Transkribus Webinar for Beginners (English)

Hello everyone. All right, welcome to today's beginner's webinar. Thank you so much for joining. Um, you joined Transcribus probably because you were in a similar situation where you had a historical document with kind of tricky handwriting to read and you were struggling with it. had material and documents and you wanted to extract this information maybe and work with it in a different way and so transcribus uses AI to help you automatically read this document to process it and it basically analyzes the text and then gives you machine readable layout so it does the reading for you and helps make your work easier. What we'll have a look at today is first an introduction to transcribus to the software and then we'll have a look at the workflow the standard workflow. So how to upload your documents how to start the AI recognition and how to pick the right model for your documents and we'll even have a small look at the advanced workflow because maybe some of you already know you can even train your own custom model but more about that later. And at the end uh I hope we'll have time for questions but of course you can always send your questions to the zoom chat. So who is here with you today from the team? My colleague Areno who is working as a solution adviser is here with me and working in chat. So as a solution adviser he has a lot of knowledge can answer your questions and do your best to reply in the chat immediately provide you with helpful links helpful information. And then there's me. I'm Helina. I work in the marketing and comms team. So I usually do the webinars. I also work in social media and I go to events. So again, maybe we've met at Rootste two weeks ago and I'll do my best to uh help you get started with transcribus today. So let's have a look at the introduction. How can we make your work with historical documents a breeze? So with transcribus what you can do is you can use the transcribus AI for text recognition and automatic transcription. So maybe you've had exactly these questions and these struggles that your transcript is never accurate but working with transcribers you can't maybe find the type of model can't really type find the type of writing and that's why you were joined today. Maybe you have already used it and you know what kind of language, what kind of script you're looking for, but you still don't really know how to pick the right model. Or maybe you're just starting out and you don't even know uh what questions you might have and you really just want to get a first idea of how to work with transcribers and how to get those transcriptions. So in that case, you're definitely in the right place today. By the end of this session, you will know how to upload and transcribe a document, but also how to manually edit the results. Maybe there were a few mistakes in the automatic transcription and you'll know how to pick the right model for your material and also when in what case it makes it makes sense and also how to train a custom model. Um again just to remind you transcribus is your AI part LA designed to simplify your time consuming time consuming and laborous work with historical documents. So we really try to make your work with historical documents easier. Um in what way do we do that? Well, with Transcribus, you cannot only transcribe but also search your documents because it is machine readable text. It is also machine searchable. So, you can use full text search to search for specific information within your material. You can also use text classifications for names, places, and dates. And you can even recognize structure because often it's not just about the text itself, but also the structure that the text comes in. and to extract the information in a structured form as well. Now what does this look like in practice? Let's have a look. So we have three different workflows. Um the first one is a quite quick one. It's to be able to get really fast results with readym made models. So this basically works in a drag and drop way where you drag and drop pages into it kind of like often um uh translation tools work. Level two is then a guided workflow where you can improve accuracy by picking a right model. So this is our standard workflow. And then level three is the custom uh the custom training where you train your own model. So the first one is really one where you use transcribus out of the box. You drag and drop your page in similar to translation tool is for quick results kind of like this. So transcribus scans your page and gives you a result immediately. The second level which is a guided one which is our standard workflow. You can use the existing models for your material and we will have a look at this one today and then we will have a sneak peek at our custom model training and I'll show you how you can start training your own custom models as well. But first again we will have a look at our guided level the standard workflow. Now how do we start with transcribus? So we've gone through uh the steps, we've gone through the workflows and now we really want to start at the beginning uploading our documents. Now first we will have to take a look at where you will be doing the uploading and the working in the transcribus web app for some context and on the um uh transcribus web app once you log in this is what you will see. So you will see three main areas here on the side underneath home. Home is of course your landing page. The three main areas are then a desk space which is your central workspace. So here you will manage documents and collections. And this is also where you will start your workflow today. Then you have the models area. Here you can browse and select different AI models for text recognition. And then we have sites. So sites is where you can publish your collections online. But we will get more into that a bit later. We'll take it one step at a time. So first the important area is the desk area because again this is where we work. This is where we upload and organize our documents. Now in transcribus everything lives inside a collection. So these collections are kind of like folders or projects and within these folders are your documents. So for me example because I work in marketing and I do webinars one of my collections one of my projects is a webinar. So within the webinar I then have different collections and within uh different documents I'm sorry and within these documents I have different pages. Now you can have multiple collections with multiple documents and the amount of pages really depends on what type of documents you uploaded. So if you have a letter might just be one page. If you have a book probably maybe hundreds of pages. Now that we know how the documents are structured, let's see how easy it is to bring it into transcribus and we will have a look right into the plat platform and jump on over here. So once we log in, as you can see here, we're now in my home in the landing page and I will move into the desk workspace where my collections are. Here I have an overview of my favorite collections and I can also just move to the general overview where all my collections are um uh situated. So we'll move to my collection called beginner's webinar. But if you're starting out and you want to create a new collection, you simply click here in the top right corner. Click new correction uh new collection. Give it a name. So we'll call this webinar test. create the collection. And now you can already see the area where you can upload your documents. So I'll click on upload. And here you have a overview of what formats you can upload and what size they can be. You can here also upload entire folders and just drag and drop them in. Now let's see. I have prepared a few pages for this webinar and I'm just dropping them in here. We can see that it's uploading. We see the loading bar progressing and we do not do want to sort them before continuing. Enter document title. This is then again the document test webinar. Then we also have a name for the document. And now we can already see we are in my collection called webinar test and we see the document with the three pages that I drag and dropped in called document test webinar. So just in a few seconds your material is safely stored and now ready for text recognition. And now that your documents are uploaded and organized in the collection the next steps is to turn these images into text. So now you will start with the automatic text recognition. Now how does that work? So again in this case we'll use our public AI models to process our documents and do the automatic text recognition. to start the automatic text recognition. Um, we have to use these models because they convert the printed and handwritten text into digital text using our pre-trained AI. So, these models were trained with machine learning. And in this case, it's important to choose the right right approach, which is the public models because they are pre-trained and ready to use. They're quite easy. So you just have to select one and um be mindful of uh the language that you pick. But this is the easiest way and the right approach for the standard workflow. When choosing the public models, we have over 300 public models available for handwritten as well as printed text and for over three uh for over 30 languages and scripts. So in this case um it is really important to filter for models because we have over 300 models available. Um the important thing is to find the right one and this is also a question that we got during registration is how do I pick the right model? There are so many public models available. How do I know which one to pick for my material? And you can do that by using this filter option in the transcribus web app. So you can see here an overview of all the models that we have. We have as you can see um the last time I took the screenshot we had 330. I think it's even more now. Um but you see that this is quite a large number. And you can use these filter options on top. Oh, I'm sorry. I think there is missing a screenshot here but I can show it to you right in the platform if we go to models and public. All right, there we go. As mentioned, it's already quite a much higher amount of public models that are available. But you can simply use the filter option right here to pick a handwritten or printed text and even just filter by selecting a language here. So once you select a language for example English and you filter for handwritten texts and a specific century let's say 17 to 19 you can see that the amount of public models has immediately reduced to just a handful. So that way you have a much easier time of looking through the public models that are available to you. Let's go back. Now if you click on the model description, you will also see further information about the model on the right side. So this will give you more information and a further indication of if the model performs well for your language or for the text that you're working with. So let's say you work with a Spanish document. In this case, you have selected the colossa espanol. You can see here that it has been trained on almost 40,000 pages and it has a cer of only 4.8. So the cer is also another indication of how well the model might perform um next to of course the language and the time frame. But the CR is the character error rate. So that means that only 4.8 out of 100 characters were recognized incorrectly. Meaning during the training there was an over 95% accuracy during the training. And this gives you an indication that this might work quite well for the documents that you're working with. One of the more newer technology that we have are the super models. And what is new and impressive about them is that they are very large and generally more versatile models. So what does that mean? This means that you can use the models for different scripts and languages and also mixed material. So as you can see here we have the text titan one which has been trained on six different languages can see here German and then five other ones. So the super models usually really provide the best out of the box performance. Now once you found a model that matches your material whether it's a public model for a specific script we have had the question about working with German uh documents so we have a lot of German models um it's actually one of the stronger languages that we support so we have one for German Kurand sitelin also frau so we have models for very specific scripts or you use one of our more general purpose um super models you can apply directly to to your documents. Now, the progress of doing this is quite simple, but I think to make this a bit clearer, we will switch over to the web app again and have a look at what this looks like in practice. So I will move to my desk workspace and to my collections and I'll move to my beginner's webinar or actually let's do our webinar test. Let's open our pages that we just uploaded. So in my case I have here my document and I want to recognize all three pages at once. I can do so by selecting the uh document just here by clicking on it or I can also open it. Now I see the overview of the separate pages and now I can either select one or maybe all three depending on how many you want to process at once. Now you just click on process with AI and you can see this slide in on the right side. Now here you select the process type. So right now it's uh the process type that is selected is layout but I want to recognize the text. So I'll click on text and then I have here already a pre-selected model. I probably have used this before um the last time I did a text recognition. But if I deselect it, I can see here again these filter options. Now these are a bit smaller um but they kind of uh work the same way than in the general model overview I showed you before. So here you can filter for specific models if you already know the model name. For example, here I have a few of my favorite models. Or you can go through the model list. Again, you can use the filter to search for a specific language um handwritten or printed text. And this way um select the best model for your material. So again when speaking about best model for your material it's really important to consider what type of documents are you working with what kind of language is it? What kind of script is it? What kind of time frame is it written in? And then select the model that fits exactly this document type the best. This gives you uh the best possible outcome of the automatic transcription. Now in my case I will now use the text titan. Again, these are usually the models that provide the best out of the box performance. And now I just simply click on start recognition and the recognition process starts which I can see here indicated with this scanning bar. Now it might uh take a bit of time to process the pages. This depends uh on the general um um traffic I think maybe is the right word uh that is happening in the background. So um what is happening in the background right here is that transcribus is analyzing these documents to see where is text on the page where is the layout on the page and then converts this handwritten or printed text into machine readable form. Now because this is so loading, we'll just go back to the documents that we have already prepared for today in our collection called beginners webinar and I will open my 17th and 18th century English records. And here you can see I have quite a few pages. Um let's just have a look at this one right here. Now once you open the page this is what you will see. You will see the image that you uploaded on the left and the automatic transcription output on the right side. Now it can happen especially when using public models that there are some mistakes. This happens because these models have seen a lot of training data during their training process. But we know that every handwriting is different. There might be very specific scripts. So in this case the model has of course not seen the exact handwriting that you are working with. So there might be a few mistakes but we can correct this quite easily with our uh keyboard in the editor. Let's see. We will go to our document to the next page. As you can see I can use the sidebar to navigate. And here we have prepared some uh mistakes um ahead of time. And if I go through the automatic transcription, I can zoom in here a bit. I can see for example in this line right here that there was a mist mistake in the transcription and it does not say menu row, but the letter that is written here is actually an M and then a dot. So I can simply click in the text where there is a mistake and use my keyboard to correct it. If we look further down, we can see here that it does not say treasurer but treasurer. So I can simply use my keyboard to correct this. So this way you can use the document editor to correct the automatic transcript, but there are even more options to work with the document editor. Oh, close this so we have a bit of a better view. One other thing that you can correct is the layout. So you can see here the layout part of that document editor. What you see here encased in this green line is the text region. Now the text region is the box that encloses all the handwritten text contained in this image. The line in the blue underneath the text is called baseline. And that is one of the most important reference points for text recognition. And what you can do here is because we can see it has recognized quite well this large body of text and even this marginelia here on the side. Now again this is something that we prepared ahead of time. So these mistakes we prepared but just so you know how you can correct this and edit this. So you can also correct the layout here on the left. Not just the automatic transcription output on the right, but also the layout recognition here because we also want to have this text recognized in the margins. We can add another region and baselines. So to add a region, I go here on the left side in the settings bar. I click on add region. Click add text region. I click once to start drawing the region and click again to end drawing the region. And you can see here I have a new region drawn and it immediately reflects in the um text editor part in the right side of this editor. Now there are no lines there is no transcription yet but we can change that quite easily by drawing these baselines. So again these baselines are reference points for text recognition. Now I can draw the line by clicking on add line clicking once to start drawing the baseline and continuing and then double click to end drawing it. And you can see here see you see here that an empty line appears. So there's no automatic transcription yet but there is a line where the text can go. We can continue drawing these baselines. I just continue real quick. And finished. And we have four additional lines. Now what you can do, the easiest way to add the automatic text recognition is first of all save the changes. And then you can start another text recognition to automatically recognize the text here. But of course you can also manually enter the text. Now I have started the transcription but I can also just use my keyboard. In this case it might be faster and Carlton add the text manually myself of the and so on and so forth. Now you can even let's go back to the selection mode. selection mode. You can even split regions and combine them. So in this case, we have one large body of text and we can see here this is all region two. But if I want to have the structure reflected better in the transcript and I say okay now this is one paragraph but I think this is a separate paragraph and I want this to be reflected. I can select the text region and then split it by clicking H like in a vertical cut and then clicking. You can see this blue bar appearing on the left side. Clicking where you want to split the region. And now we see that this is its separate text region. Maybe we're saying, "Okay, it doesn't really make sense that this text region right here is before the marginalia on the side." And we can fix the layout order in the transcript by using our layout tree here on the right side. And I can move the regions, entire regions, just drag and drop it. And the order has changed. And I can see now that this paragraph is region three and the marginelia is region four. Now one more important thing to know. There we go. That was already a teaser. One more important thing to know is that you can add tags to your documents. Now what are tags? So you can um you can add two types of tags. Structure tags and layout uh textural text. I'm sorry. structure text and textural text. So structural texts help define the hierarchy structure of a document. So they can identify for example titles, paragraphs, marginelia and other structural elements and these help maintain the original layout and the formatting of the documents during the transcription. So meaning and context is preserved. The textual tags on the other hand are used to identify specific textual elements for example names or dates or locations for example and this labeling can also be useful for creating a database and it can also help provide context. So to keep it short these tags serve as labels or markers that are applied to specific elements within your document. So you can apply them to your layout by selecting the region like this using the right click on your mouse and then you should see this overview of structure tags. In this case I'll select paragraph and you can see here paragraph written in the top left corner and also paragraph appearing with the transcript. Let me tag the marginelia as well. We select marginelia. And this way you can add uh more information. Now for textural tags, you need to make sure that this um area, this toggle, this button is enabled and it's blue. You can see here if it's disabled, I'm selecting to enable it. And to add the text tool tags, you simply mark the text in your automatic transcription. And now you can see these tags appearing. So I will tag this as a date. And then we also have for example a person here. So this way you can add more information. You can use these tags to also further edit text. So if there is text written in superscript, you can also update it as well with this tagging feature. Now one other thing that is important here is settings. And if you open settings, you have a few configuration options such as different visibility to for example also show uh line polygons or to remove the baselines. You can see it reflected immediately here in the image on the left side. And you can also increase label size or change the colors of region or baselines. So maybe um there is a bit of a difficulty of distinguishing between colors. So that way we want to make uh it more accessible and more more easier for you to work with the documents in the way that works best. Sometimes also we have for example here the highlight the baseline highlight in orange that maybe if your document is really quite tinted it might be hard to differentiate from the background. So that's what you can change here as well. There are also configuration options for the text, where to center the text, how to align the text, and we have configuration options for tags. So here, for example, you have an overview of all the tags that you have in your collection. And you can choose to disable them for these documents. So that will not be shown anymore. Let me show you. So I disabled marginalia and it cannot be seen in this overview. enable it again and it should be back right here. And of course, you can also add new tags. So, if there are any tags missing, you can click on edit tags in collection settings and simply add new ones. If for example, you're working with uh let's say what example do we have? Maybe a recipe book and you want to add a tag for specific ingredients or a similar thing. Then you can simply of course manually add uh different tags to really reflect the type of material that you're working with. Now that we have reviewed and corrected our text, so we say it's accurate, it's not just accurate, it's also digital and searchable. Um and because again because it is digital, you can also use full text search. But what do we do? Uh what do we want to do now? We have revised it and corrected it and now we want to export it. So let's save our changes. Make sure that everything that we have corrected is saved and let's go back to the overview of our records. So I have corrected the transcript. I am satisfied with it and I want to work with it uh in a different way. So exporting can make a lot of sense if you use the text outside of transcribus for example to include it in a publication with colleagues who don't use transcribus or and we had that question as well in our registration questions. Uh there was a question on how to work with a literary corpus. So for example, if you want to export um uh export the different documents and work with it in a different for example digital humanities tool uh different software further for analysis you can use the export feature that way. So you select the pages you want to export. Um let's just do these three pages for now and then you click on the action button right here. You select export and now you have an overview of the different export features uh export formats that you have. So let's say for example a document and here you can see you can even export the tags from your document. You can also choose another PDF for example and simply click on start the export and as shown in the popup right here the job has started and you will be sent the exported pages via email. Now the export options, the export formats that are available to you um depend on the subscription plan you have but there are a lot of different formats uh PDF, Word, page XML as well. Um again some are only available starting with the paid plan with the scholar plan such as um exporting it as um an Excel file. So if you're working with tables for example, this is available with the pay plan. But if you're still working on the material, it's usually best best to stay within transcribers for now. All your changes are saved in the documents. Um and that way you can keep all your layout, your tags, your version history intact. Maybe you are planning to further edit the documents at a later date, train a model, maybe upload more documents or publish them. So exporting is something that you don't have to do right away, but now you know how it works and how you can get the documents out of transcribers as well. Now let's go back to the slides. So, we've had a look at how we can use our public models. But what do you do when public models don't give you quite the result that you need? So maybe you've tried multiple public models. You have experimented also with using baseline models first. You have also tried some of the advanced settings. We have more resources on how to work with advanced settings in our help center as well. But the recognition is still not really quite there and that is when you can train a public a custom model. So let's quickly compare again. We have the public models on the one hand. They are pre-trained and ready to use. They work very well in many cases but they are still general purpose. The custom models are trained on your specific material. So they can really learn the um specifics of the handwriting that you're working with, the quirks of your handwriting, also regional styles or unique document layouts. So when are custom models useful? They're useful when you're working with documents that have those regional charact characteristic specific handwriting, specific vocabulary, the individual writing styles, maybe also demanding layouts. Um, and they are very versatile for individual applications. So in these cases, we choose the custom models. The custom AI model training uses specific examples to help the AI accurately recognize and understand the writing in your training data. So training a custom model in transcribus is all about giving the AI examples of the text you wanted to recognize together with correct transcriptions side by side. So you give examples to transcribus to teach the model to understand the particular handwriting. Now we do have um our advanced workflow here. So the one step that is different from the standard workflow that we saw previously is really this training part and the training workflow works as such. So we recommend to first do a pre-recoognition with a public model. So let's say you have uploaded a diary because we'll use a diary as an example. You have uploaded a diary and you want to train a custom model for the specific handwriting of the person who wrote this. To do this you need again these examples to train the model. Now how do you get these examples? it takes quite some time to manually get a certain amount of pages. We recommend around 20 pages as a starting point to train the model. So there's a difference of course in the time and effort that goes into manually transcribing 20 pages for this training data or to use a pre to use a public model for pre-recoognition. Then accurately correct the text, save this as ground truth and then start the first training. So the ground truth is the training data and this training data is again accurately transcribed text paired with its image. Now how to start the model training is you use these pages. So I have a screenshot here. We'll move into the web app um shortly but just to show you you have these pages of let's say again 20 pages as a starting point of correctly transcribed material. So you have used a public model, you have applied it, you corrected the transcription, it is now accurate for the training so that the model can learn from this accurately transcribed text can learn from this data that you're presenting to it. You will select these pages that you have saved as ground truth that are transcribed correctly and then you click again on this action button that we used first to export documents. But instead we click on train model and then on text recognition model. And I will show you what that looks like in the web app right here. Now we will use in this case our diary. So we have this diary of Marjgerie Fleming who was a child author. And in this case I will select my let's go with more than 20 pages. We have 25 pages of ground truth. I click on the action button. I click on train model. And then I select text recognition model. So we do have as you can see more options to train a model but for now for today's beginner's webinar we'll stick with the text recognition model. So I click on text recognition model and now we have already selected the training data. So you can see here let me zoom in maybe it's a bit much the training data we have selected that already. You can see here again we recommend at least 20 pages of transcribed material. We have done that. So we can click on next. The next step is the validation data. So the training data were the pages um that represent my material. These are the pages that were correctly transcribed because the AI learns from them. Now the validation data is a group of examples that allow for a neutral evaluation of the model. So what does this mean? So these are the pages that the transcribus AI compares the results with to determine how accurate the model will be during the training. So in this step, 10% are automatically selected of your set. In this case, it's only two pages. We recommend just sticking with the 10% for now. You can do a manual selection and change this, but generally speaking, this should work quite well. Then click on next. And now what's left is just the model setup. So we'll give this a name. We'll call this Marjorie Diary um webinar model. We can give this a description. Marjgery Scottish check author. In this case, we recommend to also give information about the material itself. So, for example, that it is um handwritten text, the time period that it was written in, maybe even a specific script. So, if you're working with German um text, and you're working with Kurand, for example, to specify that as well. Um then the last thing that is mandatory is the language. So we'll select English. And now we can have a look and check the information that we have entered which is everything is correct and we can already start the training. So the process that takes the longest for model training is usually preparing the training data. But again, you can really save time with doing a pre-recoognition with a public model first correcting this and then uh uh using these pages for for the training data. Now that we have trained the model or we started the model training, where can I find my custom model once uh the training has finished? We go again to the models area, but instead of the public models, I go to my models. And here is an overview of all the custom models that I have trained. So these are only visible to me unless I share them with someone. And we can see here a model that we have also prepared ahead of time. You can see here the overview of the model and this is also where you can evaluate the model. So we mentioned the CER the character error rate already previously. The character error rate is an indication of the accuracy during the training. So for example, we have here an accuracy rate of 38.9. So almost 39% meaning that 39 characters out of 100 were recognized incorrectly. That's quite a high number. We can see that the training pages that we used with this model, the amount of pages we used were only 19. So that's not that much of pages or not that many pages that a model can learn from. So we always recommend to add more pages to the training. The more pages the model sees during the training, generally speaking, the accuracy improves and I can show to you what that looks like back in the slid set. I think we'll have to skip maybe a few. Yes. So we have seen already the different training steps and we can have a look at the evaluation of the model. So this is a screenshot of the model that we already saw. So the accuracy during the training was only 61%. And we can also see this here in the learning curve. So the learning curve shows the process of the accuracy. The yaxis represents the character error rate. So with the training starting and then progressing, the curve goes down as the model improves. The xaxis represents what we call an epoch or um a full round of training. And each of these rounds of training helps the model to better understand uh the data. So it's basically practice rounds. So the more practice rounds the model has, the better it gets. But if there is no further improvement, the model stops the training. Now what we recommend in this case cuz the accuracy is not quite good is to retrain the model and simply correcting the transcription in document does not improve or retrain the model. Accuracy only increases by training new model versions with updated ground truth. What does that mean? So we had this previous workflow where it said first recognition or pre-recoognition with public model. But to retrain a model, we now recommend a recognition with your model. So that would that then be Marjgery diary version one. That would be your model. Uh accurately correct the text, save as ground truth, start a new training, recognize more pages with the new and improved model. So these added pages that you have corrected, you keep adding to the training progress. So let's say I have had 20 pages of ground truth first. Now I can use my model to recognize 20 more pages. Accurately correct these 20 more pages. Save them as ground truth. And now start the new training with 40 pages of ground truth. With the second iteration with 40 pages of training data, the model will probably already have a lower CER, meaning it will perform better. Now I can recognize even more pages. Let's say I can recognize 40 more pages, add those to the existing 40 and start a new training with 80 pages of ground truth. So repeating this process will improve your model and we can see here the effect of this. So we have here the version two. So speaking of training pages, previously we had 19. Now we have 164 pages that we used for training. Meaning the model has seen a lot more variation, a lot more handwriting during the the training process. And we can also see here the accuracy has improved by a lot. So the accuracy is now over 90%. We can also see that reflected in the training stats. And we have a very nice uh table here. Now this is an indication of how many pages of training data we recommend for a reliable model. We always uh say to aim for a character error rate of under 10%. With a character error rate of under 10% you should already have a quite reliable custom model. And retraining a model is a normal part of the uh training process. So don't get discouraged if the first model with maybe 25 pages, 30 pages has a CER of over 10%. This is quite normal. Um just simply add more correctly transcribed pages to the training data. start a second iteration of training and it should improve already by quite a lot. Exactly. So it is a normal part of the process and each attempt brings you closer to your ideal results. And that brings us slowly to the end. Um now you know how to upload and transcribe a document successfully, how to manually edit results, how to pick the right model, and even how to train and improve a custom model. So we've gone through the core workflow from uploading your documents, recognizing text with public or super models, editing and even tagging your text and even training a custom model when you need this extra accuracy. But there is even more that you can do in transcribus. It's not just text recognition. We also have as mentioned before different types of models that you can train such as baseline models. baseline models can be trained to accurately recognize text lines in your material. So if you have material that has um a bit of a tricky text layout in terms of sort of lines that are slanted, quite long or quite short um then we recommend training a baseline model or using a baseline model. And using this will definitely also improve the text recognition because when the layout isn't correctly recognized, the AI has a hard time also recognizing the text correctly because it needs to know where is the text written on the page to accurately reflect that in the automatic output. We also have field models that you can train. So they can also be trained to automatically recognize and even mark certain layout components. So they can also be trained to automatically tag certain layout components in your material. Now when you're working with specific document layouts, we also had the questions about church records, maybe tabular data, um layout, field models as well as table models are quite useful to train because they help you to extract the information uh in an accurate way. So not just the text but also the the structure that the text and the information comes in. So with the table models we can also train transcribers to automatically recognize the rows and columns in your training data. We have uh previous hel previously held webinars on field and table models. We have also recorded these webinars and they are uploaded to have been uploaded to our YouTube channel. So you can go to our YouTube channel and have a look in the webinar playlist and find the webinar recordings there. So if this is interesting to you, if you're working with documents with more complex layouts and finding out more about baseline field or table models, you can have a look at our help center and our YouTube channel. And um oh yes, here we have a nice preview of what that looks like to have the text also extracted in tabular tabular data tabular form. And I also mentioned previously that we have a publishing feature which is called transcripted sites. So you can create a searchable online database of your documents that uh everyone can access your documents online without the need for programming knowledge or extensive um IT resources and you can share entire collections with either a private group or public group from anywhere. So, transcribed sites offers a very simple and efficient way to publish your collection online. Again, you can also find more information about that on our help center or on our YouTube channel. We have done a previous webinar and we have helpful guides uh there as well. And with this, I can now uh wrap up slowly the webinar and we can have maybe some final questions at the end. So just to wrap it up with transcribus, we really want to help you save time and provide a tool that can be learned to use quite quickly. AI should be able to be understood and controlled as well. Technology should be for everyone and not just tech experts. So, we really want to help you save time uh when it comes to manual transcription by using transcribus AI to reduce the effort and time involved and free up time for working with the content of your documents. Spend time on researching on analyzing and have to spend less time on manually transcribing. Um it can be quite difficult also for the untrained eye. Um and it can be uh the historical documents can be damaged by frequent use. So we try to provide easier access to old documents digitized, readable and again searchable. We know that sometimes technology can be a bit challenging and learning a new technology can be challenging. So we really try to keep a low threshold access where you don't need an installation again don't need too much of IT experience. We try to provide these webinars also to help you get started of course. Um, and we really want to help you achieve accurate text recognition with the existing models that we provide uh for free in our free plan and also our super models which are new technology uh with uh quite good out of the box performance and even the option of training custom models which the custom model training is also even part of our free subscription plans as well because we want to give everyone the opportunity to work with historical documents. Um, so it's really not just about reading text. It is a main part, but it really is about uncovering historical um, information. So our goal is to provide the best tools for making history accessible. And I keep saying we, but who are we? Who is transcrib? Um, transcribus itself comes from the university sector. So we started as a project at the University of Insburg in Austria. um funded by the European Union. We are now a cooperative with over 250 members. So we are communityowned and even the first AI cooperative that exists in this way. And being a cooperative by our statute statutes, we cannot distribute profit. So purpose of our profit is our motto. Everything is reinvested into transcribus. Trying to make the platform better, develop new features and tools and implement them such as for example field and table models is one of the more newer additions. Uh we're also working on a new feature that is called end toend models which will be coming this year. So um we always try to improve it, make it better for our amazing user community. uh that is also helping us be better by providing very helpful feedback. And as you can see, we have many amazing members uh co-owners in the cooperative who are all also very involved in improving the platform uh from all across the world and different sectors uh and where we try with them to create a even better transcript tool and software for for amazing users. So yeah, now it's your turn. Uh you can head over if you haven't already to transcribus atapp.transcribus.org. Give it a try. Try it out. We have helpful resources in our help center. Um we also have regular webinars and the webinar recordings again are on YouTube. We will also host a transcribus user conference this year in September. Call for papers is still open. If you have an interesting project with transcribus that you want to share, feel free to uh propose uh or to hand in a proposal. Um the ticket sale should start next week, so maybe we'll see each other in person in September. So, we're really excited to to connect with our user community at the conference as well. And of course, we're uh also active on social media. We always post updates of new features, new models, new webinars there as well. So feel free to connect with us there as well. Now let's see. I think we have maybe a few more minutes for questions. Are there questions that were unanswered that we can maybe have a look at in the platform or have a >> I hopefully everything everything nobody can complain about it. >> Wow. Amazing. Thank you. for answering everything and thank you for everyone for also um asking questions. I mean this is what why we do the webinar. We really want to give you the opportunity to ask questions right away. Hopefully have them answered right away. Um so thank you all so much for joining today. I think we can now wrap up the webinar. We're we're ending quite on time. Very punctual. Very nice. Um again we'll have regular webinars. So, if there were any questions that were still unanswered, feel free to join the next one, ask questions, then we also have a contact form on our help center. Um, if there are more specific questions if you have any struggles with your documents once you've tried it out and there's just something that you can't figure out how to do it, feel free to head on over there. Uh, look for answers or contact our help center. And I think with this we can wrap it up again. Thank you so much. Hopefully I'll see you at one of the next webinars as well. We will send you the recording of today's webinar via email probably next week. And enjoy your morning, your evening, your afternoon. And uh it was great to have you here. Thank you and goodbye. See you next time.