逐字稿文字預覽
New Google AI updates are absolutelyinsane and this week was one of thebiggest weeks Google has had all year. Five things dropped, a new voice modelthat talks and thinks at the exact sametime, an AI agent that got its own Google account and now runs a wholehousehold, a brand new image tool builtstraight into Google Docs and Slides, a personal story app that went live foreveryone in the US, and a map of 9billion changes to human DNA. I want togo through all five, but I want to start with the one that changes how you workbecause I think most people are going tomiss why it matters. On September 15th, Google launched Gemini 3.8 live andGemini 3.8 live extended thinking.Google calls them their most advanced live dialogue models yet. They speak 97languages, they're built on Gemini 3Pro, and they do something no voice assistant has ever really done well.Hey, if we have not met already, I amthe digital avatar of Julian Goldie, CEO of Goldie Agency. While he's out therehelping clients grow, I am here keepingyou up to date on everything AI. Julian reads every comment, so drop one belowand he will see it. Here's the problemwith every voice AI you've used up to now. You ask it something, it needs tolook something up, and it goes quiet.Dead silence. You sit there wondering if the thing crashed. Then 30 secondslater, it comes back with an answer.That gap killed voice AI for real work. Nobody wants to talk to something thatfreezes mid-sentence. Gemini 3.8 livefixes that. It runs the look up in the background while it keeps talking toyou. Google's own words describe it asrunning tool calls without interrupting the conversation. The model can say,"Let me check that." and actually gocheck while still speaking. The extended thinking version goes further. Itnarrates progress as it works, so it'lltell you what step it's on while it's doing it. Think about a real persondoing a task for you, they don't gosilent. They say, "Okay, I'm pullingthat up now, give me a second. All right, I see it." That's the behaviorGoogle just built into a model. Googlealso shipped Gemini 3.5 Transcriber alongside it. That's a dedicatedspeech-to-text model covering 85 pluslanguages with low error rates. Now, before I get into the agent that got itsown Google account, and this one isgenuinely strange, I want to saysomething quick. Voice models like Gemini 3.8 live only pay off if you knowhat to point them at. That's the wholegame now. Inside the AI ProfitBoardroom, we've got a 30-day roadmap for exactly this, how to set up voiceagents that handle your follow-ups, yourclient calls, your inbox hands-free. We run four coaching calls every week, andright now a lot of those calls arepeople working through Gemini andWorkspace automations live. Bring your setup, we'll fix it on the call. There'sa prompt library built for Google'stools specifically, daily tutorials that walk you through new releases like thisone step by step, so you're notguessing. And business owners in there, a lot of them already running Geminiagents in their business today. There'sa member map, so you can find people near you doing the same thing. Link inthe comments and description, or go toaprofitboardroom.com. Okay, the agent with its own Google account. OnSeptember 17th, Google Labs announced anew version of CC. CC started inDecember 2025 as a personal productivity agent. It got popular fast with earlytesters. In May, Google brought it intothe Gemini app as Daily Brief. Now, Google rebuilt it completely. CC is nowan agent for families and households,and here's the part that made me stop. CC now has its own Google account, notyour account, its own, with its ownemail address, its own identity, its own permission set. Up to six people canjoin one CC. Each person picks what theyshare with it. Emails from your kids' school, the swim center, the vet, youcan forward things to CC's emailaddress, or you can set it so certainsenders always go to CC automatically. Google says CC will even suggest sendersto add each week. Then CC sorts throughall of it and builds one shared daily brief for the whole household. Googlecalls it Your day ahead. It connects tocalendar and tasks. It adds events automatically based on what peopleshare. But, it doesn't stop at reading.It does things. CC can fill outpermission slips. It can fill in registration PDFs. It can build schoolsupply shopping lists. It can planweekly meals. It can work out drive times between activities. So, you know,if you can actually make both. It cancreate shared docs and sheets. And if it's missing something it needs, itasks. Then, it updates its group memoryso it doesn't have to ask again. All right, third one. This one is moreimmediately useful for most peoplewatching. Google picks. It wentgenerally available on September 1st. It's Google's AI image creation andediting tool. And it's built directlyinto workspace. It runs on Nala Banana, Google's image model. The thing thatmakes picks different from most AI imagetools is object-level editing. Most image AI works one shot. You type aprompt, you get an image. And if onething is wrong, you start over. Newprompt, new image, new everything. Then, you lose the parts you liked. Picks letsyou select a single object inside theimage and change just that. The rest ays put. You can edit text inside animage. You can translate words inside animage. You can isolate one element and swap it out. It also generates severaloptions from one prompt so you pickinstead of re-rolling. You get to it at picks.new. It works as a standalone appon the web and it's built into docs andslides already. Select an image inside a doc or a slide, click once and you'rediting with the full picks tool setwithout leaving the tab. Driveintegration is rolling out. And Google said you'll be able to open and editnearly any image saved in Drive with oneclick. And it collaborates like the rest of workspace. Share a link. Yourteammate edits the same image. Same as adoc. Dream beans. It's a Google Labs app. It launched b
本次執行流程與問題處理
下達指令
預期流程
- 自動字幕優先，Whisper 作為備援
- 來源選擇與品質檢查
- 文字校正與句子單元
- 雙語翻譯
- 雙語時間軸與學習單元
- 筆記與發布
- 公開網頁發布
1. 下載與語音辨識
工具：videonote-transcribe-url（主辨識管線：下載、Whisper 辨識與選稿）
處理：下載音訊、以 Whisper 辨識並選出最佳逐字稿
產出：yt_dlp_meta.json（影片元資料）、audio_16k_mono.wav（16kHz 單聲道音訊）、transcript.whisper.json（Whisper 辨識結果）、transcript.json（正式逐字稿）、transcript_selection_report.json（來源選擇報告）
耗時：658.7 秒
關係：依序執行
結果：完成
2. 全文快速校讀
工具：videonote_rapid_transcript_review.py
處理：先速讀完整逐字稿，再修正高可信突兀詞句並列出待確認項目
產出：rapid_transcript_review_report.json（產出檔案）
耗時：40.6 秒
關係：依序執行
結果：完成
3. 建立句子單元
工具：videonote-build-sentence-units
處理：將逐字稿切成帶時間軸的句子單元
產出：sentence_units.json（句子單元）
耗時：0.7 秒
關係：依序執行
結果：完成
4. 時間軸稽核
工具：videonote_timeline_audit.py
處理：檢查時間軸與音訊對齊
產出：timeline_audit_report.json（時間軸稽核）
耗時：0.7 秒
關係：依序執行
結果：完成
5. 品質檢查
工具：videonote_controlled_quality.py（檢查時間軸與品質）
處理：檢查語言、時間軸與可讀性品質
產出：controlled_quality_report.json（品質報告）、language_sanity_report.json（語言檢查）
耗時：0.7 秒
關係：依序執行
結果：完成
6. 播放時間軸回填
工具：videonote_rebase_display_timeline.py
處理：將網頁播放時間軸回填為校正字幕時間
產出：display_timeline_rebase_report.json（時間軸回填報告）
耗時：0.7 秒
關係：依序執行
結果：完成
7. 雙語翻譯
工具：vnt-sentence（句子雙語翻譯）
處理：將句子翻譯成繁體中文
產出：sentence_bilingual_transcript.json（雙語逐字稿）
耗時：92.7 秒
關係：依序執行
結果：完成
8. 雙語時間軸與學習單元
工具：videonote_learning_timeline_v2.py（建立雙語時間軸與學習單元）
處理：建立雙語時間軸與學習內容
產出：sentence_bilingual_timeline.html（雙語時間軸頁）、learning.json（學習單元）、japanese_learning.json（日文學習單元）
耗時：99.0 秒
關係：依序執行
結果：完成
9. 生成筆記摘要
工具：videonote-generate-note（生成筆記摘要）
處理：生成筆記摘要與重點
產出：note.md（筆記 Markdown）、note.html（筆記 HTML）
耗時：140.6 秒
關係：依序執行
結果：完成
10. 建立筆記頁面
工具：videonote-render-html（產生筆記 HTML）
處理：產生可閱讀的筆記 HTML 頁面
產出：note.html（筆記 HTML）
耗時：0.6 秒
關係：依序執行
結果：完成
11. 發布公開頁面
工具：videonote-publish（發布公開頁面）
處理：發布到公開網站與每日列表
產出：公開筆記網址、result.json（管線結果）
耗時：0.7 秒
關係：依序執行
結果：完成
手動來源字幕
自動來源字幕
Whisper 字幕
- 尚未產生
Qwen 字幕
- 尚未產生
Parakeet 字幕
- 尚未產生
MOSS 字幕
- 尚未產生
聲學時間證據（PyAnnote / WhisperX）
- 尚未產生
校正前後逐字稿
- 校正前 transcript.source.json 純文字 JSON
- 校正後 transcript.json 純文字 JSON
目前待解決問題與初步改善方向
- 逐字稿全文快速校讀未完成
已保留自動字幕/原始逐字稿並繼續生成筆記。；未完成的全文速讀項目可在精進模式回聽確認。；RAPID_REVIEW_STATUS=REVIEW_REQUIRED_INVALID_OUTPUT
- Qwen3-ASR 未完成，需確認 qwen_asr_lab 環境與音訊後重試精進模式。
- 音訊下載不穩定，需更新 cookies / PO token 後重試精進模式。