxtype/models (unless a file of that name is
there), enable the packaged voxtype.service user unit (it starts with the
graphical session; firstboot runs this with the user manager up but no
session, so no --now), and mark Omarchy's "Install Dictation" invitation
done, as its install would. The existing step then adds the context cut to
the new config. A user who removes Dictation doesn't get it back.

The voice bar's error now leads with starting voxtype.service; the
README gets a Voice input section, and omamobile-voice and voxtype-bin
are optional dependencies. test/setup-user.sh covers a new user, a second
login, a user with their own config and model, and no voxtype or no model.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BeEcdVeSfyLCbijF8XQF1d
The draft's voice line says Voxtype and base.en come in the image, so
dictation works from the first login with no setup and no download;
VOICE-PLAN's engine notes record how the image ships it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BeEcdVeSfyLCbijF8XQF1d
# Conflicts:
#	packages/omamobile-shell/version
… sizes, removal, licence

From the round 40 review of #355:

- The once-marker was taken before `systemctl --user enable voxtype.service`
  ran, so an enable that failed (firstboot carries on without a user
  manager) was never retried. The config, the link and the invitation
  stay once-only; the enable gets its own marker, set only when it
  succeeds, and is retried at each login while voxtype and the model are
  there. A user who removed Dictation (voxtype-bin gone) still gets
  nothing. A failed cp or ln no longer ends the whole setup run.
  test/setup-user.sh: a failed enable retried once and then left alone,
  and no retry once voxtype is removed.
- The README now counts the model twice (the image's own copy of the
  repository in /var/lib/omamobile/repo holds the package archive): about
  370 MB unpacked, 325 MB compressed with voxtype-bin. It also says how to
  remove dictation fully (pacman -R omamobile-voice after Remove >
  Dictation) and that the first build needs huggingface.co.
- LICENSE.whisper carries whisper.cpp's notice (the ggml authors) next to
  OpenAI's.
- The PKGBUILD notes the huggingface.co build dependency and makepkg's
  source cache; makepkg has no fallback for one source, so a copy on our
  server would replace the URL rather than back it up.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BeEcdVeSfyLCbijF8XQF1d
@derekross
derekross merged commit 87d7839 into main Nov 2, 2026
2 checks passed
@derekross
derekross deleted the voice-out-of-the-box branch November 2, 2026 22:41
derekross added a commit that referenced this pull request Nov 2, 2026
…, voice and weather)

Fingerprint unlock is now on by default with no consent step, so the
install warning no longer names it; the runner's data guards keep
stock's data safe. G5 is done in #352 (install-phone refuses slot B
without --reinstall), so it is marked done and dropped from M0 and §4.3.
Voice ships in the image (#355) and weather uses Open-Meteo, so the
install needs no extra download or key for either.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BeEcdVeSfyLCbijF8XQF1d
robertzas pushed a commit to robertzas/omarchy-mobile that referenced this pull request Nov 14, 2026
…out of the box (derekross#355)

A fresh image ships Voxtype (voxtype-bin from [omarchy]) and Whisper's base.en model (new omamobile-voice, fetched at build time from a pinned commit with a pinned sha256; MIT, licence notices shipped), so dictation works from the first login with no setup and no network. omamobile-setup-user does Omarchy's Install > AI > Dictation steps without the download, once per user: Omarchy's Voxtype config if the user has none, the shipped model linked into ~/.local/share/voxtype/models, voxtype.service enabled behind its own marker that is set only when the enable succeeds (retried at later logins), and Omarchy's install invitation marked done. A user's own config or model is never touched, and a user who removed voxtype is left alone. Existing phones don't change (optional dependencies only). The image grows by about 325 MB compressed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BeEcdVeSfyLCbijF8XQF1d
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant