進階特色Advanced features
把即時翻譯變得更實用的那些細節。The details that make live translation genuinely useful.
同時轉錄麥克風Transcribe your mic
所有即時模式加上 --mic 即可同時轉錄自己的麥克風語音，雙向模式自動啟用。Add --mic to any live mode to also transcribe your own voice; auto-on for bidirectional.
會議主題感知翻譯Topic-aware translation
可指定會議主題（如「ZFS 儲存管理」），讓 LLM 依領域上下文精準翻譯專業術語。Set a meeting topic (e.g. "ZFS storage") so the LLM translates domain jargon accurately.
自動偵測 LLM 伺服器Auto-detect LLM server
支援 Ollama、LM Studio、Jan.ai、vLLM、LocalAI、llama.cpp、LiteLLM，自動辨識伺服器類型。Detects Ollama, LM Studio, Jan.ai, vLLM, LocalAI, llama.cpp and LiteLLM automatically.
互動式選單 + CLIInteractive menu + CLI
新手友善的選單介面，進階用戶可用命令列參數直接啟動；選單最後顯示等效 CLI 指令。A beginner-friendly menu, or launch directly via CLI; the menu prints the equivalent command.
WebUI 瀏覽器介面Web UI
--webui 在瀏覽器操作所有功能，支援即時字幕、離線處理、講者辨識、摘要，手機 / 平板也可使用。--webui drives everything in the browser (live, offline, diarization, summaries), phone/tablet friendly.
背景降噪Background denoise
即時模式可加 --denoise 啟用背景降噪，提升嘈雜環境的辨識品質。Add --denoise in live mode to clean up noisy environments before recognition.
Qwen3-ASR（實驗）Qwen3-ASR (experimental)
離線處理中文／英文／韓文錄音時可選，中文會議字錯率 28.78% → 15.75%。GPU 伺服器或本機都能跑，詳見使用說明。Optional for offline Chinese / English / Korean files: real Chinese meetings CER 28.78% → 15.75%. Runs on a GPU server or locally; see the guide.
GPU 伺服器版本檢查GPU server version check
伺服器上的辨識服務是獨立的程式，不會跟著本機升級。連上時自動比對版本，不一致會提示但不中斷作業。另有選填的自動更新（預設關閉）。The server's recognition service is a separate program that doesn't upgrade with the client. Versions are compared on connect: a mismatch warns you without interrupting the job. Optional auto-update is off by default.
關鍵字即時通知Keyword alerts
設定關鍵字，即時辨識出現時自動發出通知，可用於追蹤會議重點，或線上課程摸魚時讓系統在「請實作」「這個會考」時自動提醒。支援全螢幕警示特效、瀏覽器推播、音效提示（警示 / 柔和可選）、懸浮字幕閃爍，同一關鍵字冷卻機制避免重複通知。Set keywords and get notified the moment they're spoken: track key topics, or get pinged on "let's implement this" / "this is on the exam". Full-screen effects, browser push, sound (alert / soft), overlay flashing, and a per-keyword cooldown.
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、滑鼠穿透模式、字幕切換淡入淡出動畫，單語 / 雙語自動切換高度。（感謝 OSSLab 熊大提供建議）A translucent desktop overlay (PyQt6) over any app: auto-scaling text, drag & resize, click-through, fade animations and auto height for mono/bilingual lines.
台灣華語朗讀（v2.27.0）Taiwanese Mandarin text-to-speech (v2.27.0)
WebUI「輸入來源」選「文字內容朗讀」：貼上文字或選文字檔（本工具的摘要只念「重點摘要」），按「開始」後跟辨識一樣進入字幕畫面，每一段開始念才顯示那一段，字幕模式、懸浮字幕都能用；選「文字轉語音檔」則不播放、直接存成 MP3／WAV。聲音依性別分組，語速 0.8～1.5 倍（音調不變）、段落停頓可調；每一句顯示合成與等待的時間，可以「從這段念」重念任一句；可以播放到這台電腦的喇叭，或在開著網頁的那台電腦的瀏覽器播放（伺服器沒有喇叭也能用），也可以同時存檔。在自己的 GPU 伺服器或 Apple Silicon Mac 上用 VoxCPM2 合成，不使用作業系統內建語音。念法以教育部《重編國語辭典修訂本》為主、g2pW 依上下文判斷破音字（垃圾念ㄌㄜˋ ㄙㄜˋ），管理者可以加自訂發音（預設「和」念ㄏㄢˋ）。內建 8 個聲音（女聲、男聲各 4 個，AI 依文字描述產生、不是真人錄音），裝好就能念；也可以匯入自己的：管理者匯入一段 10～20 秒的台灣華語錄音與逐字稿，必須取得錄音者的書面同意。Choose "文字內容朗讀" (read text aloud) as the WebUI input source: paste text or pick a text file (this toolbox's summaries are read from their key points only), press Start, and the usual subtitle view shows each segment as it is spoken, including subtitle mode and the floating overlay. "文字轉語音檔" (text to audio file) skips playback and saves MP3 or WAV. Voices are grouped by gender; speed is 0.8-1.5x with pitch kept, and the pause between segments is adjustable. Each sentence shows its synthesis and wait time, and any sentence can be replayed from that point. Play on this computer's speakers, or in the browser of the computer viewing the page (works on servers without speakers), and optionally save to a file. Synthesis runs on your own GPU server or Apple Silicon Mac with VoxCPM2, never the OS voices. Readings follow Taiwan's Ministry of Education dictionary, with g2pW resolving polyphonic characters in context; admins can add custom pronunciations. 8 built-in voices (4 female, 4 male, AI-generated from text descriptions, not real people) work out of the box; to use your own, an admin imports a 10-20 second Taiwanese Mandarin recording with its transcript, with the speaker's written consent.
預設用 VoxCPM2，是為了速度與效能：GPU 伺服器比說話快一點、Mac 本機大約跟說話一樣快，兩種機器都能跑。它主要學的是大陸念法，所以我們下了不少功夫讓它念台灣話：念法以台灣日常說法為準、教育部辭典為基礎，破音字依上下文判斷，辭典跟日常說法不同的改用日常說法（市場、強制、參與、擷取、液化、亞洲、包括、角色）；用近 400 句台灣華語句子自動合成、再辨識回來逐字比對，並把將近 5,000 句裡指定的念法逐一檢查是不是台灣日常說法，找出念錯的地方一一修正；模型不認得的罕用字一律標注音，念錯的地方少了約八成。
另外可以選 BreezyVoice（MediaTek）：台灣口音道地，但合成速度慢，不適合即時（合成時間約音訊長度的 1.2～2.5 倍），只在 GPU 伺服器提供，適合轉成語音檔、不趕時間的時候。VoxCPM2 is the default for speed and efficiency: slightly faster than real time on the GPU server and about real time on a Mac, on both kinds of machine. It learned mostly Mainland readings, so we put real work into making it read like Taiwan: readings follow everyday Taiwanese usage, built on the Ministry of Education dictionary with polyphonic characters resolved in context, and where the dictionary differs from how people in Taiwan actually speak, everyday usage wins; we auto-synthesized about 400 Taiwanese Mandarin sentences, transcribed them back and compared character by character, and checked every assigned reading in nearly 5,000 sentences against everyday Taiwanese usage, fixing every misreading found; rare characters the model does not know always get a phonetic hint. Misreadings dropped by about 80%.
You can also choose BreezyVoice (MediaTek): a native Taiwanese accent, but slow to synthesize and not suited to live reading (about 1.2-2.5x the audio length). GPU server only; best for saving to audio files when time is not pressing.
讓別的系統把語音變成可引用的逐字稿Let other systems turn speech into citable transcripts
jt-live-whisper 可以當成語音處理的後端：別的系統把錄音丟進來，拿回帶時間、帶發言者、校正過的逐字稿，再拿去做自己的事，例如產生會議摘要、決議與待辦，而每一條都指得回「誰在第幾分鐘說的」。辨識、講者分離與校正都在你自己的機器上跑，錄音與逐字稿不會離開內網。jt-live-whisper can act as a speech backend: another system submits a recording and gets back a timed, speaker-tagged, corrected transcript to build on: meeting summaries, decisions and action items, each traceable to who said it and when. Recognition, diarization and correction all run on your own hardware; nothing leaves your network.
| 拿到什麼What you get | 用途Why it matters |
|---|---|
| 原始辨識（每段含開始 / 結束時間、語言、信心值）Raw ASR: per-segment start/end time, language, confidence | 時間軸不會變，用來當「指回原音」的錨點A stable anchor back to the original audio |
| 校正後文字（含「這段有沒有被改過」）Corrected text: with an "edited" flag per segment | 簡繁、標點、專有名詞修對；改動幅度異常時自動保留原文Fixes script, punctuation and proper nouns; abnormal edits are rejected and the original kept |
| 發言者對應（S1、S2… 代號）Speaker mapping: S1, S2 … labels | 把每一句掛到人身上，姓名由你那邊對應Attribute every line to a person; you supply the real names |
為什麼分三層而不是一份合併好的結果：校正後的文字會變（換校正等級、更新詞彙庫重跑），而原始辨識的段號與時間不會變。把引用綁在段號上，重跑校正之後已經產生的決議與待辦仍然指得回同一段原音；時間若複製到校正層去，就會多一份可能不一致的資料。Why three layers instead of one merged result: corrected text changes (different correction level, updated glossary, re-runs) while the raw segment numbers and timings do not. Bind citations to the segment number and your decisions survive a re-correction; copying timings into the corrected layer would create a second source of truth that can drift.
實測（GPU 伺服器、28.3 分鐘的真實會議）：辨識 37 秒、講者分離 2 秒、校正 67 秒，總共 1 分 52 秒跑完，約為即時的 15 倍速。送件當下就開始拉檔，輪到處理時音訊已經在本機（預抓耗時 0 毫秒）。Measured on a real 28.3-minute meeting with a GPU server: 37 s recognition, 2 s diarization, 67 s correction; 1 min 52 s end to end, about 15× faster than real time. Audio is pre-fetched at submission, so the fetch stage costs 0 ms.
已上線的整合：jt-doc-toolsIn production: jt-doc-tools
jt-doc-tools 已經接上 jt-live-whisper。使用者在 jt-doc-tools 裡上傳一段會議錄音，整個流程是這樣走的：jt-doc-tools is already wired to jt-live-whisper. A user uploads a meeting recording there, and this is what happens:
| 做什麼Step | 誰做Where | |
|---|---|---|
| 1 | 在「會議錄音轉逐字稿」上傳錄音或錄影（m4a / mp3 / wav / mp4 / mkv…）Upload audio or video to Meeting recording → transcript | jt-doc-tools |
| 2 | 辨識、講者分離、逐字稿校正Recognition, diarization and transcript correction | jt-live-whisper |
| 3 | 逐字稿可邊聽邊看：波形圖、點波形跳播、播到哪段就亮起來、每位發言者一個顏色、點代號直接改成人名Play along with the transcript: waveform, click-to-seek, the current line highlights, a colour per speaker, click a label to rename it | jt-doc-tools |
| 4 | 一鍵轉交「會議摘要」，產出摘要、決議、待辦、風險與章節One click to Meeting summary: key points, decisions, action items, risks and chapters | jt-doc-tools |
| 5 | 每一條決議與待辦都指得回原文的第幾段、誰在第幾分鐘講的Every decision and action item cites the exact segment: who said it, at what time | 兩邊共同Both |
第 5 點是整個整合的重點，也是為什麼要分三層：摘要可以被驗證。使用者看到一條「決議：下週前完成 A」時，點下去就能聽到當時是誰說的，而不是只能相信 LLM 沒有捏造。Step 5 is the whole point, and the reason for the three layers: summaries are verifiable. When a user sees "Decision: finish A by next week", they can click through and hear who actually said it, instead of having to trust that the LLM didn't invent it.
已上線的整合：jt-vc-portalIn production: jt-vc-portal
jt-vc-portal（Jitsi Meet 會議管理系統）把會議錄影交給 jt-live-whisper，拿回逐字稿與會議摘要。它自己沒有摘要功能，所以摘要也在這邊做：jt-vc-portal (a Jitsi Meet meeting portal) hands its meeting recordings to jt-live-whisper and gets back the transcript and the meeting summary; it has no summarizer of its own, so the summary is done here too:
| 做什麼Step | 誰做Where | |
|---|---|---|
| 1 | 會議結束後上傳錄影（影片直接傳），會議室、主持人、與會者一起送上After the meeting, upload the recording (video as-is) along with room, host and participants | jt-vc-portal |
| 2 | 取出音軌、辨識、講者分離、逐字稿校正、會議摘要Extract audio, recognize, diarize, correct, and write the meeting summary | jt-live-whisper |
| 3 | 取回逐字稿與會議摘要（重點摘要、決議、待辦、風險、議題）Fetch the transcript and summary (key points, decisions, action items, risks, topics) | jt-vc-portal |
| 4 | 每一條決議與待辦都附錄影裡的時間點與講者Every decision and action item carries its time in the recording and the speaker | 兩邊共同Both |
錄影在 jt-live-whisper 這邊處理完就刪除；逐字稿與摘要在對方取回並確認後刪除。jt-live-whisper deletes the recording once it's processed, and the transcript and summary once the portal has fetched and acknowledged them.
上面兩個整合是透過 REST API（jtlw_api/）接進來的。v2.25.1 起它隨 jt-live-whisper 一起發佈，伺服器版裝好就有；要不要啟用由你決定，設定方式見完整手冊「啟用 REST API」。Both integrations above connect through a REST API (jtlw_api/). Since v2.25.1 it ships with jt-live-whisper and is ready on a server install; it stays off until you enable it; see "Enable the REST API" in the manual.
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），可同時啟用多個平台、自訂發送間隔與內容（含時間 / 原文 / 譯文）。通用 API 支援 Body 範本（{{text}} 變數）搭配自訂 Headers。Forwards live subtitles to Telegram / Slack / Discord / Teams / LINE / Nextcloud Talk / a custom API, multiple at once, with custom intervals and content. The custom API supports a body template ({{text}}) with custom headers.
雙向語音口譯（英中，v2.28.0）Two-way voice interpretation (English-Chinese, v2.28.0)
英中雙向即時模式可以把譯文念出來：對方說的英文翻成中文念給你聽（耳機），你說的中文翻成英文念給對方聽（經虛擬麥克風送進會議）。任何線上會議軟體都能用（Zoom、Microsoft Teams、Google Meet、Jitsi Meet…）：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；下圖用 Jitsi Meet 示範。字幕與逐字稿照舊。合成在自己的 GPU 伺服器，第一段約 0.3～0.7 秒就出聲；GPU 忙、合成比說話慢時先存一段再播，不會播到一半斷掉。念給你聽的中文被系統音訊錄到時，GPU 伺服器先判斷是中文就不辨識；念過的句子 20 秒內又被錄回來（耳機漏音、對方沒有回音消除）也會略過，不會把自己的話翻來翻去。跟不上時給對方的優先、等太久的不念（字幕照樣顯示）；預設語速 1.2 倍，口譯中在字幕頁底部就能調快調慢；每一句標口譯狀態，還沒念給對方的可以取消。英文用內建的 2 個 AI 英文聲音（不是真人錄音，預設男聲），數字、金額、IP、版本號先換成念得對的寫法。預設對方只聽到英文，也可以同時送出你的原聲（念英文時自動調小）。念給對方聽要有虛擬麥克風：macOS 要先安裝 BlackHole 2ch（免費，brew install --cask blackhole-2ch），Linux 自動建立；Windows 要先安裝 usbip-win2（免費、開放原始碼，驅動由微軟簽署；.\install.ps1 -InterpMic），之後開始時自動建立「jt-live-whisper Interpreter Mic」、結束時移除。念給我聽三個平台都不用另外安裝。In English-Chinese two-way live mode, translations can be spoken: what the other side says in English is read to you in Chinese (headphones), and what you say in Chinese is read to them in English through a virtual microphone. Any online meeting app works (Zoom, Microsoft Teams, Google Meet, Jitsi Meet…): it listens to the meeting audio your computer plays and speaks through a microphone, so the meeting app needs no integration; the picture below uses Jitsi Meet. Subtitles and transcripts work as before. Synthesis runs on your own GPU server and the first audio arrives in about 0.3-0.7 s; when the GPU is busy and synthesis is slower than speech, a little audio is buffered first so playback never stalls. When your Chinese playback is picked up by the system audio capture, the GPU server detects it is Chinese and skips it; lines spoken in the last 20 seconds that come back (headphone bleed, no echo cancellation on the other side) are dropped too, so your own words are never translated back and forth. When falling behind, lines for the other side go first and stale lines are skipped (subtitles still show); speech runs at 1.2x by default and can be made faster or slower from the bottom of the subtitle page while interpreting; each line shows its status and lines not yet spoken to the other side can be cancelled. English uses two built-in AI-generated English voices (not recordings of real people; male by default), with numbers, money, IP addresses and version numbers rewritten so they are read correctly. By default the other side hears only English; you can also send your own voice alongside (lowered automatically while English is playing). Reading to the other side needs a virtual microphone: on macOS install BlackHole 2ch first (free, brew install --cask blackhole-2ch); Linux creates one automatically; on Windows install usbip-win2 first (free, open source, Microsoft-signed driver; .\install.ps1 -InterpMic), after which the "jt-live-whisper Interpreter Mic" is created when you start and removed when you stop. Reading to you needs nothing extra on any platform.
