VoxType

KO | EN

gitlite — search

#python #react #ai #typescript #llm #go #golang #android #rust #docker #nodejs #windows #kubernetes #vue

(★ 10)

Voice typing for Windows. local & cloud ASR (Seed-ASR 2.0 )豆包语音输入法,FireRedASR2

VoxType 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">VoxType</h1>

Voice typing for Windows. Press, speak, paste. 
 Windows voice typing tool — hold to speak, release to paste, local & cloud ASR

Current version: <code>v0.9.9</code> &nbsp;|&nbsp; 🇨🇳 <a href="doc/README_zh.md">中文版</a>

https://github.com/user-attachments/assets/36243dc2-cfc8-41fb-b0cf-6e558f02cd5e

---

## Highlights

- **Press and Speak** — Default CapsLock long-press to record, release to auto-paste to current window; short press toggles Caps Lock normally
- **Local-first** — C++ directly calls sherpa-onnx without Python; optional cloud ASR / LLM providers can be enabled explicitly
- **Real-time HUD** — Bottom floating capsule window during recording, 5 volume bars responding to sound
- **Dual VAD Options** — Silero VAD (lightweight) / FireRed VAD (high precision F1 97.57), intelligently skips silence
- **LLM Correction (Optional)** — Supports DeepSeek / OpenRouter / SiliconFlow and other providers, one-click configuration
- **Cloud ASR (Optional)** — Supports Volcano Engine (Doubao), Baidu Cloud, Qwen ASR (`qwen3-asr-flash-realtime`), Xiaomi MiMo ASR (`mimo-v2.5-asr`), and experimental Doubao IME ASR as alternative backends, with optional fallback ASR

## Quick Start

### 1. Download Models

```powershell
.\download_models.ps1
```

Interactive menu for model selection, auto-downloads and extracts to `models/` directory (~3GB disk space).

> Silero VAD and FireRed VAD are built-in, no download needed.

### 2. Build

```powershell
.\build.bat
```

Requires Visual Studio 2022 (C++ desktop development workload).

### 3. Run

```powershell
.\build\run\x64-release\VoxType.exe
```

Right-click the tray icon to open Settings, hold the hotkey to start recording, release to recognize and paste.

### 4. Package for distribution

```powershell
.\build.bat --package
```

This creates verified assets under `build\packages`: a Portable `.7z` and an
x64 per-machine MSI. Both originate from the same canonical runtime payload;
the packager re-extracts and hashes each result before publishing it.

- The MSI defaults to `Program Files\VoxType`. Choose **Advanced...** during
  setup to select another directory. That selection is retained by later MSI
  upgrades.
- Installed builds store `config.json`, downloaded ASR/punctuation models, and
  logs in `%LOCALAPPDATA%\VoxType`, so upgrades do not write into Program
  Files or remove user data.
- Portable builds include `portable.flag` and keep those files beside the
  extracted executable instead.
- Packages without a configured release certificate are deliberately named
  `-unsigned`; `--require-signing` refuses to publish without the signing
  environment variables documented in `packaging/windows/UPGRADE_CONTRACT.md`.

---

<details open>
<summary>Settings Guide</summary>

**General tab**
- `Start VoxType when I sign in to Windows` registers the current user's
  Windows Run entry. It is off by default and can be safely enabled for either
  the MSI or Portable build; saving after moving a Portable folder corrects its
  stored executable path.

**Recognition tab**
- `ASR Backend` — Select between `Local (sherpa-onnx)`, `Volcano Engine`, `Baidu Cloud`, `Qwen ASR`, `MiMo ASR`, and `Doubao IME (Free)`
- `Fallback` — Optional backup ASR backend: `Local`, `Baidu Cloud`, `Qwen ASR`, `MiMo ASR`, or `Doubao IME (Free)`. If the primary backend ends in an operational failure such as timeout, network error, auth/config error, or model load error, VoxType retries the same raw PCM with the fallback backend before showing the final result. `Too short` and `No speech detected` do not trigger fallback.
- `ASR model` — Speech recognition model (only for Local backend):
  - `FireRedASR2 CTC` — Fast, suitable for daily input
  - `FireRedASR2 AED` — Better quality, more accurate for long sentences
  - `SenseVoiceSmall` — Lightweight model, suitable for low-resource machines
- `Model folder` — Model file storage directory
- `Threads` — Inference thread count, `auto` uses CPU core count (max 8)
- `Enable VAD` — Enable voice activity detection, checks for speech before recording, skips ASR when no voice detected to save time
- `VAD model` — Voice activity detection model (requires Enable VAD):
  - `Silero VAD` — Lightweight and fast, accuracy F1 95.95
  - `FireRed VAD` — High precision (F1 97.57, false alarm rate 2.69%), model only 2.2MB
- `Punctuation` — Post-recognition processing:
  - `Disabled` — No processing, outputs ASR raw text
  - `Auto punctuate` — Local CT-Transformer auto-punctuation (no network required)
  - `Auto punctuate + LLM` — Local punctuation + cloud LLM correction (requires configuring provider and API Key in LLM tab, see LLM Correction section below)
- `Hold hotkey` — Click the input box then press the hotkey to record
  - `Esc` cancels this recording, `Backspace/Delete` clears the hotkey
  - Default CapsLock: Short press toggles Caps Lock, long press 300ms triggers voice input

**LLM tab**
- `Provider` — Provider dropdown, built-in DeepSeek / OpenRouter / SiliconFlow presets, auto-fills fields below on selection
- `API Base URL` — Provider API address (auto-filled by presets)
- `API Key` — API key, stored encrypted with DPAPI in local config
- `Model` — Model name (e.g., `deepseek-v4-flash`)
- `Test Connection` — Test if API connection is working, results shown in Status area below
- `Extra Params` — Additional JSON parameters, merged into request body (format: `"key":"value","key2":"value2"`). Preset providers auto-inject thinking mode disabled parameters, users can add custom fields here
- `[+]` / `[−]` — Add/delete custom providers (presets cannot be deleted)

**LLM Prompt tab**
- `System Prompt` — Multi-line editor, customize LLM correction system prompt (leave empty to use built-in default)
- `Basic Fix` / `Deep Fix` — Preset buttons, one-click fill for different correction intensity System Prompts

**Cloud ASR tab**
- `Provider` — Select between `Volcano Engine (Doubao)`, `Baidu Cloud`, `Qwen ASR (DashScope)`, `MiMo ASR (Xiaomi)`, and `Doubao IME (Free)`, controls below update dynamically
- **Baidu Cloud**: `API Key` / `Secret Key` (DPAPI encrypted) + `Language Model` (Mandarin/English/Cantonese/Sichuanese) + `Test Connection`
- **Volcano Engine (Doubao)**: `API Key` (DPAPI encrypted) + `ASR Mode` + `Model Version` + `Language` + `Test Connection`
  - ASR Mode: `bigmodel_nostream` (recommended, highest accuracy) / `bigmodel_async` (best latency) / `bigmodel` (real-time partial)
  - Model Version: `Seed-ASR 2.0 (duration)` / `Seed-ASR 2.0 (concurrent)` / `BigASR 1.0 (duration)` / `BigASR 1.0 (concurrent)`
  - Hotwords ID/Name, Correct ID/Name — Reference hotword and correction tables from the self-learning platform
  - Use history as context — Sends recent recognition results as dialog context for improved accuracy
  - Read input field context — Reads current input field text as ASR context (UIA/MSAA/WM_GETTEXT layered fallback, input field priority, history fallback)
- **Qwen ASR (DashScope)**: `API Key` (DPAPI encrypted) + `Base URL` + `Model` + `Language` + `Chunk ms` + `Test Connection`
  - Default model: `qwen3-asr-flash-realtime`
  - Turn detection is fixed to Manual for push-to-talk usage; Server VAD settings are intentionally hidden
- **MiMo ASR (Xiaomi)**: `API Key` (DPAPI encrypted) + `Base URL` + `Model` + `Language` + `Test Connection`
  - Default Base URL: `https://token-plan-ams.xiaomimimo.com/v1`
  - Default model: `mimo-v2.5-asr`; audio is uploaded as WAV via `/chat/completions`
- **Doubao IME (Free)**: no API key field. The experimental provider registers a Doubao IME-style device, stores device credentials with DPAPI-encrypted token, encodes PCM to Opus, and uses the unofficial `frontier-audio-ime-ws.doubao.com` WebSocket protocol. Availability and terms are not guaranteed.
  - Doubao IME bypasses local VAD and relies on the IME service's own segmentation. Partial HUD updates and final text are accumulated across cloud-side segments during one hotkey hold, so long recordings are pasted as one combined result after release. When selected as Fallback, Doubao IME replays the same raw PCM through a recorded request and writes refreshed credentials back to the saved config.
  - Streaming partial HUD for Qwen, Volcano Engine, and Doubao IME is display-only clear-page: it shows live text within three body lines, then clears previous HUD text and restarts from the current last sentence; the new page keeps accumulating until it exceeds three body lines again, while the final paste text stays complete.
  - Diagnostic probe: run `.\tools\doubao_ime_probe.bat` to compile a small console probe that reuses saved Doubao IME credentials when available, performs a live protocol check, and when the bundled sample wav exists, performs a real speech recognition check. Add `--streaming` to send WAV frames with a live drain thread and validate partial/final streaming behavior; add `--fresh` to force temporary re-registration.
- Cloud ASR backends handle recognition remotely; when VAD is enabled, Qwen and Volcano Engine use local streaming VAD trim, batch cloud backends use batch VAD trim before upload, and Doubao IME uploads raw PCM/Opus without local VAD. Local punctuation models are still bypassed for cloud backends

</details>

<details>
<summary>Supported Models</summary>

| Model | Type | Features |
|---|---|---|
| FireRedASR2 CTC int8 | ASR | Fast, suitable for daily input |
| FireRedASR2 AED int8 | ASR | Better quality, more accurate for long sentences |
| SenseVoiceSmall int8 | ASR | Lightweight, suitable for low-resource machines |
| CT-Transformer Punctuation int8 | Post-processing | Chinese/English punctuation auto-completion |

Model download links can be found in the `download_models.ps1` script, or manually from [sherpa-onnx releases](https://github.com/k2-fsa/sherpa-onnx/releases).

</details>

<details>
<summary>LLM Correction</summary>

Optional cloud LLM text correction added since v0.2.0. Disabled by default, requires manual enable:

1. Settings → LLM tab → Select provider, enter API Key
2. Settings → Recognition → Punctuation set to `Auto punctuate + LLM`

Features:
- Preset providers auto-inject thinking mode disabled parameters
- Extra Params supports custom JSON fragments merged into request body
- API Key encrypted with DPAPI storage
- Backward compatible with old config format

</details>

<details>
<summary>Project Structure</summary>

```
src/
  app/              — Entry point, globals, Win32 resources
  asr/              — ASR clients, batch/streaming sessions, ASR result dispatch helpers
  audio/            — Local ASR engine, audio capture, WASAPI, FireRed VAD, streaming VAD trim
  ui/               — HUD, hotkey handling, Settings window
  core/             — Shared utilities, LLM correction, input context reading
dll/                — Runtime DLLs (sherpa-onnx, onnxruntime, etc.)
third_party/        — Headers and import libraries
models/             — Model files (not committed to git)
build.bat           — Visual Studio 2022 build script
download_models.ps1 — Model download script
ARCHITECTURE.md     — Detailed architecture description
AGENTS.md           — Development collaborator notes
CHANGELOG.md        — Version change log
```

</details>

<details>
<summary>Known Limitations</summary>

- Local, Baidu, and MiMo results are finalized after recording; Qwen, Volcano Engine, and Doubao IME can show partial HUD during recording, with final text pasted after release
- Text injection primarily via clipboard + Ctrl+V, admin privilege windows may block
- Model files are large (~3GB), first load takes a few seconds

</details>

<details>
<summary>Changelog</summary>

See [CHANGELOG.md](CHANGELOG.md)

**Recent Updates:**
- **v0.9.9** — Restored the shared active-recording HUD animation for Local and other batch ASR backends, so their input-level bar lights up and moves again without changing the v0.9.3 capture-first cloud startup order
- **v0.9.8** — Canonical CMake/Ninja release pipeline with verified Portable and MSI packages, MSI upgrade/folder-selection support, portable data isolation, and Start with Windows setting
- **v0.9.7** — Privacy-safe, bounded ASR diagnostics with structured primary/fallback lifecycle events; streaming primary failures reported before release are now deferred until the full PCM is available, so configured fallback is no longer bypassed
- **v0.9.6** — Doubao IME can now be selected as a fallback ASR target through a recorded-PCM helper, including credential refresh/writeback, cloud timing, Settings support, and raw-PCM replay without local VAD trim
- **v0.9.5** — Recognition-tab fallback ASR backend, serial primary-to-fallback orchestration for batch and streaming failures, shorter fallback-enabled streaming final wait, result metadata for debug/LLM, and local-model preload when Local is configured as fallback
- **v0.9.4** — Experimental Doubao IME (`doubao_ime`) streaming cloud ASR backend, vendored static Opus 1.6.1, credential bootstrap/reset UI, protocol/WAV/streaming diagnostic probe, Doubao long-recording aggregation fixes, and shared clear-page streaming partial HUD for Qwen/Volcengine/Doubao IME
- **v0.9.3** — Volcengine rapid recording head-audio loss fix, streaming VAD double-processing fix, connection reuse/timeout tuning, active request fast cancel, and Qwen/Volcengine startup/stop cleanup
- **v0.9.2** — HUD DPI-aware rendering, Volcengine/Qwen WebSocket double-close and data-race fixes, Qwen activeClient UAF fix, retry abort checks
- **v0.9.1** — Cloud ASR architecture stabilization, Xiaomi MiMo ASR (`mimo-v2.5-asr`) backend, Qwen/Volcengine streaming sessions, shared streaming/batch VAD trim core, Baidu same-PCM retry and token refresh retry, source tree reorganization
- **v0.9.0** — Qwen ASR (`qwen3-asr-flash-realtime`) backend, true streaming send with partial HUD, Manual turn detection by default, Qwen watchdog/replay retry, shared ASR session/dispatcher/result architecture
- **v0.8.7** — Volcengine retry recognition: connection-loss audio buffering + full-PCM retry, adaptive finalize timeout, unified ASR error classification, no-text close handling
- **v0.8.6** — Volcengine connection reuse optimization (fast consecutive recording latency reduced from ~1.8s to ~0.4s), 3s expiry detection + hSession connection pool cleanup, time-gated internal retry, progressive external retry
- **v0.8.5** — Input field context (UIA/MSAA/WM_GETTEXT, read input text as ASR context), context logic refactor (input field priority, history fallback, remove window title), remove accelerate score, settings UI reorganize
- **v0.8.4** — Fix Volcengine no-speech drainThread.join() blocking 17+ seconds (close WebSocket before join), watchdog thread handle leak crash, SendMessage(WM_PASTE) UI thread blocking
- **v0.8.2** — Fix Volcengine nostream/async result loss, long recording truncation, WinHttpCloseHandle deadlock, logic deadlock, short audio false timeout, HUD timer race
- **v0.8.0.1** — Fix Volcengine nostream no-speech hang (WinHttpWebSocketReceive 1ms timeout not working)
- **v0.8.0** — Streaming VAD (real-time speech detection during recording, skip silence), Volcengine nostream/async long recording hang fix, HUD speech detection visual feedback, watchdog auto-renew during recording, code quality fixes
- **v0.7.5** — Volcengine WebSocket hang fix (forceAbort), connection prewarm, connection expiry rebuild, WinHTTP proxy/no-proxy fix, timeout tuning, reduced logging verbosity
- **v0.7.4** — WeChat Chinese IME paste fix (WM_CHAR bypass IME), IMM32 input method state switching, Unicode SendInput fallback, Force Unicode Input menu
- **v0.7.3** — WASAPI Shared Mode capture (48kHz→16kHz resample + waveIn fallback), Debug Mode console with per-stage timing, Baidu ASR response parsing fixes

- **v0.7.2** — Thread safety fixes (volcengine thread join, Baidu token mutex, VolcDebugLog mutex), SSL cert verification restored, `g_volcAudioCs` leak fix, utility functions deduplicated to `utils.h`, model downloader async, `AsrEngine::lock` encapsulation
- **v0.7.1** — `bigmodel_nostream` speed optimization (skip intermediate receives), WinHTTP connection reuse, `ExtractJsonStr` escape fix, thread safety fix
- **v0.7.0** — Volcengine ASR full API parameter support, hotwords/correction tables, dialog context, `corpus` merge fix, `context` format fix
- **v0.6.2** — Settings UI style unification: `UiStyle` namespace, consistent row spacing across all tabs
- **v0.6.1** — ASR model preload on startup, DELAYLOAD for DLLs, idle memory reduced to ~12 MB
- **v0.6.0** — Source code refactored from single-file to multi-module architecture
- **v0.5.0** — Cloud ASR UI overhaul, async mode fix, Shortcut merged into Recognition
- **v0.4.0** — Volcengine (豆包) streaming ASR integration via WebSocket
- **v0.3.0** — Baidu Cloud ASR integration
- **v0.2.1** — Multi-provider preset system, LLM Prompt independent Tab, Extra Params
- **v0.2.0** — FireRedVAD integration, cloud LLM correction module
- **v0.1.4** — C++ direct sherpa-onnx calls, removed Python dependency

</details>

---

## Acknowledgments

- [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) — ASR / VAD / Punctuation inference engine
- [FireRedASR](https://github.com/FireRedTeam/FireRedASR) — High precision Chinese ASR model
- [FireRed VAD](https://github.com/FireRedTeam/FireRedAudio) — High precision VAD model
- [Silero VAD](https://github.com/snakers4/silero-vad) — Lightweight VAD model
- [onnxruntime](https://github.com/microsoft/onnxruntime) — ONNX inference engine
- [Baidu Intelligent Cloud](https://ai.baidu.com/tech/speech/asr) — Baidu Cloud ASR API
- [Volcengine Speech](https://www.volcengine.com/docs/6561/1354869) — Volcengine (豆包) streaming ASR API
- [Alibaba Cloud Model Studio Qwen ASR](https://help.aliyun.com/zh/model-studio/qwen-asr-realtime-interaction-process) — Qwen ASR realtime API
- [Xiaomi MiMo](https://platform.xiaomimimo.com/docs/zh-CN/usage-guide/Speech-Recognition) — MiMo ASR API
- [push-2-talk](https://github.com/yyyzl/push-2-talk) — MIT-licensed Doubao IME protocol research reference
- [Opus](https://opus-codec.org/) — Audio codec used by the Doubao IME experimental provider

VoxType

VoxType 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

VoxType

VoxType 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README