TIO Voice Genie User Guide

File	When you use it
`1_安裝說明`	This guide, in printable form
`2_申請金鑰與儲值`	You have to do this the first time, about 10 minutes
`3_安裝.bat`	Double-click it and it installs everything
`4_開始使用.bat`	Opens the menu. Normally you just double-click the "TIO語音精靈" shortcut on your desktop (it runs this file); double-click this one only if there's no desktop shortcut
`5_使用說明`	How each feature works; look it up any time
`6_診斷.bat`	Run it when something goes wrong, then screenshot the window and send it in
`7_解除安裝.bat`	When you don't want it any more
`8_刪不掉時救援.bat`	Only when the folder refuses to delete

Item	Notes
Windows 10 / 11	Right now it's Windows only — capturing "whatever the PC is playing" uses a Windows-only technology, so a Mac can't do it
Python 3.12 or newer	Fine if you don't have it: when the installer sees it's missing, it installs it for you, and it doesn't need administrator rights
An internet connection	It talks to Google's servers, so it can't work offline
A Gemini API key	You apply for your own — see the next section

If this happens	What to do
You can't find the "Unblock" tick box	That means the file isn't marked and won't be blocked — just double-click it
You'd already unzipped it, then went back and ticked "Unblock" on the zip	That doesn't help — the mark was copied onto the unzipped files long before. Use the right-hand one above instead (tick it on `3_安裝.bat`), or delete the folder and unzip again
An "Open File - Security Warning" window pops up	Just click Run. It only asks this once; as soon as the installer starts, it clears the mark
A blue "Windows protected your PC" screen appears	First click More info, then Run anyway
The warning has no way to proceed at all, and the message mentions Smart App Control	This is Windows 11's protection feature running in enforcement mode. You can fix this yourself (tested on a laptop, 2026-09-19): delete the folder you already unzipped → right-click the zip → Properties → tick Unblock → OK → unzip it again, then double-click `3_安裝.bat`
The warning has no way to proceed at all, and the message mentions an administrator or Group Policy	This PC is locked down by your organisation's IT policy, and you can't get around it yourself. Take a screenshot of that window and send it, together with the `6_診斷.bat` screen, to your organisation's IT staff

What to watch	Why
When the balance hits zero, every key stops at once	All the keys under the same billing account stop together. If that happens halfway through a meeting, the captions simply stop. Turn on automatic top-up and set a monthly cap.
The daily quota resets at midnight US Pacific time	That works out at about 3–4 pm Taiwan time, not Taiwan midnight. Use up the quota in the morning and you wait until that afternoon for it to come back.
The quota belongs to the "project", not to the key	Hand out several keys from one project to different people and everyone shares one quota, so you crowd each other out. Each person should create their own project under their own Google account.

The six features

Find the one that matches your situation, then follow it. If you're not sure, number 2 is the one people use most.

The .json that feature 2 produces can go into feature 4 to be translated (no re-transcribing — you only pay for the translation), or be rebuilt with feature 5 (no cost at all)

1　Live meeting captions

The meeting is happening now and you want captions as it goes

Press 1 → Enter
It asks which sound to capture:
1 Online meeting (Teams/Meet/Zoom) — captures what comes out of the PC's speakers, so you get everything the other side says. Strictly speaking, it records Windows' default playback device, so the meeting app's speaker must be set to the same device (see the note below)
2 In-person meeting — captures the microphone (two extra questions follow; see below)
3 Both — you're in an online meeting and also want your own voice picked up (it also asks which microphone to use)
It asks for proper nouns → type the terms that will come up with a single space between them, like "深耕計畫研發處會計室" (Higher Education Sprout Project, Office of Research and Development, Accounting Office). This step really matters — leave it blank and it hears 會計室 (Accounting Office) as 會計師 (accountant)
It asks what this meeting is called → type anything, or just press Enter
Press Enter to start

Chose the microphone? It asks two more questions

Which microphone? (only asked if the PC has two or more) 1 (or just Enter) = let the program choose: it follows Windows' default, and switches over if you unplug or plug in a headset partway through. Any other number = stick with that one microphone the whole time: if it's unplugged the captions stop, and it reconnects by itself when you plug it back in. If you chose "Both" and fixed a microphone, unplugging the headset only stops the microphone side; whatever the PC is playing still gets captioned.

Is there sound from loudspeakers in the room that also needs capturing? (only asked for 2 In-person meeting) Normally press 1. Press 2 only if you need to capture what comes out of loudspeakers: full capture, bypassing Windows' noise suppression — but there's no guarantee it will pick them up; see hybrid meetings.

For a meeting with several people, don't use a headset (it's designed to sit right by one person's mouth): use a conference microphone if you have one, otherwise the laptop's built-in microphone. About 6 seconds after you start, a "🎚 收音檢查" ("sound check") appears on screen, showing whether sound is actually reaching the program — here's how to read it.

[00:12] Right, let's begin today's Office of Research and Development meeting.
[00:19] First item is the interim report for the Higher Education Sprout Project.
[00:24] Let me add something here — the deadline for clearing exp

Grey text = still being recognised; it keeps jumping and correcting itself, and that's normal. White text = locked in. Each sentence locks as soon as it's finished, so you don't have to pause.

How do I stop when the meeting ends? Press Ctrl + C in that black window (hold Ctrl down, then press C).

Files are saved to your desktop, two at a time: .md for people to read and paste straight into Word; .txt as a backup, so the captions survive even if the PC crashes.

Live captions can't tell who is speaking

That's a current Google limitation. If you need it to say "the Director said" or "the Secretary said", use feature 2 below.

Meeting in Teams? Set its speaker to Windows' default device by hand

When you choose 1 Online meeting, the program records Windows' default playback device, so the meeting app's speaker must be set to the same device. The new Teams has no option to follow the system default (tested on one laptop, 2026-09-19): go into Teams' device settings and pick the device that is currently Windows' default by hand. If you pick a different one, the whole meeting will be recorded without the other side's voices. The "聲音來源：…" ("audio source: …") line on screen at the start shows which device is being recorded.

2　Recording to transcriptmost used

The meeting is already recorded and you want text with speakers and timestamps

Press 2 → Enter
A file picker opens → find your recording, audio or video (mp3, m4a, wav, or video files such as mp4 and webm (videos recorded in a browser) all work)
It asks for proper nouns → again separated by single spaces, like "深耕計畫研發處會計室教學組長" (Higher Education Sprout Project, Office of Research and Development, Accounting Office, Teaching Section Chief)
It asks for speaker names → type them in the order they first speak, e.g. "Chair Wang, Ms Li" (separate the names with commas, so a name can contain spaces, such as a full name; if no name contains a space, spaces work too). If you don't know, just press Enter and they become 講者1 / 講者2 ("Speaker 1" / "Speaker 2"). Recordings over 28 minutes are split into segments, and the names only apply to segment 1 (see below)
Wait for it to finish. An hour of audio takes about 1–3 minutes.

When it's done it asks whether you want to open the file; press 1 and it opens automatically. The file is on your desktop and looks like this:

**[00:00:00] Director**

Right, let's begin today's Office of Research and Development meeting. First item is the interim report for the Higher Education Sprout Project.

**[00:00:09] Secretary**

Let me add something here — the deadline for clearing expenses has already been pushed back to 31 October.

A heading ends with 〔補轉〕 ("re-transcribed")?

That sentence was missed when the text was tidied up, so the program cut that short stretch out of the original audio and transcribed it again on its own to put it back (with the bundled demo file, the Secretary's line may get this). The file header also gains a line saying how many there are. The speaker and time come from the first recognition pass, so it's worth checking them against the recording. If a line can't be recovered, the header says "可能不完整" ("may be incomplete") and lists the time range.

Someone speaks English in a Chinese meeting (a visiting guest's address, say)

The English stays in the transcript as spoken. But the speaker-separation pass often fails to hear English in the middle of a Chinese meeting, so the speaker and time for that stretch may be wrong (it often gets merged into the previous speaker's paragraph). The header will say "這 N 秒的錄音裡有聲音，逐字稿卻沒有標到這段時間" ("there's sound in these N seconds of the recording, but the transcript doesn't cover that time") — listen to that part of the recording. (Tested with a home-made meeting on 2026-09-25: the old version quietly dropped the whole English passage and said nothing in the header.)

Proper nouns really do help

Tested: leave them blank and it hears 會計室 (Accounting Office) as 會計師 (accountant), and writes 深耕計畫 as 深耕計劃. Fill them in and it corrects itself automatically. A good standing list: 深耕計畫研發處教務處會計室教學組長核銷期中報告

Recordings longer than 28 minutes are split automatically

Google's official documentation states: with speaker separation and timestamps switched on, the model only processes 30 minutes of audio at a time. So the program splits long recordings into segments of 28 minutes or less, leaving a little margin.

But shorter segments don't mean nothing can go wrong. Testing with long recordings turned up two problems: a whole stretch missing from the middle of a segment (seen in 28-minute segments), or the text flooded with the same handful of sentences, over and over (seen in both 45-minute and 28-minute segments). So once it finishes, the program checks six things automatically: whether a whole stretch is missing from the middle of a segment, whether the output of the two recognition passes disagrees badly, whether the text is heavily repeated, whether something the first pass heard has gone missing from the tidied-up text (missing lines are first re-transcribed on their own and marked 〔補轉〕; only the ones that can't be recovered are listed), stretches where the recording has sound but the transcript doesn't cover that time (reported when there is sound for more than 10 seconds in a row, or at least 6 seconds of sound in total within any 30 seconds of the uncovered time; music and applause get listed too — a quick listen tells you), and whether the times of the paragraphs overlap or go backwards (when two sets of content are placed on the same stretch of time, speakers and times around there may not match). If there's a problem, the file header and the screen name the segment and the time range, and tell you which segment to cut out and run again on its own.

No warning doesn't guarantee there's no problem. Known blind spots: very short gaps (under 10 seconds in a row and under 6 seconds in total within any 30 seconds), very quiet speech (close to the background noise), and a whole stretch covered by one very long segment with a normal amount of text (a segment over 1 minute with almost no text still has its middle checked), and a recording that is noisy from start to finish, with speech hardly louder than the background (for example background music or a broadcast running throughout; when the program judges that this check can't see well, the header adds an ℹ️ line saying "This recording is almost never quiet…", which is only a note, not a gap). For important meetings, spot-check anyway.

Each segment works out its speakers on its own, so it can't tell whether a speaker in one segment is the same person as in another: from segment 2 onwards, speakers are written like "第2段講者1" ("segment 2, speaker 1"). The speaker names you type before transcribing apply to segment 1 only (the old version applied them by number to every segment, so segment 2 could end up with the wrong names). To make the names consistent, use feature 5: it shows each speaker's first line, and you can give the same person the same name in different segments.

Long transcripts made before 2026-09-16 should be re-run

The old version had none of the completeness checks described above, so when something went wrong it said nothing at all. A real case: in a 2 hour 57 minute meeting, one segment lost 14 minutes from the middle entirely, and in another about nine tenths of the text was the same handful of sentences repeating — neither produced any warning.

Re-running is the same as the first time: press 2 and pick the same recording, then check the file header for warnings. ⚠️ Note that shorter segments do not fix these two problems (28-minute segments have run into both) — the checks are what catch them, and even the checks have blind spots, so spot-check anything that matters.

3　Live bilingual captions

Following a talk in another language, or putting foreign-language captions on a Chinese talk

Press 3 → Enter
It asks where the sound comes from: 1 whatever the PC is playing (YouTube, an online seminar)/2 the microphone (a visiting speaker giving a talk in the room)/3 both (a hybrid meeting, in person + online)/4 a file you've already saved. With 2 or 3 it also asks which microphone to use, and with 2 whether there's loudspeaker sound in the room to capture (normally press 1) — the same as feature 1
It asks "fast, or accurate?"
1 Accurate (recommended): waits for the whole sentence before translating, about 2–3 seconds behind, normal word order
2 Fast: translates as it listens, about 1 second behind, but the word order occasionally jams up
Which language to translate into → 1 Traditional Chinese/2 Japanese/3 English (no need to set the speaker's language first; the program works it out by itself)

[00:02] Thank you for joining today's session on research assessment.
　　　　謝謝大家參加今天的研究評估會議。

[00:22] First, a seed funding scheme for cross-disciplinary projects.
　　　　第一，建立一個種子基金計畫，用於跨學科專案。

What's the difference between "fast" and "accurate"? Real output for the same sentence

	Fast (translates as it listens)	Accurate (waits for the sentence)
Translation	Our university has increased international collaboration rate 15% this year.	Our university's international collaboration rate rose by 15% this year.
Delay	about 1 second	about 2–3 seconds

The word-order problem in "fast" mode isn't caused by showing the text too early. The model is translating as it listens and starts before it has heard the second half of the sentence. If you want the best quality and can afford to wait, use feature 4 below.

With "whatever the PC is playing", it records every sound playing on Windows' default playback device

Close any unrelated videos or music first, or they'll be mixed in. To avoid the problem completely, use option 4 and read the audio or video file directly.

In an online meeting, the meeting app's speaker must be set to this same device too (in the new Teams you have to pick it by hand; see feature 1). The "聲音來源：…" ("audio source: …") line on screen at the start shows which device is being recorded.

Someone speaking Chinese into the microphone, translated into English? Choose "accurate"

Tested on a laptop on 2026-09-19: in this situation "fast" mode is very unreliable. Sometimes a whole passage gets no captions at all: in the tests, only 0%–19% of what was said showed up in the captions, while "accurate" mode with the same microphone got 78%–97% (very few test runs, so treat these as rough figures). Reading the built-in sample file, or capturing a news broadcast playing on the PC, works normally; but saving the microphone recording to a file and reading that is just as unreliable. Feature 6 uses "fast" mode, so the same applies there.

In "fast" mode, a line of source text only appears after a sentence-ending punctuation mark such as a full stop or question mark, a 2.5-second pause, or 90 characters. When people speak Chinese, the transcribed sentences mostly use commas, so during continuous speech the source text can appear 20 seconds or more late. That's down to how it's displayed, not the model being slow.

4　Transcript to bilingualuse this for documents overseas guests will read

Chinese minutes that an international partner needs to read, or a recording of an overseas conference that needs translating into Chinese

	Feature 3, live captions	Feature 4, this one
When to use it	Watching as you listen	Producing a document afterwards
How it translates	As it listens, like a live interpreter	The whole file together, so it can see the context
Quality	Word order occasionally odd	Much better

Press 4 → pick a file (a recording, a video, a transcript .json or subtitles .srt all work)
Choose which language to translate into
Fixed translations (strongly recommended): write them as 深耕計畫=Higher Education Sprout Project, with a semicolon ; between pairs. That keeps proper nouns consistent right through the document
What this document is about (optional): for example "a Higher Education Sprout Project meeting at the Office of Research and Development" — the model then translates more aptly
Speaker names (optional): fill them in and the finished document says "Director" instead of 講者1 ("Speaker 1"); when translating into English, speakers you haven't named are written as Speaker 1, and in Japanese as 話者1. If you pick a .json, it first prints each speaker's first line so you can match names to voices; if you already filled in names for that transcript in feature 2, just press Enter to keep them

You get four files (on the desktop, all sharing the same filename prefix):

File	What it's for
`_對照.md`	A paragraph of source, then a paragraph of translation — for people to read
`_對照表.md`	A table (time/speaker/source/translation), easy to paste into Word or Excel
`_en.md`	The clean translation-only version (into Chinese it's `_zh-TW.md`, into Japanese `_ja.md`)
`_對照.json`	Machine format, useful later if you want to rebuild the layout or translate into another language

5　Rebuild a transcriptyour lifeline when things go wrong

Runs entirely on your own PC: no network, no API call, no cost whatsoever, and it takes seconds

Every time feature 2 finishes, alongside the .md that people read it leaves a .json with the same name. That .json holds the raw data, and this feature rebuilds the .md from it.

Situation	Without this feature	With it
You transcribe a three-hour recording and it crashes halfway	Run the whole thing again, paying once more in money and time	Pick the `.partial.json` next to it and rescue the part that was already done
The speaker names went in the wrong order	Run the whole thing again	Type them again — sorted in 10 seconds
The `.md` was deleted by mistake	Run the whole thing again	Just rebuild it

It first tells you how many segments and how many speakers the file has, and it also prints the first line each speaker says, so you can work out who is who. If you just press Enter when asked for speaker names, any names you filled in back in feature 2 are kept; otherwise they're 講者1, 講者2 ("Speaker 1", "Speaker 2").

Recordings longer than 28 minutes: speakers from segment 2 on are numbered separately

Feature 2 split the recording into several segments (each 28 minutes or less) and recognised them separately. Each segment works out its own speakers, and from segment 2 onwards they're listed like "第2段講者1" ("segment 2, speaker 1"), kept apart from segment 1. The first lines it prints help you tell people apart: for the same person in different segments, type the same name.

In long transcripts made before 2026-09-25, "Speaker 1" in one segment isn't guaranteed to be the same person as "Speaker 1" in another, and there's no way to tell them apart here.

6　Live voice translationlike wearing interpreter headphones

A visiting speaker's talk, an international online meeting, a foreign-language video with no subtitles — you'd rather not keep staring at captions, and just want to hear it in Chinese

Bilingual captions still run on screen as usual (grey is the source, blue is the translation), and everything is saved when you finish. The voice is an extra, not a replacement for the captions. You can also go the other way and translate into Japanese or English (for example, listening to a Chinese talk and wanting an English voice; there's no need to set the speaker's language first; the program works it out by itself).

No 6 on your menu?

This feature only exists in the LVT edition (the edition offered for download now). If your menu stops at 5, what you have is the old original edition — download it again from the download page.

First, one thing to understand: why you choose "two routes"

When you choose whatever the PC is playing (Teams/Meet/Zoom/YouTube) as the sound source, the program records whatever Windows' default playback device is playing right now. So "I don't want to hear the original audio" can't be solved by muting — mute it, and the program can't record it either.

The only way is to send the two sounds out through two different outputs: the original audio goes to an output you can't hear (the program can still record it), and the interpreter's voice goes to the one you can hear. That way you only hear the translation, and the two sounds never overlap; the interpreter's voice also can't be recorded back in, which would leave the model translating what it has just said itself.

When it starts, the program automatically switches Windows' sound output to ①, and switches it back automatically when it ends. The program can't change the meeting app's speaker setting — see the note below

With the equipment I have, what do I pick for each route?

Your situation	① Original-audio route (you can't hear it)	② Interpreter route (you can hear it)
You want to hear the interpreter through headphones (closest to interpreter headphones)	An output that makes no sound: a digital (optical) output with nothing plugged in, HDMI to a monitor that doesn't play sound, or a virtual audio cable	Headphones (worn)
You want to hear the interpreter through speakers	Headphones (plugged in, but not worn)	Speakers
The PC has a virtual audio cable installed (for example the free VB-CABLE)	The virtual audio cable (it never makes a sound anyway, so it's the most reliable)	Speakers or headphones, either is fine
The PC has only one sound output	This mode can't be used, and the program tells you so straight away. Plugging in headphones usually gives you a second route (headphones as ①, speakers as ②, and don't wear the headphones): on 2026-09-19, wired headphones worked on one laptop; Bluetooth headphones haven't been tested yet. On some PCs the speakers and headphones are the same output, so plugging them in adds nothing. Plug them in, go back to the menu and press `6` again, and if the list has one more entry, you can use it. You can also install a virtual audio cable (see the row above), or use "microphone" or "audio/video file" as the sound source instead; those two don't have this restriction
The program says "查不到可以用的播放裝置" ("no usable playback device found")	Possibly no speakers or headphones are connected, or the sound system isn't responding at that moment. Connect them, press Enter to go back to the menu and press `6` again; if that still doesn't work, close the menu window and open it again

Not sure which one you can't hear? That's fine: once you've chosen, the program plays one sound on each route so you can listen. If you got it wrong you can choose again on the spot, rather than finding out when the meeting starts.

How to use it

Press 6 → Enter
Where the sound comes from: 1 whatever the PC is playing (Teams/Meet/Zoom/YouTube)/2 the microphone (a visiting speaker giving a talk in the room)/3 a file you've already saved. With 2, if the PC has two or more microphones, it also asks which microphone to use once you've chosen where to hear the interpreter (same as feature 1)
If you choose 1, set the two routes in turn: first choose ① which output the original audio goes to (the one you can't hear), then ② where you'll hear the interpreter (the one you can hear). The names in the list are the names in Windows' sound settings, for example "Speakers", "Headphones", "Digital Output"
Test: the program plays one sound on each route. You should not hear the first, and you should hear the second. If that's right, press 1; if not, press 2 to choose both routes again; if you didn't catch it, press 3 to play them again
If you choose 2 microphone or 3 file, it asks instead where you want to hear the interpreter: pick the headphones you're wearing (the list marks the current default). The program plays one test sound and asks whether you heard it where you wanted: press 1 for yes, or 2 to choose again. If the PC has only one usable playback device, it doesn't ask which one: the screen says in yellow that the interpreter will come out of that device, and that to listen on headphones you should plug them in first and choose 6 again from the menu; with the microphone as the source, it also asks whether to carry on (1 continue/0 back, to go and plug in headphones first). If the screen says in yellow "查不到播放裝置清單" ("couldn't get the list of playback devices"), it doesn't ask either, and the interpreter comes out of Windows' default playback device
Which language to translate into: 1 Traditional Chinese (Taiwan)/2 Japanese/3 English (no need to set the speaker's language first; the program works it out by itself)
Give this session a name (just press Enter to use today's date) → press Enter to start

This mode needs two routes. Please set each one:

① Which output should the original audio (the meeting sound) go to?
You will NOT hear this one; it's only used for recording. Pick an output with no speakers or headphones attached,
such as a digital output with nothing plugged in, or headphones that are plugged in but you won't wear.
1 Speakers
2 Headphones
3 Digital Output
0 Back
3

② Where do you want to hear the interpreter? (You DO need to hear this one)
1 Speakers
2 Headphones
0 Choose again
2

[Test] One sound on each route, to check they aren't swapped:
　① Original-audio route "Digital Output"… you should NOT hear this one
　② Interpreter route "Headphones"… you SHOULD hear this one

1 Right: didn't hear the first, heard the second
2 Wrong, choose both routes again
3 Didn't catch it, play again
0 Back

Is that right? 1

The device names on screen are only examples; on your PC you'll see the names of your own devices. The list for ② never includes the one you chose for ①, and it also leaves out any route that is actually the same playback device as ①. If that leaves nothing, the program asks you to pick a different original-audio route.

If you choose 2 microphone, the screen looks like this (with 3 file it's the same, just without the "please pick the headphones you're wearing" reminder):

Where do you want to hear the interpreter?
Please pick the headphones you're wearing. If you pick the speakers, the microphone will pick the interpreter up again,
and the model keeps translating what it said itself, over and over without stopping — and you are charged the whole time.

1 Speakers (current default)
2 Headphones
0 Back

2
　Test "Headphones"… you should hear one sound

1 Yes
2 No, choose again
0 Back

Did you hear it where you wanted? 1

Sound source	What to watch for
Whatever the PC is playing	You set two routes (see above), and they're tested once you've chosen. Close any unrelated videos or music before you start — it records every sound playing on Windows' default playback device at the time (once it starts, that's route ①)
Microphone	Please wear headphones, and choose those headphones for the interpreter in the menu (it plays one test sound once you've chosen; if the PC has only one usable playback device, it doesn't ask which one, but reminds you to plug in headphones and asks whether to carry on). If you pick the speakers, the microphone picks the interpreter up again and the model keeps translating what it said itself, over and over without stopping, while you are charged the whole time — the program can't prevent that
A file you've already saved	This problem doesn't arise. The file is read directly, so the original audio is never played at all, and all you hear is the translation anyway. It still asks where to play the interpreter (unless there's only one playback device)

Teams, Zoom and Webex speakers need setting separately — and setting back afterwards

Meeting apps have their own speaker settings, which don't necessarily follow Windows.

New Teams: before you start, always go into Teams' device settings and set the speaker to route ①. Its speaker list has no option to follow the system default, so it doesn't follow the program's switch (tested 2026-09-19). If you don't set it, the original audio never reaches ①, nothing is captured, and there's no translation.

Zoom/Webex: if the routes are set but you can still hear the original audio, go into that app's audio settings and set the speaker to route ① directly.

Google Meet and YouTube in Chrome follow the system, so you don't need to do anything.

When you finish (or stop partway), remember to set it back. The program switches Windows back to your original speakers, but it can't change the meeting app. If you set the meeting app's speaker to ①, change it back yourself to the one you normally listen on, or you won't hear anything in your next meeting.

Don't turn the volume of route ① down to 0: on some PCs, the captured volume drops along with it. In testing, 8% still translated normally, so there's no need to turn it up; this is just a precaution.

How do I stop it? Press Ctrl + C in the black window. The screen first says "收到 Ctrl+C，正在收尾" ("Ctrl+C received, wrapping up"), and it takes about 3–5 seconds to save the files before it ends. If you chose "whatever the PC is playing", the sound output switches back to your original speakers automatically: the screen says "聲音輸出已改回「…」" ("sound output switched back to …") and reminds you to switch the meeting app's speaker back too. If you had set Windows' "default communication device" separately, that is switched back as well.

If you chose "whatever the PC is playing" and it can't switch back, the screen says in yellow "沒辦法自動把聲音輸出改回原本的喇叭" ("couldn't automatically switch the sound output back to your original speakers"): click the speaker icon at the bottom right of the taskbar and pick your original speakers yourself; the program will also try again the next time you open it. If it crashes or is force-closed, it switches back automatically when you return to the menu (or next time you open the menu).

Things you'll run into

What happens	Why
The voice is 1–2 seconds behind the speaker	Simultaneous interpretation always lags a little behind the speaker
When the speaker talks too fast, the voice skips a few sentences	The program would rather skip than let your ears fall further and further behind. The captions are complete — just read the saved file afterwards; at the end, the screen tells you how many seconds were skipped
The voice flips between male and female, or sounds different after a long pause	A limitation of the model itself, which Google's official documentation also states: the voice may shift after a long pause, and it may assign the wrong gender based on how the speech starts
The voice may cut off mid-sentence every 8 minutes	The program switches to a new connection every 8 minutes; the captions aren't affected
When a speaker talks Chinese into the microphone and you translate into English, it's very unreliable, and sometimes a whole passage doesn't get translated	It's the same model as feature 3's "fast" mode — see feature 3. In this situation, switching to feature 3, choosing "accurate" and reading the captions is more reliable

Feature 6 always uses "fast" mode (Google's simultaneous-interpretation model), because only that mode has a voice to play. It costs about NT$71 an hour — see Cost.

What you see	What it means	What to do
有收到聲音 (sound received)	Loud enough to be recognised	Nothing
聲音偏小 (sound is weak)	Captions may miss words or be inaccurate	Microphone: move it closer to the speaker, or use another microphone; PC audio: turn the playback volume up (see below)
幾乎沒有聲音 (almost no sound)	There won't be any captions	Check the microphone, and whether the headset plug has worked loose
電腦現在幾乎沒有在播 (the PC is hardly playing anything)	The PC isn't playing sound (the online meeting isn't on, or the sound is going to another speaker)	Make sure the online meeting's sound is playing from this PC
沒有收到資料 (no data received)	The device may have been unplugged	Plug it back in; the program reconnects automatically

Four things to check	What to do if it fails
Are the people in the room being recognised	Move the microphone towards whoever is speaking, or switch to a USB conference microphone; don't use a headset lying on the table
Are the online participants being recognised	Switch to option 1 (the caption PC joins the meeting too and you choose "Both"). Turning up the loudspeakers doesn't help
Does the same sentence appear twice	That's echo — use headphones on the caption PC, or turn its loudspeaker down
Is the caption text big enough	In the caption window, press `Ctrl` + `+` to enlarge it (you can press it repeatedly) and `Ctrl` + `-` to shrink it. As a rule of thumb, the character height in millimetres should be about 9 × the distance in metres to the furthest seat

Feature	Model	Input price	Output price	Per hour
① Live meeting captions	gemini-3.5-transcribe-live	$3.50/1M ($0.005/min)	$21.00/1M ($0.004/min · text)	NT$17
② Recording to transcript	gemini-3.5-transcribe (two passes) + gemini-3.5-flash (alignment)	$2.00/1M ($0.003/min); alignment $1.50/1M	$12.00/1M ($0.002/min · text); alignment $9.00/1M	~NT$20
③ Live bilingual · accurate	gemini-3.5-transcribe-live + gemini-3.5-flash-lite	①＋$0.30/1M	①＋$2.50/1M	~NT$18
③ Live bilingual · fast	gemini-3.5-live-translate-preview	$3.50/1M ($0.0053/min)	$21.00/1M ($0.0315/min · audio)	NT$71
④ Transcript to bilingual doc	gemini-3.5-flash	$1.50/1M	$9.00/1M (including thinking tokens)	~NT$16 (per hour of transcript, measured)
⑤ Rebuild transcript layout	no model call	—	—	NT$0
⑥ Live voice translation	gemini-3.5-live-translate-preview	same as ③ fast	same as ③ fast	NT$71

	Free tier	Paid tier
Cost	No charge	You pay for what you use
Audio used for training	Yes	No
Human reviewers may read it	Possibly	No

Google engine	Feature	Coverage
transcribe-live	Feature 1 live meeting captions	98.1%
transcribe-live	Feature 3 "accurate", source text	97.9%
transcribe	Feature 2 recording to transcript	99.9%
live-translate-preview	Feature 3 "fast", source text (feature 6 interpretation uses the same model; not measured separately)	98.4%
flash	Feature 4 transcript to bilingual (translation)	100%
flash-lite	Feature 3 "accurate" (translation)	100%

Turn what is said in a meeting into text you can edit

Quick start: three things

This is the screen you'll see when it opens

Installing

What your PC needs

Installation steps

Blocked by Windows: "Unblock" it first, then install

Not unzipped yet (the cleanest way)

Already unzipped (the drawing below shows this case)

Getting a key and turning on billing

Three things people trip over

The six features

1　Live meeting captions

2　Recording to transcriptmost used

Recordings longer than 28 minutes are split automatically

3　Live bilingual captions

What's the difference between "fast" and "accurate"? Real output for the same sentence

4　Transcript to bilingualuse this for documents overseas guests will read

5　Rebuild a transcriptyour lifeline when things go wrong

6　Live voice translationlike wearing interpreter headphones

First, one thing to understand: why you choose "two routes"

With the equipment I have, what do I pick for each route?

How to use it

Things you'll run into

Setting up a hybrid meeting (in person + online)

Option 1: the caption PC joins the online meeting too, and you choose "Both"recommended

Option 2: run a cable from the room's sound system (the most reliable, but you need the equipment)

Option 3: all you can do is mic the loudspeakers (a fallback, no guarantees)

Which microphone for the people in the room

Reading the "🎚 收音檢查" (sound check) on screen

What it costs

Which model each feature uses, and what it costs

Things to know before you use it

What to do when something goes wrong

Don't want it any more? How to remove it

Ticked for you by default

Not ticked by default (other programs may use them too)

Limitation	How to work around it
Live captions can't tell speakers apart	When you need speakers labelled, use feature 2
On feature 3 with "fast", or on feature 6, translating Chinese into another language, the source text is occasionally a whole block of Simplified Chinese (roughly once in three sessions; the translation is unaffected)	Confirmed unavoidable over 29 tested sessions — this is model-side behaviour. If you need that source text, paste the `.txt` into any Simplified-to-Traditional converter, or use feature 3's "accurate" mode (feature 6 has no "accurate" mode to switch to)
On feature 3 with "accurate" and "translate into Traditional Chinese", when the speaker is speaking Chinese, the source text comes out in Simplified Chinese (for example people speaking Chinese in the room at a hybrid meeting; seen once in testing, 2026-09-19)	The translation is still in Traditional Chinese; it just amounts to rewriting the Chinese. If you need that source text, paste the `.txt` into a Simplified-to-Traditional converter
On feature 3 with "fast", or on feature 6, it's very unreliable when someone speaks Chinese into the microphone and you translate into English (the same passage sometimes goes completely untranslated; tested on a laptop, 2026-09-19)	Switching to feature 3, choosing "accurate" and reading the captions is more reliable
Recordings over 28 minutes are split, and it can't tell whether a speaker in one segment is the same person as in another	From segment 2 onwards, speakers are labelled like 第2段講者1 ("segment 2, speaker 1"), and the names you type before transcribing apply to segment 1 only. Use feature 5: look at each speaker's first line and give the same person the same name
English spoken in a Chinese meeting may get the wrong speaker and time (it often gets merged into the previous speaker's paragraph)	The English is kept as spoken, and the header lists the stretches where "the recording has sound but the transcript doesn't cover that time" — listen to those parts of the recording
Feature 6 needs two sound outputs when it captures "whatever the PC is playing"	Plug in headphones — this usually adds a second output (wired headphones worked on one laptop; Bluetooth hasn't been tested yet) — or install a virtual audio cable; or use the microphone or an audio/video file as the sound source instead
Feature 6's voice occasionally flips between male and female, and skips a few sentences when the speaker talks too fast	The male/female flipping is a limitation of Google's model itself; the skipping is deliberate on the program's part, so that your ears don't fall further and further behind. The captions are complete — for anything official, go by the saved file
Windows only	Capturing "whatever the PC is playing" uses a Windows-only technology, so Mac and Linux can't do it

What happens	What to do
You click it, a window flashes up and disappears, then the menu appears	That's normal. The program opens a separate window for the menu, and its own window closes straight after
After you press Enter the screen sits still for tens of seconds, as if it has frozen	It hasn't frozen: it's loading the speech components, which have to be read back in from disk if they haven't been used for a while. The screen shows the progress item by item (for example `· 語音處理元件（scipy）… ✓ 27.2 秒`, "speech-processing component (scipy) … ✓ 27.2 s"); as long as you can see that, all is well. Warming it up before the meeting avoids the wait
Straight after installing, features 1/3/6 say "載不起語音處理元件（scipy）" ("couldn't load the speech-processing component (scipy)")	Windows 11's Smart App Control first blocks newly installed components that don't have a cloud reputation yet (seen on one laptop, 2026-09-20). Wait a few minutes and try again — it usually lets them through on its own. To check, double-click `6_診斷.bat` and see whether those lines say `BLOCKED` or `OK`; use it once they're back to `OK`. Feature 5 works offline and doesn't need those components, so you can still use it meanwhile
No captions appear at all	Check the PC is actually playing sound; for an online meeting, remember to choose `1` and not `2`. The meeting app's speaker must be set to Windows' default playback device: the new Teams has no option to follow the system default, so pick the same device by hand in Teams' device settings (tested on one laptop, 2026-09-19). The "聲音來源：…" ("audio source: …") line on screen at the start shows which device is being recorded. Feature 6 is the exception: when you choose "whatever the PC is playing", set the meeting app's speaker to route ①, the one you chose in the menu, not the usual Windows default
You choose the microphone and the program says "沒有可用的聲音來源" ("no usable audio source") and exits	This PC has no microphone attached, or Windows privacy settings are blocking it (Settings → Privacy & security → Microphone)
You choose "microphone" with headphones plugged in: which microphone does it pick up?	On some PCs, plugging in headphones that have a microphone (a headset) automatically switches to the microphone on the headphones (for example laptops with a single combined headset jack), and it only switches back to the PC's microphone when you unplug them (tested on one laptop, 2026-09-19). If the headphones are plugged in but not worn and you choose "microphone", it picks up the headset mic hanging to one side: unplug the headphones, or switch back to the PC's microphone under Settings → System → Sound → Input. The screen shows which one it's using at the start (the "麥克風：…" ("microphone: …") line). If the PC has two or more microphones, the menu asks which microphone to use, and you can simply pick the PC's built-in one there
The transcript header says "可能不完整" ("may be incomplete")	There are six causes: ① nothing recognised in the later part of the recording (silence, music, noise) ② the same words repeated heavily, so they were dropped when the transcript was tidied up ③ a problem in one segment: a whole stretch in the later part or the middle of that segment not recognised, the two recognition passes not agreeing, or the text flooded with the same handful of sentences, over and over ④ a line the first pass heard is missing from the tidied-up text, and re-transcribing it on its own didn't bring it back (the header gives the time and the speaker) ⑤ the recording has sound, but the transcript doesn't cover that time (over 10 seconds in a row, or at least 6 seconds in total within any 30 seconds; typically English spoken in a Chinese meeting, music or applause — the header gives the time range; when the recording is noisy from start to finish and speech is hardly louder than the background, this check can't see well, and the header adds an ℹ️ note that does not count as a gap) ⑥ the times of the paragraphs overlap or go backwards (typical of very long recordings that mix several languages; the header gives the time range — listen to that stretch against the recording). For ①, ② and "nothing in the later part of a segment", the header says which minute recognition got up to; otherwise, it names the segment and the time range where the problem is. If you need that part, cut it out and run it again on its own (for ②, ③ and ④ the header also suggests how to cut it). If it was the later part or the middle of a segment that wasn't recognised and the cut-out part comes back with content, it had been skipped; if nobody was really speaking, the result will be empty. Passages marked as repeated were not really said that many times, so don't quote them word for word.
You closed the window partway through a transcription	The program deletes the temporary recording and the cloud copy before exiting. For recordings over 28 minutes, the segments already finished stay in `.partial.json` — rescue them with feature 5. Under 28 minutes it's processed as one segment, so closing it while recognition is still running means the whole thing has to be run again. If something goes wrong only after that segment has fully finished (a failure writing the file, say), the `.partial.json` is still left, so have a look on the desktop for it first
The network drops partway through transcribing or translating	When the connection is completely down (a Wi-Fi outage, say), the program waits for the network to come back: it tries again every 20 seconds, for up to 5 minutes, and carries on as soon as it returns (the screen says "The internet connection is back"); if the network is merely unstable and drops for a moment, it just retries a few times as before. If it is still down after 5 minutes: transcribing stops — for a recording over 28 minutes, the parts already fully transcribed still become a transcript (the header lists the missing parts); a recording of 28 minutes or less (or one that stops in its first part) produces nothing, so transcribe the whole thing again once the network is working. The exceptions: if it drops at the speaker-segmentation step, the program switches to an estimate and still finishes; if it drops while re-transcribing missed passages, those are skipped and listed in the header. Translating stops, and what has been translated is still saved, with the untranslated paragraphs marked (the same happens if you press Ctrl+C while it is waiting; if nothing has been translated yet, Ctrl+C leaves no file). When the input to translate is a recording, the transcribing and translating steps can each wait up to 5 minutes
The same sentence appears twice	Echo in a hybrid meeting: the meeting PC's speakers and microphone are picking each other up. Switch to headphones
The people in the room are captured, but not the online participants	The online participants have to be taken straight from the PC: have the caption PC join the meeting too and choose "Both" (see option 1 under hybrid meetings). Turning up the loudspeakers doesn't help
The caption text is too small	In the caption window, press `Ctrl` + `+` to enlarge it (you can press it repeatedly) and `Ctrl` + `-` to shrink it; there's a reminder of this on screen at the start as well. `Ctrl` + `0` is the official "reset size" key, but it clashes with the 自然輸入法 (Going Natural) Chinese input method, so we don't recommend it
The captions suddenly stop (the headset plug came loose, or it was unplugged and plugged back in)	The program reconnects automatically and shows "✓ 麥克風接回來了" ("✓ microphone reconnected") or "✓ 電腦聲音接回來了" ("✓ computer audio reconnected") (or "✓ 電腦聲音改從這顆錄：…" ("✓ computer audio is now recorded from: …") when it switches to another device); if that never appears, press `Ctrl`+`C` to stop and start again (all the captions so far are saved)
Chinese shows up as boxes or garbled characters	Open it with "Windows Terminal" instead (search for Terminal in the taskbar); the older Command Prompt handles Chinese fonts poorly
A passage in the transcript just says "（整理後的逐字稿略去了這段）" ("this passage was left out of the tidied-up transcript")	Usually that speaker said nothing but "mm" and "right" all meeting. The line is kept so that the speaker names you typed don't all shift out of line
The folder won't delete, saying it needs administrator rights	It's almost never a rights problem. ① Close any File Explorer window that has this folder open ② restart and delete it before anything else ③ only then double-click `8_刪不掉時救援.bat`
The key was pasted wrong, has expired, or you want to change it	Press `9`, "API key settings", in the menu: you can switch to another key, check whether the current one works, or clear the key saved on this PC
The screen says "API 金鑰有問題" ("there's a problem with the API key")	For live captions, first press `Ctrl`+`C` to stop, then go back to the menu and press `9` to check the key or change it. When a transcription or translation doesn't succeed, the menu checks with Google once by itself, and if that confirms the key is bad, it asks you straight away whether you want to change it
The screen says "Google 專案這個月的花費上限到了" ("the Google project has reached this month's spend cap")	It's not a network problem: this Google project has spent up to the monthly cap set in AI Studio, and Google pauses it until the 1st of next month. To restore it now: at `aistudio.google.com`, open Spend on the left → Monthly spend cap → Edit spend cap and raise the cap (this needs owner or editor access to the project; if it still can't connect after you raise it, wait a few minutes and try again). Live captions reconnect by themselves; a transcription or translation that stopped halfway has to be run again
Feature 6 is running, but there's no interpreter voice at all from the headphones or speakers	First check that the line at the top of the screen, "🔊 翻譯語音將從「⋯」播出" ("🔊 the translated voice will play from '…'"), names the device you're actually wearing or listening to. If it doesn't, go back to the menu, run feature 6 again and pick the right device at "where do you want to hear the interpreter"; if it does, turn that device's volume up, or check in the speaker icon at the bottom right of Windows that it isn't muted. At the end, the screen says "共送出約 N 秒到「⋯」" ("sent about N seconds to '…'") — that's how much was sent to the device, which doesn't guarantee you could hear it, so a number there doesn't prove the device is fine
Feature 6 with the microphone: after the speaker stops, the captions keep repeating the same sentence without stopping	The interpreter voice is coming out of the speakers and the microphone is recording it back in, so the model keeps translating its own voice — and you are charged the whole time. Press `Ctrl`+`C` to stop first; then put on headphones, pick the headphones you're wearing at "where do you want to hear the interpreter", and start again
No sound from the speakers after using feature 6	Finishing normally, pressing `Ctrl`+`C` and clicking the window's X all switch back to your original speakers automatically. If it crashes or is closed from Task Manager, it switches back automatically when you return to the menu (or next time you open the menu); if you can't wait, click the speaker icon at the bottom right and pick your original speakers yourself. If you set the meeting app's (Teams/Zoom/Webex) speaker to ① as described, the program can't change that — switch it back yourself
Feature 6 is set up, but you can still hear the original audio	The meeting app (Teams/Zoom/Webex) has its own speaker setting. Go into its audio settings and set the speaker to route ① directly; the new Teams has no option to follow the system default, so set it before you start. Remember to switch it back to the one you normally listen on afterwards
Feature 6 (or feature 3 with "fast") keeps running after you press Ctrl+C	It first shows "收到 Ctrl+C，正在收尾" ("Ctrl+C received, wrapping up"), and takes about 3–5 seconds to save the files before it ends. You can also end it at that point with the X in the top-right corner of the window, but the tidied-up `.md` won't be produced (the word-for-word `.txt` is still there), and the menu closes as well; feature 6's sound output still switches back automatically
You want to know what's actually wrong	Double-click `6_診斷.bat` and it lists whether Python, the packages, ffmpeg, the key and the files are all there, and it actually finds the microphone and computer-sound recording devices and opens each one for a moment (no sound is kept). Just screenshot it and send it in

Quick start: three things

This is the screen you'll see when it opens

Installing

What your PC needs

Installation steps

Blocked by Windows: "Unblock" it first, then install

Not unzipped yet (the cleanest way)

Already unzipped (the drawing below shows this case)

Getting a key and turning on billing

Three things people trip over

The six features

1 Live meeting captions

2 Recording to transcriptmost used

Recordings longer than 28 minutes are split automatically

3 Live bilingual captions

What's the difference between "fast" and "accurate"? Real output for the same sentence

4 Transcript to bilingualuse this for documents overseas guests will read

5 Rebuild a transcriptyour lifeline when things go wrong

6 Live voice translationlike wearing interpreter headphones

First, one thing to understand: why you choose "two routes"

With the equipment I have, what do I pick for each route?

How to use it

Things you'll run into

Setting up a hybrid meeting (in person + online)

Option 1: the caption PC joins the online meeting too, and you choose "Both"recommended

Option 2: run a cable from the room's sound system (the most reliable, but you need the equipment)

Option 3: all you can do is mic the loudspeakers (a fallback, no guarantees)

Which microphone for the people in the room

Reading the "🎚 收音檢查" (sound check) on screen

What it costs

Which model each feature uses, and what it costs

Things to know before you use it

What to do when something goes wrong

Don't want it any more? How to remove it

Ticked for you by default

Not ticked by default (other programs may use them too)

1　Live meeting captions

2　Recording to transcriptmost used

3　Live bilingual captions

4　Transcript to bilingualuse this for documents overseas guests will read

5　Rebuild a transcriptyour lifeline when things go wrong

6　Live voice translationlike wearing interpreter headphones