本頁On this page 為什麼Why 核心功能Features 部署Deploy 畫面Screenshots 安裝Install 升級Upgrade

聽得見的每一句話都即時變成你看得懂的字幕Every word you hearlive, in words you know

jt-live-whisper 是 100% 全地端的 AI 語音工具箱：即時轉錄、即時翻譯、錄音檔批次處理、講者辨識、會議摘要與台灣華語朗讀。所有 AI 模型都在你自己的設備上執行，資料不經過任何雲端服務。 jt-live-whisper is a 100% on-device AI voice toolkit: live transcription, live translation, batch audio processing, speaker diarization, meeting summaries and Taiwanese Mandarin text-to-speech. Every AI model runs on your own hardware; nothing ever touches the cloud.

三行指令安裝Install in 3 commands 已經裝過？升級指令Already installed? Upgrade 前往 GitHub 專案Go to GitHub
WebUI 英中雙向對話模式 / WebUI bidirectional chat
即時翻譯字幕Live subtitles 中日韓英互譯ZH / JA / KO / EN 講者辨識Speaker diarization AI 會議摘要AI summaries 雙向字幕Bidirectional 終端機 CLITerminal CLI WebUI 介面Web UI 懸浮字幕Overlay subtitles 關鍵字通知Keyword alerts 台灣華語朗讀Taiwanese Mandarin TTS 雙向語音口譯Two-way voice interpretation macOS · Windows · Linux

資料留在自己手上，才能安心開機密會議Your data stays yours, safe enough for confidential meetings

語音辨識、翻譯、講者辨識、摘要全部使用自有設備上的 AI 模型，不需要任何雲端 API Key，也不會把語音或會議內容上傳給第三方。適合企業內部會議、機密討論，以及任何在意隱私的場合。Speech recognition, translation, diarization and summarization all run on AI models on your own hardware; no cloud API keys, nothing uploaded to any third party. Built for internal meetings, confidential discussions and anyone who cares about privacy.

為何打造 jt-live-whisperWhy we built jt-live-whisper

一套工具，從即時翻譯到會後摘要全包，全程不離開你的設備。One toolkit, from live translation to post-meeting summaries, all without leaving your device.

完全地端執行Fully on-device

辨識、翻譯、講者辨識、摘要全部用自有設備上的 AI 模型，無需雲端 API Key、不上傳任何資料。Recognition, translation, diarization and summaries all use local AI models: no cloud API key, no uploads.

隱私安全Private & secure

會議內容、語音資料全程留在自有設備，適合企業內部會議與機密討論。資料保護與合規 →Meeting content and audio stay on your machine, ideal for internal and confidential discussions. Data protection & compliance →

零月租成本Zero subscription

不需要付費雲端 API（ChatGPT、Claude、Gemini 等），所有採用的 AI 模型皆為自由開源。No paid cloud APIs (ChatGPT, Claude, Gemini…); every model is free and open source.

不限應用程式Any application

採用系統音訊層級擷取，理論上任何軟體的聲音輸出都能處理（Zoom、Teams、Meet、YouTube、Podcast 等）。macOS 13 以上使用系統內建的 ScreenCaptureKit，免安裝虛擬音訊驅動、免建立多重輸出裝置。Captures at the system-audio level, so virtually any app's output works (Zoom, Teams, Meet, YouTube, podcasts…). On macOS 13+ it uses the built-in ScreenCaptureKit: no virtual audio driver, no Multi-Output Device.

功能完整Complete toolkit

從即時轉錄翻譯、離線音訊處理、講者辨識到 AI 摘要，一套搞定。From live transcription & translation to offline processing, diarization and AI summaries, all in one.

三行指令安裝Three-command install

貼上三行指令，安裝腳本自動下載並編譯所有 AI 模型和相依套件，macOS、Windows、Linux 都支援。The installer downloads and builds every AI model and dependency for you on macOS, Windows and Linux.

緣起：某次參加原廠的線上技術課程，全程英文授課，聽得七零八落。為了補足英文聽力的不足，乾脆動手打造了這套工具來即時翻譯，結果功能越做越多，就變成現在這個樣子了 XD Origin: Struggling to follow an all-English vendor training session, the author built a tool to translate it live, then kept adding features until it became what you see today. 😄

核心功能Core features

即時、離線、講者、摘要，會議全流程都覆蓋。Live, offline, speakers and summaries: the whole meeting workflow.

即時語音轉錄翻譯Live transcription & translation

擷取系統音訊，地端 AI 即時辨識並翻譯成繁體中文字幕顯示於終端機。開會、看影片、聽 Podcast 即時翻譯。Captures system audio and transcribes & translates it to subtitles in real time: meetings, videos, podcasts.

離線音訊檔批次處理Offline batch processing

支援 mp3 / wav / m4a / flac，使用 faster-whisper 離線轉錄翻譯，適合會後補做逐字稿。可選用 Qwen3-ASR（實驗）：中文會議字錯率 28.78% → 15.75%，中英夾雜更準；GPU 伺服器（v2.23.0）或本機（v2.24.0 起，Apple Silicon 用 MLX）都能跑。mp3 / wav / m4a / flac via faster-whisper, perfect for post-meeting transcripts. Optional Qwen3-ASR (experimental): Chinese meeting CER 28.78% → 15.75%, much better on code-switching. Runs on a GPU server (v2.23.0) or locally (v2.24.0+, MLX on Apple Silicon).

講者辨識Speaker diarization

自動辨識音訊中的不同講者並以不同顏色標示，支援自動偵測或手動指定講者人數。Detects different speakers and color-codes them; auto-detect or set the speaker count manually.

AI 會議摘要AI meeting summary

地端 LLM 整理出重點摘要、決議、待辦、風險與議題，每一條都附逐字稿的時間點，引用經過比對，另附校正逐字稿。A local LLM extracts key points, decisions, action items, risks and topics (each citing the transcript timestamp, with citations checked), plus a corrected transcript.

時間軸互動逐字稿Interactive transcript

HTML 逐字稿內嵌音訊播放器與波形圖，點波形即跳到該時間點，播放時對應段落即時高亮。HTML transcript with an embedded player and waveform: click to seek; the current line highlights as it plays.

接進你的系統Integrates with your systems

可當成語音後端，已接上 jt-doc-tools 與 jt-vc-portal：上傳錄音或會議錄影 → 辨識、分離講者、校正 → 產出會議摘要與決議，每一條都指得回誰在第幾分鐘說的。全程在自己的機器上跑。Use it as a speech backend: already wired into jt-doc-tools and jt-vc-portal: upload a recording or meeting video → transcribe, separate speakers, correct → summaries and decisions, each citing who said it and when. All on your own hardware.

伺服器模式Server mode

./install.sh --server 裝在無桌面的主機上，WebUI 以 systemd 服務常駐、開機自動啟動，區網其他電腦用瀏覽器就能操作。可限制來源網段、加上 TLS。./install.sh --server installs on a headless box; the Web UI runs as a systemd service on boot and anyone on the LAN can drive it from a browser. Source ranges and TLS are configurable.

雙向字幕 + 16 種模式Bidirectional & 16 modes

英中 / 日中 / 韓中雙向，同時擷取系統音訊與麥克風，對方外語翻中文、自己中文翻外語；另支援台語（台灣閩南語）辨識，共 16 種功能模式。EN↔ZH / JA↔ZH / KO↔ZH bidirectional from system audio + mic, plus Taiwanese Hokkien ASR; 16 modes in total.

台灣華語朗讀Taiwanese Mandarin TTS

把文字、講稿、摘要念出來，邊念邊顯示字幕，也能存成 MP3（v2.27.0）。模型 VoxCPM2 支援中、英、日、韓等 30 種語言，目前朗讀針對台灣華語調校（夾雜英文也行），雙向口譯也用它念英文。在自己的 GPU 伺服器或 Apple Silicon Mac 合成，可用內建 AI 聲音或匯入取得同意的錄音；也可選台灣口音的 BreezyVoice（較慢）。Read text, scripts and summaries aloud with live subtitles, or save MP3 (v2.27.0). The VoxCPM2 model supports 30 languages, including Chinese, English, Japanese and Korean; read-aloud is tuned for Taiwanese Mandarin (mixed-in English works) and interpretation uses it to speak English. Runs on your GPU server or Apple Silicon Mac with built-in AI voices or a consented voice you import; BreezyVoice offers a native Taiwanese accent (slower).

雙向語音口譯Two-way voice interpretation

英中線上會議：對方說的英文翻成中文念給你聽、你說的中文翻成英文經虛擬麥克風念給對方聽（v2.28.0）。邊合成邊播，第一段約 0.3～0.7 秒出聲；自己念的聲音被錄回去會自動略過，不會翻來翻去。English-Chinese online meetings: what they say in English is read to you in Chinese, and what you say in Chinese is read to them in English through a virtual microphone (v2.28.0). Streaming synthesis starts in about 0.3-0.7 s, and your own playback picked up again is dropped so nothing loops.

看完整功能詳解 →See all features →

兩種操作介面Two ways to operate

習慣終端機，或偏好瀏覽器圖形介面，兩種都行：同一套功能、同一份設定。Prefer the terminal or a browser GUI? Both drive the exact same features and config.

終端機 CLI + 互動式選單Terminal CLI + interactive menu

直接 ./start.sh 進入互動式選單，逐步引導完成所有設定；進階用戶可用命令列參數一行啟動。Run ./start.sh for a guided interactive menu, or launch directly with CLI flags for power users.

  • 新手友善的選單，音訊裝置全自動偵測Beginner-friendly menu; audio devices auto-detected
  • 選單最後顯示等效 CLI 指令，下次可直接用The menu prints the equivalent CLI command for next time
  • 字幕直接顯示在終端機，最輕量Subtitles right in the terminal, the lightest option

瀏覽器 WebUIBrowser Web UI

./start.sh --webui 在瀏覽器中完成所有設定與操作，不需記指令。即時 / 離線功能全包。./start.sh --webui does everything in the browser; no commands to remember. Covers live & offline.

  • 聊天 / 字幕模式切換、淺色 / 深色主題Chat / subtitle modes, light / dark themes
  • 各階段即時進度（辨識 / 講者 / 校正 / 摘要）Live progress for each stage (ASR / diarize / correct / summary)
  • 支援遠端觀看（密碼保護）、手機 / 平板Remote viewing (password-protected), phone / tablet

三種部署方式Three ways to deploy

一台電腦就能跑；要更快，再加一台 GPU 伺服器；要給整個團隊用，架成伺服器。Run on one machine, add a GPU server for speed, or host it as a server for your whole team.

單機模式Single machine

一台 Mac、Windows PC 或 Linux 桌機即可完成所有處理，不需要額外硬體。One Mac, Windows PC or Linux desktop handles everything, no extra hardware.

  • macOS Apple Silicon（M1–M4）：mlx-whisper Metal GPU 加速，辨識約 1–3 秒macOS Apple Silicon (M1–M4): mlx-whisper Metal GPU, ~1–3 s recognition
  • Windows + NVIDIA GPU：安裝程式自動啟用 CUDA，辨識約 0.5–1 秒Windows + NVIDIA GPU: installer auto-enables CUDA, ~0.5–1 s
  • Linux：PipeWire / PulseAudio 直接擷取系統音訊，免裝虛擬音效卡；有 NVIDIA GPU 時自動走 CUDALinux: captures system audio via PipeWire / PulseAudio with no virtual sound card; uses CUDA when an NVIDIA GPU is present
  • 無 GPU / Intel Mac：CPU 辨識，搭配 small 模型可用No GPU / Intel Mac: CPU recognition with the small model

本機 + GPU 伺服器Client + GPU server

本機負責音訊擷取與介面，辨識與講者辨識交給區網內的 GPU 伺服器（系統音訊與麥克風兩路都可送遠端）。The client captures audio & UI; recognition and diarization run on a LAN GPU server (both system audio & mic).

  • 離線辨識快 5–10 倍，即時辨識約 0.3–0.5 秒Offline 5–10× faster; live recognition ~0.3–0.5 s
  • 伺服器可為 DGX Spark 或裝有 NVIDIA GPU 的 Linux 主機（RTX 4090 / 5090 亦可）DGX Spark or any Linux box with an NVIDIA GPU (RTX 4090 / 5090 work too)
  • 伺服器離線自動降級為本機處理，不中斷使用Auto-falls back to local if the server is offline, no interruption

伺服器模式Server mode

裝在無桌面的 Linux 主機（install.sh --server），WebUI 常駐，區網內的電腦用瀏覽器上傳錄音處理（遠端操作需要管理密碼）。Install on a headless Linux host (install.sh --server): the Web UI runs as a service; computers on your LAN upload recordings from the browser (remote use needs the admin password).

  • 搭配 GPU 伺服器（建議）：2 vCPU、4 GB 記憶體、30 GB 磁碟即可，辨識與講者辨識都在 GPU 伺服器上（實測 37 分鐘會議，記憶體峰值約 0.4 GB）With a GPU server (recommended): 2 vCPU, 4 GB RAM, 30 GB disk is enough; recognition and diarization run on the GPU server (measured: 37-min meeting peaks at ~0.4 GB RAM)
  • 不搭配 GPU 伺服器：改在 CPU 上辨識，建議 8 核、8 GB 以上；長錄音要花數倍於錄音長度的時間Without a GPU server: CPU recognition; 8 cores and 8 GB+ RAM recommended; long recordings take several times their length
  • 已與 jt-doc-tools、jt-vc-portal 串接Already integrated with jt-doc-tools and jt-vc-portal
實際畫面Screenshots

畫面導覽See it in action

點任一張圖可放大。Click any image to enlarge.

即時英翻中字幕（macOS）
01

即時翻譯字幕Live translated subtitles

擷取系統音訊，地端 AI 一邊聽一邊辨識並翻成繁體中文，字幕即時顯示於終端機，並附上翻譯速度標籤與音訊波形。開會、看影片、聽 Podcast 都能即時跟上。Captures system audio and transcribes & translates it live, with speed badges and an audio waveform right in the terminal, so you can keep up with meetings, videos and podcasts.

WebUI 設定頁
02

WebUI 瀏覽器介面Web UI in your browser

--webui 在瀏覽器中完成所有設定與操作：即時字幕、離線處理、講者辨識、摘要全包，辨識模型依裝置自動推薦，各階段即時進度顯示，手機 / 平板也能用。With --webui, do everything in the browser: live subtitles, offline processing, diarization and summaries, with device-aware model recommendations, live progress, and phone/tablet support.

WebUI 字幕模式 - 雙向
03

雙向字幕模式Bidirectional subtitles

英中（en_zh）、日中（ja_zh）與韓中（ko_zh）雙向：同時擷取系統音訊與麥克風，對方外語翻中文、自己中文翻外語，適用於雙語線上會議。電影風格黑底大字，一眼看清。EN↔ZH, JA↔ZH and KO↔ZH: captures system audio and your mic at once: their language to Chinese, your Chinese to theirs. Big cinema-style captions for bilingual calls.

離線處理選單
04

離線批次處理Offline batch processing

匯入錄音檔即可離線轉錄翻譯。互動式選單依序引導模式、辨識位置與模型、翻譯引擎、講者辨識與摘要設定，最後顯示等效 CLI 指令方便下次直接用。Import a recording for offline transcription. The interactive menu walks you through mode, engine, diarization and summary, then prints the equivalent CLI command for next time.

講者辨識結果
05

講者辨識Speaker diarization

自動辨識音訊中的不同講者，以不同顏色清楚標示，支援自動偵測或手動指定 2–20 位講者。誰在什麼時候說了什麼，一目了然。Identifies different speakers and color-codes them; auto-detect or set 2–20 speakers manually. Who said what, when, at a glance.

AI 會議摘要：每一條都附逐字稿時間點
06

AI 會議摘要AI meeting summary

重點摘要、事件與影響、決議、待辦（負責人、期限）、風險、未決問題、議題時間軸、發言統計。每一條都附逐字稿的時間點，而且會比對引用的原文是不是真的講到這件事；點時間點就跳到依據的那段原文。分析與 jt-doc-tools 的會議摘要是同一套。Key points, events & impacts, decisions, action items (owner, due), risks, open questions, a topic timeline and speaker statistics. Every item cites a transcript timestamp, and each citation is checked against what was actually said; click a timestamp to jump to the source. Same analysis as the jt-doc-tools meeting summary.

時間軸逐字稿 HTML
07

時間軸互動逐字稿Interactive timeline transcript

時間逐字稿 HTML 內嵌音訊播放器與波形圖，可直接點波形任意位置跳至該時間點；播放時對應段落即時高亮，對照聆聽超方便。另可輸出 SRT / WebVTT 字幕檔。The HTML transcript embeds a player and waveform: click anywhere to seek, and the matching line highlights as it plays. SRT / WebVTT export too.

關鍵字即時通知
08

關鍵字即時通知Keyword alerts

設定關鍵字，即時辨識出現時自動全螢幕警示 + 音效提醒。可追蹤會議重點，或線上課程在「請實作」「這個會考」時自動提醒，內建冷卻機制避免重複通知。Set keywords and get a full-screen alert plus sound the moment they're spoken, great for tracking key topics, with a cooldown to avoid repeats.

懸浮字幕
09

懸浮字幕Overlay subtitles

桌面半透明字幕覆蓋視窗（PyQt6），可疊加於任何應用程式上方。字體依視窗大小自動縮放、可拖曳移動與調整大小、支援滑鼠穿透與淡入淡出動畫，單語 / 雙語自動切換高度。A translucent desktop overlay (PyQt6) that floats over any app: auto-scaling text, drag & resize, click-through mode and fade animations.

字幕轉發到 Telegram
10

字幕轉發Subtitle forwarding

即時字幕自動轉發到通訊平台：Telegram / Slack / Discord / Teams / LINE / Nextcloud Talk / 通用 API。可同時啟用多平台、自訂發送間隔與內容（含時間 / 原文 / 譯文）。Forwards live subtitles to Telegram / Slack / Discord / Teams / LINE / Nextcloud Talk / a custom API, multiple at once, with custom intervals and content.

台灣華語朗讀的設定畫面
11

台灣華語朗讀Taiwanese Mandarin text-to-speech

把文字、講稿、會議摘要念出來，邊念邊顯示字幕，也能直接存成 MP3。念法以台灣日常說法為準（夾雜英文也行），在自己的 GPU 伺服器或 Apple Silicon Mac 合成；內建 8 個 AI 聲音，也能匯入取得同意的錄音。Read text, scripts and meeting summaries aloud with live subtitles, or save them straight to MP3. Readings follow everyday Taiwanese usage (mixed-in English works), synthesized on your own GPU server or Apple Silicon Mac with 8 built-in AI voices or a consented voice you import.

12

線上會議雙向語音口譯Two-way voice interpretation in online meetings

英中雙向會議裡把譯文念出來：對方說的英文翻成中文念到你的耳機，你說的中文翻成英文、用合成語音送進會議。只有你的電腦要裝，對方照常開會、什麼都不用裝。下圖是 Mac（Bianca，說英文）與 Windows（Kevin，說中文）兩台電腦用 Jitsi Meet 實際開會的畫面；任何線上會議軟體都能用（Zoom、Microsoft Teams、Google Meet…）：jt-live-whisper 聽的是電腦播出的會議聲音、說的話經麥克風送進會議，不需要會議軟體配合。In English-Chinese meetings the translations are spoken: the other side's English is read to you in Chinese in your headset, and your Chinese is sent into the meeting in English with a synthesized voice. Only your computer needs it; the other side joins the meeting as usual with nothing to install. Below: a real Jitsi Meet call between a Mac (Bianca, English) and a Windows PC (Kevin, Chinese). Any online meeting app works (Zoom, Microsoft Teams, Google Meet…): jt-live-whisper listens to the meeting audio your computer plays and speaks through a microphone, so the meeting app needs no integration.

線上會議雙向語音口譯：Mac 與 Windows 兩台電腦用 Jitsi Meet 開會 Two-way voice interpretation in a Jitsi Meet call between a Mac and a Windows PC

三行指令安裝Install in 3 commands

安裝腳本自動下載並設定所有地端 AI 模型和相依套件。首次約 10–20 分鐘。The installer downloads and configures every local AI model and dependency. First run takes ~10–20 min.

macOS：終端機貼上即可macOS: paste into Terminal
mkdir -p ~/Apps/jt-live-whisper && cd ~/Apps/jt-live-whisper
curl -fsSL https://raw.githubusercontent.com/jasoncheng7115/jt-live-whisper/main/install.sh -o install.sh
bash install.sh
Linux（Ubuntu / Debian）：終端機貼上即可，會用 sudo 補齊系統套件Linux (Ubuntu / Debian): paste into a terminal; uses sudo for system packages
mkdir -p ~/Apps/jt-live-whisper && cd ~/Apps/jt-live-whisper
curl -fsSL https://raw.githubusercontent.com/jasoncheng7115/jt-live-whisper/main/install.sh -o install.sh
bash install.sh            # 無桌面主機加 --server
Windows：以系統管理員開啟 PowerShellWindows: open PowerShell as Administrator
mkdir C:\jt-live-whisper -Force | Out-Null; cd C:\jt-live-whisper
irm https://raw.githubusercontent.com/jasoncheng7115/jt-live-whisper/main/install.ps1 -OutFile install.ps1
powershell -ExecutionPolicy Bypass -File install.ps1
啟動（推薦 WebUI）Launch (Web UI recommended)
# macOS / Linux
./start.sh --webui
# Windows
.\start.ps1 --webui
已經裝過：升級到最新版Already installed: upgrade to the latest version
# macOS / Linux
cd ~/Apps/jt-live-whisper
./install.sh --upgrade
# Windows (PowerShell)
cd C:\jt-live-whisper
.\install.ps1 -Upgrade

上面是預設的安裝資料夾，裝在別的地方就切到你的安裝資料夾。升級會從 GitHub 下載最新版，執行一次就好。升級完把開著的 WebUI 關掉再重新啟動，開著的還在跑舊程式（Linux 伺服器版會自動重啟服務）。macOS 與 Windows 升級後請再執行一次安裝腳本（./install.sh／.\install.ps1，不加 --upgrade），新版本需要的套件才會裝上；Linux 會自動補。Windows 若顯示「已停用指令碼執行」，改用 powershell -ExecutionPolicy Bypass -File install.ps1 -Upgrade。These are the default install folders; if you installed elsewhere, cd into your install folder. The upgrade pulls the latest version from GitHub; one run is enough. Close any running Web UI and start it again afterwards, it is still running the old code (the Linux server edition restarts its service automatically). On macOS and Windows, run the installer once more (./install.sh / .\install.ps1, without --upgrade) so packages the new version needs get installed; Linux does this automatically. If Windows says running scripts is disabled, use powershell -ExecutionPolicy Bypass -File install.ps1 -Upgrade.

不裝 LLM 也能翻譯：程式可切換為 NLLB（中日韓英互譯，模型僅限非商業用途）或 Argos（僅英翻中）離線翻譯引擎，完全不需要額外伺服器。注意：摘要功能仍需 LLM 伺服器（推薦 Ollama）。 No LLM needed to translate: switch to the NLLB (ZH/JA/KO/EN; the model is non-commercial only) or Argos (EN→ZH) offline engines; no server required. Summaries still need an LLM server (Ollama recommended).

完整安裝與使用說明 →Full install & usage docs →