สร้างแอป Local AI แปลงไฟล์ PDF เป็น Excel ภาษาไทยไม่เพี้ยน! — Transcript
Full transcript
- 0:00So, here, OCR only takes about 200
- 0:03milliseconds for 38 lines.
- 0:07You can save it to SQL or export it to
- 0:09Excel.
- 0:14Wow, this program is amazing! Hello
- 0:16everyone, this is Broom.
- 0:19Today I'm showing you how to create a
- 0:21Local AI harness for accounting.
- 0:26We can receive PDF files, scanned bills
- 0:29/receipts, and photos—all within this
- 0:31harness using GoLanguage from Google.
- 0:39Today, the big picture involves 5 steps
- 0:41: receiving files, preparing files,
- 0:44reading documents, and this is the
- 0:46highlight: the OCR model in
- 0:48ThailandCommunity has just been updated
- 0:51, and it's a great model, like Typhoon
- 0:54OCR or OpenThai 2.0, the hottest OCR
- 0:56models right now.
- 1:01So, today I'll show you how to install
- 1:03it using HuggingFace.
- 1:08I'll use one first to test the flow here
- 1:10.
- 1:12We'll also organize the program's table
- 1:14, columns, and check the formatting,
- 1:17because when we scan hundreds of files,
- 1:19bills, and receipts, we need to organize
- 1:21the table.
- 1:24Then we can export it to Excel.
- 1:27It can be a CSV file or a local SQLite
- 1:30file.
- 1:31This is another step in the backup
- 1:33process.
- 1:35Here, we'll deal with data hashing,
- 1:38cipher encryption, and chart
- 1:39verification of the files.
- 1:44This system helps make the harness
- 1:52or equipment used by each organization
- 1:54and business more secure.
- 1:59Okay, I'll open the slides and explain
- 2:02step by step what a local AI harness is
- 2:04.
- 2:10A harness allows us to change any model
- 2:12because a harness is like everyone
- 2:14building a basic engine structure.
- 2:20For example, if I build a car or a robot
- 2:23, I can change the engine, the brain
- 2:25inside, whether it's Typhoon, OpenThai,
- 2:28or various models like Google Gemini,
- 2:31GPT, etc.
- 2:39I'll show you how to use the antigravity
- 2:41model because it's a brand new model,
- 2:44Gemini 4.0, which includes carbon and
- 2:46hydrogen.
- 2:50I'll be using these two together.
- 2:53This one, hydrogen, is an orchestator.
- 2:59Actually, an orchestator is, I'll
- 3:01explain briefly, it's used to structure
- 3:04and distribute tasks to AI.
- 3:07It's called a sub-agent or agent team.
- 3:12It can be used, but I recommend using it
- 3:14for large projects, especially if you're
- 3:16running it in parallel.
- 3:22As I said, carbon and hydrogen really
- 3:24work well together.
- 3:28And there's a free quota from Google too
- 3:30, so you can test it out.
- 3:32I'll provide the download link here.
- 3:35You can download Google Antigravity
- 3:37according to your OS.
- 3:41Once finished, we'll get a page like
- 3:43this.
- 3:45Okay, let's start the steps.
- 3:48We'll use Google Antigravity, either the
- 3:51Go language or Python.
- 3:54Then, we'll follow these 5 steps.
- 3:59I'll keep it simple for everyone.
- 4:02You can copy the files in this slide or
- 4:05like this.
- 4:08I'll capture this and use it as a
- 4:10project folder for this application.
- 4:17Then, click on Google Antigravity, and
- 4:19then click the plus sign (the one I use
- 4:21when creating any application) and
- 4:23select "from this folder."
- 4:30On my computer, I'll create it on the
- 4:32desktop and name it "bin."
- 4:38Okay, so we're using billingPDF, text,
- 4:42OCR, and local AI harness.
- 4:50It might be a bit long, but we can
- 4:51shorten it to this.
- 4:55Okay, then I'll click create and open,
- 4:58and then click next.
- 5:02Now we'll be in our local project folder
- 5:04on our desktop.
- 5:08After that, this will be the worktree.
- 5:11Let me explain it simply.
- 5:14It's not complicated at all.
- 5:19We already have the main code, and if we
- 5:21want to open the worktree without
- 5:23affecting the main code, we open a
- 5:25branch, do the work, and then merge it
- 5:27into the main branch.
- 5:31What's the benefit of this?
- 5:33It's beneficial...It's very important
- 5:36when we run parallel applications,
- 5:38creating multiple features
- 5:39simultaneously or working on separate
- 5:42tasks—meaning running them in parallel—
- 5:44using orchestrators or agents.
- 5:48This relates to worktrees.
- 5:52But today, for the fastest applications
- 5:55we can start with, I recommend avoiding
- 5:57branching into lines, placing
- 5:59orchestrators, and not placing agent
- 6:01teams yet.
- 6:05I strongly advise against this, based on
- 6:07research, as it reduces performance
- 6:09unless the task is complex or requires
- 6:11running multiple features concurrently.
- 6:17Okay, I'll click the plus button to
- 6:19upload the file.
- 6:22This is my screen sharing right here.
- 6:25This is the application file, everyone.
- 6:29This one, and after that, we can
- 6:31actually
- 6:36use it, or we can use it first.
- 6:42I'll let you know first whether this
- 6:43idea is good or not, and whether it will
- 6:45really meet the needs.
- 6:51Then, everyone, I'm going to adjust it
- 6:53to turbo because it focuses on speed,
- 6:56and the data is relatively fine, so we
- 6:58can approve it.
- 7:04Turbo, okay?
- 7:07Let's
- 7:10talk
- 7:13about
- 7:16the
- 7:20all-website,
- 7:24or the structure.
- 7:32I'll tell you whether this structure
- 7:34meets the needs of our organization, our
- 7:36system.
- 7:41For example, we use Windows or Not,
- 7:43something
- 7:49like that for now.
- 7:52Then, I'll use hydrogen first, and use
- 7:55it because there are four hydrogens, it
- 7:57won't consume as much as carbon.
- 8:02Carbon is very smart, but during this
- 8:04phase, or asking about our idea, I
- 8:06recommend using something that isn't too
- 8:08focused on coding, and that you already
- 8:11have a reasonable amount of knowledge
- 8:13about, it will meet the needs better.
- 8:19Then, while they're guild-ing, I'll read
- 8:21this first because sometimes they'll get
- 8:23a pop-up like this.
- 8:33Okay,
- 8:38okay.
- 8:40And here, actually, if we use it for
- 8:42multi-users and share an office PC, I
- 8:44would highly recommend this because it's
- 8:47for real work, for work used in a real
- 8:49organization.
- 8:53And, or, this one...For running the
- 8:56project, if I were to recommend one, I'd
- 8:58suggest option two.
- 9:02However, for maximum convenience right
- 9:05now, I'd recommend option one.
- 9:08Everyone, if you want to delve deeper
- 9:11into the full-fledged option two for
- 9:13organizational use, you can contact me
- 9:15directly here to discuss and ask
- 9:17questions.
- 9:21You can also consult with me about the
- 9:23community
- 9:31version, which includes all the source
- 9:33code and class materials
- 9:41that you can download and use directly.
- 9:48A new version will be available soon;
- 9:50you can join via the link below.
- 9:56If everyone wants to consult me
- 9:58initially, they can do so normally.
- 10:03Okay, so we'll look at this application
- 10:05.
- 10:07It has five different system components
- 10:09, and I'll explain further.
- 10:17As I said, it's like a car or a robot
- 10:19with a complete frame that can run just
- 10:21by adding an engine and a brain.
- 10:28How is this different from other
- 10:30software applications?
- 10:33It can think, has memory, and can handle
- 10:35all kinds of devices.
- 10:39For example,
- 10:44it
- 10:52can create programs for accounting and
- 10:55other tasks.
- 11:04This is why it's called...well, as I
- 11:06explained, it can read files, prepare
- 11:08images, and has all the necessary
- 11:10equipment for reading bills and
- 11:12documents, preparing them for export,
- 11:14and connecting models.
- 11:27This video or this application can do
- 11:29various things: it can take photos,
- 11:31convert image files, correct tilting,
- 11:33call models, and then check the table
- 11:35arrangement to see if the data is
- 11:37correct according to the columns and
- 11:39columns.
- 11:46We check and save the data locally or
- 11:48elsewhere so we can use it later.
- 11:54It's important to be mindful of our
- 11:56personal data and not forget security.
- 11:58This is crucial because otherwise, our
- 12:00financial information could be leaked.
- 12:05Another important reason why we should
- 12:07be careful about our financial and
- 12:09accounting data is that this information
- 12:11can be sent to other servers and
- 12:17processed by various companies.
- 12:22AI doesn't just think about what a
- 12:24company's bill is; it can analyze how
- 12:26much a business typically spends on that
- 12:28type of expense.
- 12:32Our data could be publicly disclosed.
- 12:35So, I want everyone to be aware of local
- 12:38AI and use it wisely, without wasting
- 12:40tokens, and to respect privacy.
- 12:43I'll discuss this further in the next
- 12:45section. Okay, let's move on. I'll
- 12:47explain the importance of local AI with
- 12:49privacy.
- 12:52We can fine-tune models without paying
- 12:55for expensive subscription courses, GPT
- 12:57chat, or Gemini.
- 13:01We can take those and build a structure
- 13:03, then create our own harness and local
- 13:05AI for use.
- 13:11If you want to test this individually or
- 13:13discuss it further, you can join the
- 13:15community.
- 13:17But if you want my company or my team to
- 13:19help set up this system, you can contact
- 13:21us.
- 13:24I also have harness systems for
- 13:26organizations.
- 13:31Okay, so regarding the model size, I'll
- 13:33elaborate on the model size we can...
- 13:36The OCR functionality can be changed
- 13:38here.
- 13:41In the accounting field, there are two
- 13:44latest models: the SCB Typhoon and
- 13:46OpenThai 2.0.
- 13:49Both are excellent and developed by
- 13:51incredibly talented researchers in
- 13:53Thailand.
- 13:54Don't forget to like, share, and add
- 13:56them on Facebook.
- 13:59I'll provide the source code and give
- 14:01credit to the development team.
- 14:050.8 B refers to the model size of this
- 14:08AI model from SCB Typhoon/OpenThai. It's
- 14:11lightweight and fast.
- 14:130.8 B refers to the parameter size; this
- 14:16one has 800 million parameters, and this
- 14:18one has 2 billion parameters.
- 14:23In simple computer terms, this means the
- 14:26model size is approximately 1.3 GB
- 14:29versus 3.2 GB.
- 14:35The model's clarity is 8-bit, and if
- 14:37you're using a Mac, I recommend 8GB of
- 14:40RAM or more (16GB or more if you're
- 14:43using a 2-bit system) and a graphics
- 14:46card with 4GB of VRAM or more.
- 14:49There's a formula for calculating VRAM,
- 14:52it might not be perfectly precise, but
- 14:54it helps a lot.
- 14:56The safest approach is to have plenty of
- 14:59VRAM and plenty of RAM.
- 15:03Macs use processors that
- 15:09can integrate all their GPU cores and
- 15:11RAM.
- 15:14PCs without dedicated graphics cards now
- 15:16have a similar structure to Macs.
- 15:20Newer Spark notebooks are starting to
- 15:23meet these needs, so you don't have to
- 15:25worry; they'll be similar to Macs.
- 15:29Okay, so...There's also a rating system
- 15:32for how the Thai language scans perform
- 15:34, and this will be ours.
- 15:39Let's continue.
- 15:41Actually, everyone needs to install it
- 15:43first, or you can use Python.
- 15:48It's not a problem at all, but it's
- 15:50lighter in terms of running speed and
- 15:52application size.
- 15:56It will run more smoothly.
- 15:58Actually, it's very simple.
- 16:02Everyone can take my slides, the
- 16:04commands below, the slides that are
- 16:06already in the application, and paste
- 16:08them like this.
- 16:12It's like the
- 16:15AI is our assistant. Sometimes we
- 16:17understand the management of commands,
- 16:19understand the structure, and know which
- 16:21app to use; we just let it handle the
- 16:23rest. Yes, it can be done.
- 16:27For UI, I'll use this as a
- 16:28recommendation here, but I need to
- 16:31mention that right now I'm using macOS.
- 16:35Also, please use the "give me version"
- 16:38in macOS.
- 16:40Actually, you don't need to include it;
- 16:42I'll remove it.
- 16:46I didn't continue this because I wanted
- 16:48to say that we're currently using macOS
- 16:50.
- 16:52Okay, everyone, don't forget to install
- 16:55the Go version. If you want to use it
- 16:57like me, copy the command here in the
- 16:59slide and paste it.
- 17:02The next part, the model runner, is
- 17:05quite important: LMStudio or lama.cpp.
- 17:10Actually, if everyone is using an NVIDIA
- 17:13graphics card or that architecture, you
- 17:16can use VLLM.
- 17:18Here's VLLM.
- 17:22Everyone, it will be...It's about
- 17:25running AI models smoothly, fast, easy,
- 17:27fast, and cost-efficient for NVIDIA
- 17:30architecture.
- 17:33For example, CUDA, everyone, CUDA is
- 17:36from NVIDIA.
- 17:39CPUs also have it, but for AI to really
- 17:41run, it's better to run it on CUDA on
- 17:44the GPU, also from NVIDIA.
- 17:47CUDA, everyone.
- 17:49Sorry if I mispronounced it.
- 17:51Okay, then load Typhoon OCR, choose the
- 17:54model size, or use OpenType.
- 17:58For example, I have 48GB of RAM on my
- 18:01MacBook, so I'll use a 2 billion model.
- 18:07Let me check which one to load first.
- 18:11I actually want to load OpenType too.
- 18:14Let me look for a moment. Okay, I'll use
- 18:17OpenType. This one.
- 18:19And then...So, we'll have Olama pull it
- 18:22down.
- 18:24We need to check if Olama has it or not
- 18:26.
- 18:27For example, with Typhoon, they've
- 18:29deployed it, and it's great! Everyone
- 18:31has it in Olama.
- 18:33This part is also in Olama, so it's easy
- 18:36and convenient.
- 18:38You can download it here.
- 18:40The command is in the slide as well;
- 18:42just copy it from here.
- 18:44Everyone, finish the first line, then
- 18:46check and type the second line in the
- 18:48terminal like this.
- 18:51I'll show you the terminal window. It
- 18:53will look like this. Everyone, the
- 18:55terminal and
- 18:58Olama are okay.
- 19:01No problem, I'll use this one first
- 19:03because it's very easy for Olama.
- 19:07Olama isn't installed yet; you can
- 19:09install it directly through the download
- 19:11here. It's compatible with MacOS, Linux
- 19:14, and Windows. Here are the download
- 19:17links, everyone.
- 19:18Place them here, then I'll install the
- 19:21Olama file first.
- 19:24Okay, then check here for any issues.
- 19:26Follow their instructions, they
- 19:28recommend it, and then submit.
- 19:33Next, type the Olama Pool OCR model here
- 19:36for testing.
- 19:40My computer has a lot of RAM, but
- 19:42version 0.8 is fine for now, it should
- 19:45be easy.
- 19:46Everyone can run it too.
- 19:48But if you have a powerful computer, use
- 19:51this link, everyone. This link is for
- 19:54version 2.0.
- 19:56I'll check with the Thai open forum
- 19:58first to see if it's been deployed yet.
- 20:01There might be, everyone seems to know.
- 20:06I'll check with the Thai open forum,
- 20:08Professor Kae.
- 20:12This is really important, especially the
- 20:14popular and newest parts. Oh, they're
- 20:16not here yet, everyone. That's okay,
- 20:19it's not too serious.
- 20:21We'll consider this as a use case
- 20:23example that everyone can adapt
- 20:24themselves.
- 20:27You can download it from here or
- 20:29directly from [link].
- 20:32I'll skip this part for now.
- 20:35This
- 20:38model is approved.
- 20:41Okay, I'll finish step 1 first, then
- 20:43I'll go to the model above.
- 20:47It's simple, everyone.
- 20:51For this application, use
- 21:00this model.
- 21:02Answer their questions, and that's it.
- 21:05Oh, this is still in the process.
- 21:09I'll send the model for them to run.
- 21:12Wow, everyone, it's amazing!
- 21:15Actually, regarding the 4-digit code, I
- 21:18use all of them.
- 21:25Everyone, both the Google code and the
- 21:27model from the company that's always
- 21:29popular, but I'm really starting to like
- 21:31Google because you get everything from
- 21:33Google itself, like YouTube Premium.
- 21:41It's a huge ecosystem.
- 21:44If you subscribe to Google's code, and
- 21:49maybe a little bit from other things,
- 21:51maybe get a discount or something.
- 21:53This is my plan.
- 21:56That's about it.
- 21:59I'll look at the next steps. Preparing
- 22:01the files is easy. I've already given
- 22:03them an overview of how we receive the
- 22:05files
- 22:08and verify them.
- 22:10They'll write code to check this, and
- 22:12you can let them help build the
- 22:14components.
- 22:15Then add the necessary parts.
- 22:18The data files are organized to prevent
- 22:20duplication, using generated numbers.
- 22:25We're simply comparing the file
- 22:27codenames.
- 22:28Once that's done, we prepare for
- 22:30preprocessing.
- 22:32This means preparing the data for
- 22:34processing before starting the
- 22:36preprocessing process.
- 22:38This is called preprocessing–converting
- 22:41a PDF to an image, correcting image tilt
- 22:44, enlarging, sharpening, etc. After
- 22:46preprocessing, we call the model, send
- 22:49the image to the AI model to read, and
- 22:51then wait for the results.
- 22:55This is for the model to start reading;
- 22:57any model will do.
- 23:01And here, the table checking part is
- 23:03where we design the application–whether
- 23:06we want to check the table format, check
- 23:08the totals, etc. Wherever you want to
- 23:10highlight something in yellow or put a
- 23:12flag, this is how we design our harness
- 23:15.
- 23:17Actually, everyone on GitHub, especially
- 23:20those in the BoonBigNotePlus
- 23:24local group, can access it through the
- 23:26BoonBigNotePlus GitHub repository.
- 23:30For those in the community group
- 23:36here, you can download and use the
- 23:38application directly.
- 23:42Once you understand the basic principles
- 23:44, you can further develop this
- 23:46application.
- 23:49You can clone it from GitHub, fork it,
- 23:51or help develop it further on my GitHub
- 23:54repository so it can be included in the
- 23:56community group.
- 23:59This is regarding the distribution on
- 24:01GitHub.
- 24:03If anyone wants to use GitHub but
- 24:05doesn't know how, you can ask me on my
- 24:07profile.
- 24:10If you're in the SchoolBoonBigNotePlus
- 24:12group, you can contact me privately, and
- 24:14I'll migrate you free of charge.
- 24:18This is step 5; check the components,
- 24:21look at the yellow warning indicators,
- 24:23and save.
- 24:25It's simple. For exporting, you can
- 24:27choose whether you want an Excel file, a
- 24:30CSV file, or an SQLite file, or whether
- 24:32you want to save it on your computer.
- 24:34You can design it however you like.
- 24:37I've designed it to back up the data to
- 24:40SQLite and export to Excel as well.
- 24:43Let's look at the application's results
- 24:45.
- 24:46Okay everyone, this is the repository
- 24:48I'm talking about.
- 24:54For private communities, it includes VPS
- 24:56setup and, you know, the observe feature
- 24:59, which is for organizational use.
- 25:06I'll reveal it to everyone soon, but
- 25:08this is just a brief overview.
- 25:13Okay, everyone, here it is.
- 25:18It will include organizational chat, OCR
- 25:21login, and so on.
- 25:25Okay, so that's it.
- 25:29I'll be using this to download the
- 25:31models.
- 25:34It's like our OCR, the one with the Goji
- 25:37and Rama models.
- 25:40We can download the AI, it's open source
- 25:42.
- 25:45Okay, and we've already prepared for
- 25:47security, data hashing, and cipher
- 25:49implementation.
- 25:51This is about storing the key,
- 25:53specifically the file extension name,
- 25:55and backing up.
- 25:59This includes signing with crypto to
- 26:01protect the data.
- 26:05You can open the project in any disk or
- 26:07computer environment.
- 26:13The concept is the same: click "new,"
- 26:16then remember to open "diskcomputer" and
- 26:18"openfolder"(it's like opening a project
- 26:21, for example, the same thing).
- 26:29This is working within the same project
- 26:31, or even in a chat.
- 26:35Everyone follows the same principle.
- 26:39Once you're done, open
- 26:45the project, select "new project," and
- 26:47you might need to type a name.
- 26:53They're not all the same, I'll start
- 26:55with this much, then add more, and it
- 26:57will be integrated into this program.
- 27:00Everyone follows the same principle,
- 27:02regardless of the course.
- 27:04If you have any other programs, you can
- 27:07run them like this.
- 27:09Okay, then run it here.
- 27:11This is the program for those who
- 27:13already have it and want to run it.
- 27:16This is how to run it. But if it doesn't
- 27:18run, you can tell me.
- 27:20Just open the app normally.
- 27:24This is also something I've prepared for
- 27:26you.
- 27:29You can check the box or verify it
- 27:31before actually running this app.
- 27:34And this addresses common problems that
- 27:36many people may encounter.
- 27:39I've compiled them all. If you misread a
- 27:41lot, it might be due to your own
- 27:42understanding. Is it the source image?
- 27:44Why is it slow?
- 27:46Are we using a smaller model, or a
- 27:48larger one but there's not enough RAM?
- 27:51Or are we using a smaller model and can
- 27:54we adjust the resolution to 4-bit, or 4
- 27:56-bit?
- 27:58It's like a resolution level, something
- 28:00like that.
- 28:02This is some terminology.
- 28:08We also need to check if the input data
- 28:10is correct.
- 28:12This is a system.
- 28:15Okay, to summarize, this video covers
- 28:17everything from installing the Typhoon
- 28:20Orama, to
- 28:27creating security and running it in
- 28:29real-world mode.
- 28:33Today, I'm going to use Google
- 28:35Anti-Gravity.
- 28:39I recommend using the Google
- 28:40Anti-Gravity
- 28:46feature. Antigravity is quite high,
- 28:48everyone. Guild has
- 28:52this
- 28:57slash
- 29:02, and
- 29:07Guild
- 29:14has...
- 29:22okay?
- 29:28Okay,
- 29:39okay...
- 29:49Okay, I'll try selecting a PDF file
- 29:51first, which I've prepared.
- 29:54It might be in your download. Here it is
- 29:56, everyone, the invoice sheet. Let's
- 29:58take a look.
- 30:00Wow, everyone, it's here!
- 30:03And it's in fullscreen.
- 30:06Okay, 0%tilt, definitely, because this
- 30:09is a sample file.
- 30:13And it only took 200ms to find 38 lines
- 30:17.
- 30:20Wow, everyone, it has file encryption.
- 30:24Okay, regarding security.
- 30:27And the invoice is correct.
- 30:29It can also be autotable.
- 30:31Wow, it meets all the requirements. It
- 30:34seems like it knows that this is an
- 30:35invoice, and there's a sample invoice
- 30:37too.
- 30:38Wow, everyone, there's a sample invoice
- 30:41as well.
- 30:42OCR is running. It's already running.
- 30:45And we can save it to SQLite as well.
- 30:48You can export to Excel too, everyone!
- 30:51Wow, this program is perfect!
- 30:56Actually, everyone can download it from
- 30:58the private community, or for those
- 31:00already in the free community, it's
- 31:02available here.
- 31:07But I'll be developing this application
- 31:10on GitHub continuously; it's a CI&CD
- 31:12application.
- 31:15You can follow along and discuss it
- 31:17further.
- 31:19That concludes the learning in this
- 31:20video, but we can build upon it in the
- 31:22community.
- 31:26For those who really want to learn about
- 31:28AI and its real benefits, let's help
- 31:30develop a better society for Thailand
- 31:32and share good AI and good tools because
- 31:34we have them.
- 31:36The idea is already great, and we'll go
- 31:39global and sell it internationally.
- 31:41It's really not an exaggeration; this
- 31:43kind of thing only requires our
- 31:45cooperation, unity, and mutual support
- 31:48within this community.
- 31:51Okay, that's all for today.
- 31:55We'll see each other in the next episode
- 31:57or in the community.
- 32:00Goodbye for now.
- 32:01Oh, and don't forget, everyone, the
- 32:04Typhoon OCR model and OpenThai's work
- 32:06are amazing!
- 32:08And in the next video, I'll teach you
- 32:10how to train your own local AI if you
- 32:12already have data within your
- 32:14organization.
- 32:16Let's wait and see in the next video.
- 32:19Goodbye!
About this transcript
This page contains the full transcript of สร้างแอป Local AI แปลงไฟล์ PDF เป็น Excel ภาษาไทยไม่เพี้ยน! by BoomBigNose, generated from the public captions YouTube serves with the video. The transcript has 3,484 words across 648 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.