DEV Community· sunerpy·· 4 小時前AI 評分62
sunerpy 開發可全離線運行的開源 Typeless 替代工具 Voltip
I built an open-source alternative to Typeless that can also run fully offline
AI 導讀
sunerpy 開發開源語音輸入工具 Voltip,支援在本機辨識語音,亦可全程離線運行。作者以 Qwen3-ASR 0.6B 處理 16 s 中文,L40S Vulkan 用時 0.16 s,同機 8 個 CPU 執行緒用時 2.87 s;首次 GPU 執行較慢。AI 服務只接收字典處理後的辨識文字、預設指示及正確拼寫資料,不會收到音訊。
正文
What happens between "hold the key" and "text at the cursor"
- Hold Ctrl+Alt+Space, speak, let go; the text is pasted at the cursor in any app, then the clipboard is put back. Not an input method; the microphone is open only while the shortcut is held. No wake word, nothing spoken gets executed.
- The order: recognition → dictionary replacements → AI polish (optional) → replacement rules → paste.
- (Your example: "Um, the review is on Monday at ten. No, sorry, at two…" → "The review is on Monday at two. Please bring the latest numbers." The gif shows it.)
Recognition on your own machine
- Five downloadable models: Qwen3-ASR 0.6B (690 MB, 30 languages, punctuation, GPU or CPU), Qwen3-ASR 1.7B (1.7 GB), SenseVoice Small (240 MB, CPU), Paraformer (227 MB, no punctuation, CPU), live transcription (169 MB, preview only).
- Downloads: project mirror → Hugging Face → hf-mirror.com; resumable; disk space checked; every file verified against its SHA-256. Offline import via "Check and import".
- GPU only for Qwen3-ASR: Vulkan on Windows/Linux, Metal on macOS; falls back to the CPU.
- Your numbers (16 s of Chinese, Qwen3-ASR 0.6B, after the first run): L40S Vulkan 0.16 s, same machine CPU 8 threads 2.87 s, Tesla T4 Vulkan 0.28 s. First GPU run is slow (~13 s on the L40S, ~2 min on the T4), so the model is loaded in the background at start.
- x86 needs AVX2 for Qwen3-ASR. AMD/Intel GPU measurements are still on the roadmap (ask readers for numbers).
The phone as the computer's microphone
- Pairing: QR code or 6-digit code, both screens show the same safety code.
- Audio: Opus, end-to-end encrypted, streamed through a relay that cannot read it; the computer recognises with its own settings. The phone runs no local model.
voltip-server
- Same recognition, dictionary and presets on a headless Linux server; OpenAI-compatible transcription API; accepts connections from that machine only unless opened up.
What the AI service receives
- Text only: recognised text after the dictionary, the preset's instructions, the dictionary's right spellings, the app name (can be turned off). Never audio.
- Keys in the system keychain; each key sent only to its own provider. Fallback models (up to 8) only on an explicit quota-exhausted answer.
Trade-offs (why there is no wake word / silence auto-stop, admin windows on Windows,
unsigned Windows installer, macOS not notarized, iOS planned)
Links
來源:DEV Community · dev.to