Skip to content

Changelog

Jump to bottom
rcspam edited this page May 12, 2026 · 11 revisions

🌐 Language: English | Français

Changelog

This page mirrors the GitHub release notes with additional context and cross-links into the wiki for features introduced in each version.

For the raw release artifacts (.deb, .rpm, .plasmoid, tarballs), always use the Releases page. For upcoming versions, see the Roadmap v1.4+ section at the bottom.

Table of Contents

v1.3.4 (stable)

Released 2026-05-12. Hotfix release: long files transcribe correctly on every machine (CPU and GPU), dictee-transcribe no longer mixes up your tabs, and several silent failures now show clear messages.

🔗 Release page · 11 assets.

What's new

  • Long-file transcription works for everyone. Before v1.3.4, only CUDA users could transcribe files longer than 5 min — CPU users hit a model bug with no recovery. Now any file longer than 3 minutes is automatically split into chunks and reassembled, on every machine. Tested up to 54 min keynotes.
  • Live dictation: a clear limit per ASR backend, no more silent crashes. Some backends can't safely handle very long recordings (Canary runs out of GPU memory around 3 min, Parakeet hits a model bug at 5:20 min). If you record beyond the safe limit (Canary 2:30, Parakeet 4:30), dictee now stops cleanly with a clear notification and saves your audio to /tmp/dictee-long-<timestamp>.wav so you can open it in dictee-transcribe instead. Whisper and Vosk stream natively and stay uncapped.
  • dictee-transcribe: tabs stay independent. Previously, starting a transcription in one tab could overwrite the editor, audio player, speaker labels or timeline markers of another tab you were reviewing. Each transcription now stays in its own tab, even while you browse the others.
  • Translation skips now visible. When a translation can't run (no target language picked, source language matches the target), dictee-transcribe shows a red status message in your language (fr / de / es / it / pt / uk) — previously it just silently did nothing.
  • Diarization works with multilingual files even when Canary is your live-dictation backend. Before, diarizing a file whose language was different from your dictation language produced gibberish silently. Diarization now uses Parakeet (multilingual auto-detect) for these files. Costs ~5–10 s extra to load the model, much better than wrong text.

Other fixes

  • Recovery after a stopped daemon. If dictee-transcribe stops the speech daemon to free GPU memory and then runs into an issue, push-to-talk dictation now resumes automatically instead of staying broken until you restart the service manually.
  • Recovery after a stuck push-to-talk state. If a previous transcription was killed (out of memory, crash, signal), pressing F9 again now resets properly instead of silently blocking all further dictation.
  • Default cheatsheet shortcut is Shift+F9 out of the box. New users immediately get the voice-commands cheatsheet hotkey — previously it shipped as "Disabled" and required a manual choice in the setup wizard.
  • "No GPU detected" warning no longer leaks into your text when running the CUDA build on a machine without an NVIDIA driver (VM, headless server). The transcription, translation and post-processing stay clean.
  • Whisper transcriptions of files longer than ~25 min no longer time out before the daemon finishes (the internal timeout was tuned for Parakeet, which is ~10× faster). The change is invisible for Parakeet / Canary users.

Installation & packaging

  • dictee-cuda and dictee-plasmoid now install together on Debian/Ubuntu. Previously dpkg refused both at the same time because they both claimed the same plasmoid file. The widget file now ships only with the cpu/cuda variants — dictee-plasmoid keeps the auto-discovery directory for system-wide installs (same structure as the rpm, which always worked).
  • Arch CUDA install no longer downloads 1.5 GB of NVIDIA libraries on a CPU-only machine. The other formats (deb / rpm / tarball) already skipped this since v1.3.2; Arch was the last one to catch up.
  • Five missing system packages declared (ffmpeg, python3-gi, pkexec, libkf6config-bin, wireplumber). The app uses them but the package descriptions didn't list them — fresh installs on some distros were missing pieces (audio file conversion, GNOME tray, privilege prompts, KDE shortcuts, mic volume control).
  • A new internal audit script blocks the build before release if a system package is missing from the metadata — same class of regression caught at build time from now on.

Upgrading

  • No configuration migration needed. v1.3.4 reads your v1.3.3 settings and models as-is.
  • No setting to change for long files. dictee-transcribe picks the chunked pipeline automatically.

v1.3.3

Released 2026-05-07. Patch release fixing silent breakage on fresh installs across distros — particularly Arch where .install hooks were missing usermod and python-evdev was an optdepends. Also closes #5 (NVIDIA false positive on AMD) and #6 (PP-translate toggle persistence + LibreTranslate language combo).

🔗 Release page · 9 assets.

What's fixed

Cross-distro packaging consistency

  • Arch .install hooks now add input / docker groups at install time (postinst .deb / %post .rpm already did). Without this, fresh Arch installs silently broke PTT and LibreTranslate even after a successful wizard run.
  • python-evdev promoted to hard depends on Arch (was optdepends → pacman never installed it → dictee-ptt fell back to broken raw mode that leaked the PTT key to focused apps).
  • build-rpm.sh no longer ships a stale dictee.plasmoid — plasmoid generation now factored into dict_prepare_pkg_dir (build-common.sh). Every builder regenerates from source on every run; no more skew between deb / rpm / tarball.
  • udev rule shipped in mode 0660 directly — was 0620 (write-only group) which breaks python-evdev O_RDWR open. Drop-in 81-dictee-uinput.rules added on Arch where the AUR dotool package's 0620 rule kept being restored on its upgrades.
  • Postprocess venv (text2num) created in tarball install too — was missing.
  • install.sh prompts to install Docker on Arch when LibreTranslate is selected.

Plasmoid

  • docker inspect wrapped in sg docker so the LT indicator stays accurate when plasmashell's group set hasn't refreshed yet (typical right after first install).
  • LibreTranslate language combo populated even when DICTEE_LIBRETRANSLATE_LANGS is unset — falls back to querying the running container's /languages API at localhost:5000.

Setup wizard

  • NVIDIA detection strict — uses nvidia-smi --query-gpu directly (with FileNotFoundError caught), no longer fooled by /sys/class/drm/card*/device/mem_info_vram_total returning AMD VRAM > 0. Closes #5.
  • Toggle "Enable PP for translation" persists across Apply / reload — was reading the wrong variable at startup. Closes #6 (part 1).
  • _docker_use_sg auto-detected at startup, not only after Setup Docker click.

dictee binary

  • dotool wrapped in sg input when invoked from a stale-group parent shell (plasmoid Dictate button, tray click, manual CLI). dictee-ptt.service already had the wrapper; standalone invocations didn't.

Cleanup

  • dictee-debug orphan package on Arch — options=('!debug') in PKGBUILDs prevents the empty sibling.
  • restart_daemons dead-code warning during cargo build — function cfg aligned with caller cfg.

Notes for upgraders

  • Plasmoid widget on KDE: the running widget instance keeps the old QML in memory after a package upgrade. Run kquitapp6 plasmashell && kstart plasmashell (or log out / back in) for the LT-status fix to take effect on already-pinned widgets.
  • Group runtime: any session opened before the install will keep its old supplementary groups until next login. dictee handles this internally (sg wrappers), but the change won't show up in groups runtime until logout/login.

v1.3.2

Released 2026-05-05. Patch release fixing 8 user-facing bugs accumulated since v1.3.1, all surfaced during a single VM-test session on Fedora and Ubuntu.

🔗 Release page · 9 assets (same shape as v1.3.1).

What's fixed

Reinstall / install

  • uninstall.sh now prompts before deleting personal configs (new --keep-configs flag for symmetry with --keep-models). Also actually removes ~/.config/dictee.conf — the file used to slip through --purge because only the directory ~/.config/dictee/ was deleted, leaving the .conf orphaned. Result: the next install correctly skipped the wizard because DICTEE_SETUP_DONE=true survived.
  • install.sh auto-restarts dictee-ptt on reinstall when dictee.conf is preserved, so the PTT key works immediately. Robust to curl | sudo bash (auto-detects SUDO_USER).
  • dictee-reset now restarts dictee-ptt. The "Reset" button in the plasmoid / tray is finally a true reset, not just an ASR daemon kick.

Plasmoid

  • Dictation button focus regression fixed (revert of the auto-pin introduced in v1.3.0). The auto-pin assumed animation-speech-ctl would steal back the window activation — verified in source that it does not (set_accept_focus(False) + KeyboardMode.NONE). Text typed by dotool now lands in the previously focused editor again.

Cheatsheet shortcut — full refactor

  • The Mod+<PTT key> modes are now handled directly by dictee-ptt (same path as Alt+<PTT key> for translate). No more dependency on kglobalaccel or gsettings — works on KDE and GNOME identically.
  • Default shortcut changed from the unsafe Same key + Ctrl + Alt (collided with Ctrl+Alt+Fn TTY switching on Linux) to Disabled — users opt in explicitly.
  • New option: Super + same key (replaces the unsafe Ctrl+Alt+key).
  • First-run banner is now i18n in 7 languages (was hardcoded French) and shows the actual configured shortcut (was hardcoded Ctrl+Alt+H).

Tray

  • dictee-setup spawn from the tray menu / first-run dialog is now detached from the systemd transient scope (systemd-run --user --scope --collect) so it survives the tray's sys.exit() on GNOME (app-gnome-*.scope with KillMode=control-group was killing the child immediately).

Packaging

  • CUDA postinst skips ~1.5 GB pip download on CPU-only machines. The dictee-cuda package now detects whether an NVIDIA GPU is present (/proc/driver/nvidia/ or /dev/nvidia0). If not, it prints a short note and the runtime fallback (introduced in v1.3.1) takes over. Install on a CPU-only machine is now near-instant.
  • python3-evdev is now a hard dependency (Depends: / Requires:). On Fedora it was a Recommends, installed after the package's %post ran — so when the post-install script restarted dictee-ptt, evdev wasn't yet available and the daemon fell back to raw /dev/input mode where the PTT key leaked to focused applications. With evdev enforced as a hard dep, the service starts in evdev (grab + uinput) mode out of the box.

Upgrading

# Debian / Ubuntu
sudo apt install ./dictee-{cpu,cuda}_1.3.2_amd64.deb

# Fedora / openSUSE
sudo dnf install ./dictee-{cpu,cuda}-1.3.2-1.x86_64.rpm

# Arch (AUR)
yay -S dictee   # or dictee-cuda

In-place upgrade. No config migration.

v1.3.1

(superseded by v1.3.2 — kept for history)

Released 2026-05-04. Patch release fixing four user-visible bugs from v1.3.0, all hit on the first day on real-world setups (fresh Fedora VM, headless boxes without a working NVIDIA driver, two transcriptions kicked off in quick succession).

🔗 Release page · 9 assets (same shape as v1.3.0).

What's fixed

  • CUDA → CPU fallback at runtime. The CUDA .deb / .rpm now probes /proc/driver/nvidia/gpus/ at startup and silently falls back to CPU on any host without a usable NVIDIA driver — virtio VMs, headless containers, machines where the driver was removed. Before, the daemon crashed in a restart loop with cudaSetDevice failed: driver insufficient. New DICTEE_FORCE_CPU=1 environment variable forces CPU even when a GPU is present.
  • uinput auto-load on Fedora / RHEL / openSUSE. Adds /etc/modules-load.d/dictee-uinput.conf and runs modprobe uinput in the post-install hook so dictee-ptt and dotool can open /dev/uinput. Without this, the F8 push-to-talk shortcut was dead on a fresh Fedora install — the kernel module is not auto-loaded by default like on Arch / Ubuntu.
  • Setup wizard "ASR service" check is now strict. The final wizard page polls systemctl is-active for up to 6 s and checks the daemon's Unix socket exists; it can no longer report "Everything is ready!" while the service is stuck in activating (auto-restart).
  • LLM analysis tab inherits the right #N counter. When you click LLM analysis from a diarized tab after starting a second plain transcription, the result tab now reads #1 LLM: Synthèse (matching the source), not #2.
  • Transcribe button + diarize toggle + auto-translate stay disabled for the entire run, including the brief phase 1 → phase 2 transition of two-phase diarization, so a second job can no longer be kicked off mid-flight.

Upgrading

# Debian / Ubuntu
sudo apt install ./dictee-{cpu,cuda}_1.3.1_amd64.deb

# Fedora / openSUSE
sudo dnf install ./dictee-{cpu,cuda}-1.3.1-1.x86_64.rpm

# Arch (AUR)
yay -S dictee   # or dictee-cuda

The upgrade is in-place; no config migration. If you previously installed v1.3.0 and hit the dead-F8 issue on Fedora, the v1.3.1 post-install loads uinput immediately — no reboot required.

v1.3.0

Released 2026-05-03. Cumulative since v1.2.0 — 629 commits, 9 months of development.

🔗 Release page · 9 assets (CPU + CUDA .deb / .rpm, plasmoid .deb / .rpm / .plasmoid, universal .tar.gz, source tarball).

Highlights (full notes on the release page):

  • Canary-1B v2 ASR backend (NVIDIA AED, 7 languages, 12 native translation pairs ↔ EN)
  • Pipeline chunked for long-file diarization (no more 10-15 min cap on 8 GB VRAM)
  • LLM analysis pipeline on diarized transcripts: 6 built-in profiles × 14 provider presets (Ollama / OpenAI / Claude / Gemini / Mistral / DeepSeek / Perplexity / Groq / Cerebras / OpenRouter / etc.)
  • dictee-transcribe — new dedicated window for file transcription (audio player synced with timeline, multi-tab exports, LLM/Translate per-tab)
  • 5 translation backends (Canary, LibreTranslate, Ollama, Google, Bing) + refonte of the per-file Translate pad
  • 6 i18n languages complete (FR / DE / ES / IT / PT / UK)
  • Portable CUDA libs via pip venv at postinst (no system NVIDIA repo required)

Pre-release history

The v1.3 cycle went through 5 betas + 3 release candidates between 2026-04-15 and 2026-05-02. Each is preserved on GitHub for traceability:

  • v1.3.0-rc3 (2026-05-02) — LLM diarization analysis + chunked long-files + UI polish
  • v1.3.0-rc2
  • v1.3.0-rc1 (2026-04-20)
  • v1.3.0-beta1 → beta5 (2026-04-15 → 2026-04-19)

v1.3.0-rc1 — original notes

Released 2026-04-20. First release candidate. Contains 3 last-minute fixes found during the VM test round.

🔗 Release page · 8 assets (CPU + CUDA .deb/.rpm + plasmoid + tarballs).

Released 2026-04-20. First release candidate. Contains 3 last-minute fixes found during the VM test round.

🔗 Release page · 8 assets (CPU + CUDA .deb/.rpm + plasmoid + tarballs).

Fixes on top of beta5:

  • 72ec24f — Theme-aware banner on fresh KDE Fedora session (kreadconfig6 ColorScheme returns empty → fall back to Qt palette)
  • 44649bb — Wizard translation sync + 130 target languages in dropdown + combo widths aligned
  • 130ab95 — Plasmoid popup: language combo compact, dropdown popup wide enough for 5-char codes (pt-BR, zh-CN), popup closes on window deactivate

v1.3.0-beta5

Released 2026-04-19. Pre-RC candidate.

Features (beta3 → beta5):

  • f9cff13 — Plasmoid cross-process toggles: LLM / Short / Meeting switches synchronized across tray, plasmoid, and dictee.conf
  • d11c9b3 — Batch diarize binary compiled + long-file warning in dictee-transcribe when >10 min + diarization + CUDA
  • b179143 — French hour formatting ("14 h 30") + surrounded equal signs ("deux égale deux" → "2 = 2")
  • 3ccadf4, 50366af — Rules tab UI fixes (combo widths, scrollbar, split "Ajouter une règle" row)

CUDA build improvements:

  • bbca0f9 — Bundle CUDA libs via pip venv instead of OS packages — no more system-wide cuda-cudart-12 dependency
  • d60ba9f — Bundle libonnxruntime_providers_cuda.so inside the package

v1.3.0-betas

Earlier beta series (beta1 → beta4, 2026-04-15 → 2026-04-17):

Short-text keepcaps exceptions (7 languages):

  • New dialog for per-language acronym preservation (SDK, API, GPT-4…)
  • Extended match mode toggle (beyond short_text)
  • ~50 default exceptions per language

Multi-user safe:

  • UID suffix on /dev/shm state files → no more cross-user conflicts
  • keepcaps.conf now in ~/.config/dictee/ (user-specific)

LibreTranslate improvements:

  • Purge unused models button in setup
  • Volume size refreshed asynchronously
  • Docker container management from wizard

Post-processing:

  • Continuation + translate whitespace preservation
  • Version number dictation ("un point trois point zéro" → "1.3.0")
  • Decimal commas for French ("un virgule cinq" → "1,5")

Test coverage:

  • 682 post-processing tests (was ~400 in v1.2)
  • 148 pipeline tests (new)
  • CI job with isolated XDG_CONFIG_HOME

Packaging:

  • 8 → 10 release assets (added RPM variants for Fedora/openSUSE)
  • Arch PKGBUILD updated for v1.3
  • Installer supports /releases?per_page=1 to fetch prereleases

v1.2.0

Released 2025-12. Major backend overhaul.

Features:

  • 4 ASR backends unified (Parakeet, Canary, Whisper, Vosk)
  • Canary-1B v2 with built-in translation (25 European languages ↔ EN, 48 pairs)
  • Post-processing pipeline formalized into 12 steps
  • Quick backend switching via dictee-switch-backend
  • First-run wizard (was optional before)
  • dictee-test-rules CLI tool
  • Sidebar UI with dual pipeline diagrams

Breaking changes:

  • onnx-asr Python dependency removed — Canary now runs via native Rust ONNX wrapper (same code path as Parakeet)
  • Old dictee.conf format migrated automatically on first run

v1.1.0

Released 2025-09. Translation backends added.

Features:

  • LibreTranslate integration
  • Ollama integration
  • Google + Bing via translate-shell
  • Target language selector in plasmoid
  • DICTEE_LANG_TARGET environment variable

v1.0.0

Released 2025-05. Initial public release.

Features:

  • Parakeet-TDT 0.6B v3 ASR
  • KDE Plasma 6 plasmoid
  • PyQt6 tray icon
  • Setup wizard
  • Basic post-processing (rules, dictionary)
  • Keyboard shortcuts (KDE + GNOME)
  • .deb packaging only

Roadmap v1.4+

Tentative scope for v1.4 (ordered by value/complexity ratio):

  • Whisper translate — multi-target translation via task="translate" (EN only, offline alternative to Canary's 4 language pair limit)
  • Hotword boosting — bias ASR decoding toward custom names / terms without retraining (shallow fusion on TDT logits, Parakeet only)
  • Moonshine CPU backend — ultra-fast English-only ASR for low-power setups
  • CLI speech-to-text — pipe audio, get text (headless use, server dictation)
  • VAD (Voice Activity Detection) — hands-free dictation without push-to-talk
  • Streaming transcription with live text display — merge Vosk streaming mode with plasmoid live text field
  • Built-in overlay — replace external animation-speech with in-process fullscreen overlay
  • AppImage / Flatpak packaging — portable install without system packages
  • COSMIC applet + GNOME Shell extension — feature parity with KDE plasmoid on other desktops
  • Passthrough mode (video calls) — pw-loopback mix mic + app audio → diarization of the meeting
  • Silence threshold per source — adaptive RMS threshold based on source type (mic / monitor / app)

Tracked in the project memory and in GitHub Projects. Contributions welcome on any of these.

Next steps

📖 dictee Wiki

🇬🇧 Home · 🇫🇷 Accueil

Getting started / Premiers pas

Speech recognition / ASR

Translation / Traduction

Post-processing / Post-traitement

CLI

Reference / Référence

🏠 Repo · 📦 Releases · 🐛 Issues

Clone this wiki locally