Voica v0.6.1

Voica v0.6.1 — the result window stays on the screen you dictated on.

Fixed

In the "show result window" output mode, the window opened on the screen holding the mouse
pointer, not the screen where you were typing. Park the pointer on the second monitor, dictate
into a window on the first, and the text appeared over there.

Both the dictation bar and the result window now follow the focused window's screen, decided
once when recording starts. The default "insert into focused field" mode was never affected — the
text always went to the field you were typing in.

Nothing else changed; everything in 0.6.0
— the dictation bar with cancel, double-tap to start, multi-select in History — stays as it is.

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~80 MB	nothing — fully self-contained
Voica-fx.exe	~37 MB	.NET 8 Desktop Runtime installed once

Not code-signed yet, so SmartScreen may warn → More info → Run anyway
(why, and how to verify).

Full changelog: https://github.com/Inhum/voica-win/blob/main/CHANGELOG.md

Voica v0.6.0

Voica v0.6.0 — the dictation bar, safer starts, batch cleanup in History.

What's new

• A dictation bar at the bottom of the screen. A floating capsule shows a live level wave
  while you speak, with × to cancel (the recording is thrown away — nothing is sent or
  transcribed) and ✓ to stop and transcribe. While the text is being recognized the wave gives
  way to a spinner and "Transcribing…". It floats above other windows and never takes focus, so
  your text still lands in the field you were typing in. With several monitors it shows up on the
  one you're typing on, not always the main one. On by default —
  Settings → Dictation → Show the recording bar at the bottom of the screen.
• Cancelling a dictation is new. Until now, once a recording started the only way out was to
  send it.
• Exactly one indicator. With the bar on, the tray icon stays neutral in every state. Turn the
  bar off and the icon takes over again (pulsing while recording, a static accent while
  transcribing).
• Double tap to start (Toggle mode, on by default): recording begins only on two quick presses
  within 0.35 s, so a stray Right Alt no longer starts a dictation. Stopping is still a single
  press; push-to-talk and the tray's "Dictate" item are unaffected, and it can be turned off.
• Multi-select and batch delete in History — Ctrl/Shift/Ctrl+A, the Delete key, or the context
  menu; the selection and its audio go in one transaction.
• "Support the project" link in Settings → About, and a Russian guide to the vocabulary and AI
  term correction.

Also: Settings → Dictation is tidier — the long explanations now live in ⓘ tooltips — and in
History the Play button becomes Stop while audio is playing.

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~80 MB	nothing — fully self-contained
Voica-fx.exe	~37 MB	.NET 8 Desktop Runtime installed once

Not code-signed yet, so SmartScreen may warn → More info → Run anyway
(why, and how to verify).

Full changelog: https://github.com/Inhum/voica-win/blob/main/CHANGELOG.md

Voica v0.5.0

Voica v0.5.0 — model choice, self-healing AI correction, history export.

What's new

• Pick your Whisper model and language (Settings → Dictation): whisper-large-v3-turbo
  (faster, default) or whisper-large-v3 (more accurate), and auto‑detect / Russian / English.
  Forcing a language fixes short phrases that auto‑detect reads as the wrong language.
• AI term correction now heals itself. The chat model is no longer hardcoded: Voica reads the
  live model list for your key, picks the best available one, remembers it, and switches
  automatically if Groq retires or renames a model — no app update needed. You can also pin a
  specific model in the picker.
• Export your history to Markdown, CSV or JSON — the whole history with text and metadata
  (audio isn't included). CSV carries a UTF‑8 BOM so Excel shows Cyrillic correctly.
• About moved into Settings as a fifth tab, along with the updates block ("Check now" +
  check‑on‑launch).
• Better long dictations: local‑engine chunks now overlap, so a word or punctuation mark at a
  25‑second boundary is no longer split or lost.

Also: a transcription blocked for your Groq organization (HTTP 403) now tells you exactly which
model to allow.

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~80 MB	nothing — fully self-contained
Voica-fx.exe	~37 MB	.NET 8 Desktop Runtime installed once

Not code-signed yet, so SmartScreen may warn → More info → Run anyway
(why, and how to verify).

Full changelog: https://github.com/Inhum/voica-win/blob/main/CHANGELOG.md

Voica v0.4.0

Voica v0.4.0 — local offline engine.

What's new

• Local offline recognition — switch Settings → General → Recognition engine to Local and
  Voica transcribes entirely on your PC with Sber GigaAM v3 (Russian, with punctuation): no
  network, no API key, audio never leaves the machine. The model (~215 MB, int8 ONNX) downloads
  once with a progress bar and SHA-256 verification; delete it anytime in Settings → Data.
• Offline fallback — cloud selected but no internet? If the local model is installed,
  dictation just works locally (with a small notice).
• Tabbed Settings (General / Dictation / Vocabulary / Data) — compact, fits small displays.
• Model column in History, and the About window now covers both engines.

Notes: the vocabulary hint applies to the cloud engine only; AI term correction works with both
engines. ONNX conversion by istupakov/gigaam-v3-onnx (MIT).

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~75 MB	nothing — fully self-contained
Voica-fx.exe	~26 MB	.NET 8 Desktop Runtime installed once

Not code-signed yet, so SmartScreen may warn → More info → Run anyway.

Full changelog: https://github.com/Inhum/voica-win/blob/main/CHANGELOG.md

Voica v0.3.1

Voica v0.3.1 — hotfix: AI term correction works again.

Fixed

• AI term correction was silently broken. Groq removed the qwen/qwen3-32b model (requests
  returned 404), so corrections quietly fell back to the raw transcription (fail-open by design).
  Voica now uses llama-3.3-70b-versatile. The availability check in Settings also
  distinguishes a removed model (404 → “update the app”) from one blocked in your Groq org
  (403 → “allow it at console.groq.com → Settings → Limits”).

If you use AI term correction (Settings → Vocabulary), please update.

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~75 MB	nothing — fully self-contained
Voica-fx.exe	~26 MB	.NET 8 Desktop Runtime installed once

Not code-signed yet, so SmartScreen may warn → More info → Run anyway.

Full changelog: https://github.com/Inhum/voica-win/blob/main/CHANGELOG.md

Model: GigaAM v3 e2e CTC (int8 ONNX)

Model asset for Voica's local offline engine (spec §2.5). Not an app release — download Voica.exe from the latest release instead; the app fetches this model automatically when you enable the local engine in Settings.

• Model: Sber GigaAM v3 e2e CTC (Russian, with punctuation and text normalization), int8-quantized ONNX.
• Source: salute-developers/GigaAM (MIT), ONNX conversion by istupakov/gigaam-v3-onnx (MIT). Mirrored here unchanged for stable, versioned distribution.

SHA-256:

2e3fcb7a7b66030336fd10c2fcfb033bd1dc7e1bf238fe5cfd83b1d0cfc9d28e  v3_e2e_ctc.int8.onnx
e67eca3a311ad7c8813d36dff6b8eeba7ad3459fd811d6faea2a26535754a358  v3_e2e_ctc.yaml
142de7570b3de5b3035ce111a89c228e80e6085273731d944093ddf24fa539cd  v3_e2e_ctc_vocab.txt

Voica v0.3.0

Voica v0.3.0 — Dictate from the tray menu.

First release built and published entirely by CI from the tagged commit.

What's new

• Dictate — the first item in the tray menu starts/stops dictation with the mouse, no hotkey
  needed. Always works as a toggle: click to start, click to stop; ignored while transcribing.
• Menu layout now matches the cross-platform spec (Dictate · History · Settings · About ·
  Check for Updates · Quit), and a left click on the tray icon opens the menu too.

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~75 MB	nothing — fully self-contained
Voica-fx.exe	~26 MB	.NET 8 Desktop Runtime installed once

Not code-signed yet, so SmartScreen may warn → More info → Run anyway. On first run add your
Groq key (from https://console.groq.com) in Settings, then press Right Alt to dictate.

Full changelog: https://github.com/Inhum/voica-win/blob/main/CHANGELOG.md

Voica v0.2.0

Voica v0.2.0 — AI term correction, settings reset, and an important input fix.

What's new

• AI term correction (opt-in, off by default): after transcription, mangled vocabulary terms are fixed by a Groq chat model (qwen/qwen3-32b), matching grammatical case and context — e.g. «кубер стил» → kubectl. Fail-safe: if the model is unavailable, the original Whisper text is delivered. Settings shows whether the model is available for your Groq org (with a hint if it's blocked).
• Reset settings… button — back to defaults while keeping your API key, history, audio, and vocabulary.
• Live vocabulary counter N / 800 with a warning color when over budget (only the tail is sent to Whisper).

Fixed

• Random keyboard-layout switching after dictation. The injected Ctrl+V could combine with a physically held Shift/Alt into the system layout-switch chord (Ctrl+Shift / Alt+Shift). Insert now waits until your modifiers are released; the keyboard hook ignores injected input, injected keys carry real scan codes, and hotkey handling no longer runs inside the low-level hook callback.

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~75 MB	nothing — fully self-contained
Voica-fx.exe	~26 MB	.NET 8 Desktop Runtime installed once

Not code-signed yet, so SmartScreen may warn → More info → Run anyway. On first run add your Groq key (from https://console.groq.com) in Settings, then press Right Alt to dictate.

Full changelog: https://github.com/Inhum/voica-win/blob/main/CHANGELOG.md

Voica v0.1.0

Voica v0.1.0 — the first Windows release.

Dictation → punctuated text via Groq Whisper. Press a hotkey, speak, and Voica inserts clean, punctuated text into whatever field you're typing in. Bring your own Groq key.

Install

Two downloads to choose from (Windows 10 1809+ / x64):

File	Size	Needs
Voica.exe	~75 MB	nothing — fully self-contained
Voica-fx.exe	~26 MB	.NET 8 Desktop Runtime installed once

1. Download either file and run it — no installer.
2. The build is not code-signed yet, so SmartScreen may warn (“Windows protected your PC”) → More info → Run anyway.
3. On first run, add your Groq API key (from https://console.groq.com) in Settings, then press Right Alt to dictate.

Highlights

• Global hotkey dictation — PTT or Toggle; presets (Right/Left Alt, CapsLock, ScrollLock, Pause) and custom combinations (e.g. Ctrl+Shift+Space)
• Punctuation via Groq Whisper (whisper-large-v3-turbo), auto language detection
• Auto-insert into the focused field + clipboard fallback, or an editable result window
• SQLite history with audio playback and retention
• Groq key encrypted with Windows DPAPI; English/Russian UI
• No backend, no telemetry

See the README and the full changelog.