How to Train a Custom Font with Tesseract 5 in C#

How Do I Set Up the Training Environment?

Install-Package IronOcr

sudo apt install tesseract-ocr
sudo apt install libtesseract-dev

sudo apt install tesseract-ocr
sudo apt install libtesseract-dev

How Do I Prepare the Font for Training?

cd /
cd /mnt/d

cd /
cd /mnt/d

How Do I Copy the Font File to Ubuntu Font Folder?

Tesseract renders sample text in your font to build training images, so the font has to be installed on the Linux side, not just on Windows. Copy the font file to both Ubuntu font directories: /usr/share/fonts and /usr/local/share/fonts. The simplest way is to type \wsl$ in File Explorer's address bar to browse the Ubuntu filesystem from Windows, then drag the .ttf across.

Here's how the font copy should look once it lands in the Ubuntu fonts directory:

What If I Get Destination Folder Access Denied?

If File Explorer rejects the copy, run it from a root shell instead:

cd /
su root
cd /c/Users/Admin/Downloads/'AMGDT Regular'
cp 'AMGDT Regular.ttf' /usr/share/fonts
cp 'AMGDT Regular.ttf' /usr/local/share/fonts
exit

cd /
su root
cd /c/Users/Admin/Downloads/'AMGDT Regular'
cp 'AMGDT Regular.ttf' /usr/share/fonts
cp 'AMGDT Regular.ttf' /usr/local/share/fonts
exit

SHELL

How Do I Clone the Training Repositories from GitHub?

The training pipeline depends on three repositories. Clone the tutorial wrapper first, then the two upstream Tesseract repos inside it, then create the output folder:

git clone https://github.com/astutejoe/tesseract_tutorial.git
cd tesseract_tutorial
git clone https://github.com/tesseract-ocr/tesstrain
git clone https://github.com/tesseract-ocr/tesseract
mkdir tesstrain/data

git clone https://github.com/astutejoe/tesseract_tutorial.git
cd tesseract_tutorial
git clone https://github.com/tesseract-ocr/tesstrain
git clone https://github.com/tesseract-ocr/tesseract
mkdir tesstrain/data

• Tesseract_tutorial bundles the Python scripts and config files that drive each training step (text generation, image rendering, training-pair creation).
• tesstrain contains the Makefile that drives the actual training run.
• Tesseract contains the tessdata folder with stock .traineddata files used as the starting model for custom training.
• tesstrain/data is where generated .box files (character bounding boxes), .tif images, and intermediate LSTM checkpoints all land.

Here's how the clone sequence should look in the terminal:

For working with multiple language packs alongside a custom one, see our international languages guide.

How Do I Generate Training Files?

How Do I Run the split_training_text.py Script?

From the Tesseract_tutorial folder, run:

python split_training_text.py

python split_training_text.py

SHELL

The script generates one .box / .tif pair per training sample and writes them to the data folder.

Here's how the script run should look as it generates the training pairs:

How Do I Fix Fontconfig Warning?

If you see the warning Fontconfig warning: "/tmp/fonts.conf, line 4: empty font directory name ignored", fontconfig cannot resolve the font directories. Fix it by editing tesseract_tutorial/fonts.conf:

<dir>/usr/share/fonts</dir>
<dir>/usr/local/share/fonts</dir>
<dir prefix="xdg">fonts</dir>

<dir>~/.fonts</dir>

<dir>/usr/share/fonts</dir>
<dir>/usr/local/share/fonts</dir>
<dir prefix="xdg">fonts</dir>

<dir>~/.fonts</dir>

XML

Copy it to /etc/fonts:

cp fonts.conf /etc/fonts

cp fonts.conf /etc/fonts

SHELL

Then point split_training_text.py at the same path:

fontconf_dir = '/etc/fonts'

fontconf_dir = '/etc/fonts'

PYTHON

How Many Training Files Should I Generate?

By default the script generates 100 training pairs. Change the count near the top of split_training_text.py:

Sizing guidance:

• 100-500 samples are enough to confirm the pipeline works end-to-end.
• 1000-5000 samples are the working range for production accuracy.
• Training text must cover every character your font needs to recognize, ideally several times each.
• More samples mean more training time; pick the smallest count that hits your accuracy target.

How Do I Build My Custom Font Trained Data File?

From the tesstrain folder, run:

TESSDATA_PREFIX=../tesseract/tessdata make training MODEL_NAME=AMGDT START_MODEL=eng TESSDATA=../tesseract/tessdata MAX_ITERATIONS=100

TESSDATA_PREFIX=../tesseract/tessdata make training MODEL_NAME=AMGDT START_MODEL=eng TESSDATA=../tesseract/tessdata MAX_ITERATIONS=100

• MODEL_NAME is the name of your custom font (used for the output filename).
• START_MODEL is the base .traineddata you downloaded above.
• MAX_ITERATIONS caps the training run; higher values typically reduce error rate.

WORDLIST_FILE := $(OUTPUT_DIR2)/$(MODEL_NAME).lstm-word-dawg
NUMBERS_FILE := $(OUTPUT_DIR2)/$(MODEL_NAME).lstm-number-dawg
PUNC_FILE := $(OUTPUT_DIR2)/$(MODEL_NAME).lstm-punc-dawg

How Do I Fix "Failed to Load Script Unicharset"?

Download Latin.unicharset from langdata_lstm and place it in the tesstrain/data/langdata folder.

The .unicharset file defines which characters the trainer is allowed to emit. Use the file that covers every character in your font, for example Cyrillic.unicharset for Cyrillic fonts or Devanagari.unicharset for Devanagari.

Here's how a successful training run should look as tesstrain produces the .traineddata file:

How Do I Verify the Accuracy of My Trained Data File?

With 1000 .box and .tif files and 3000 training iterations, the output AMGDT.traineddata reaches a training error rate (BCER) of around 5.77%.

To test the trained model with IronOCR, point UseCustomTesseractLanguageFile at the file and read a sample image:

using IronOcr;

// Load the trained model; AutoOsd handles orientation
var ocr = new IronTesseract();
ocr.UseCustomTesseractLanguageFile("path/to/AMGDT.traineddata");
ocr.Configuration.PageSegmentationMode = TesseractPageSegmentationMode.AutoOsd;

// Preprocess so the model sees clean glyphs
using var input = new OcrInput();
input.LoadImage("test-image-with-amgdt-font.png");
input.EnhanceResolution(300);
input.DeNoise();

// Confidence reflects training quality
var result = ocr.Read(input);
Console.WriteLine($"Text: {result.Text}");
Console.WriteLine($"Confidence: {result.Confidence}%");

using IronOcr;

// Load the trained model; AutoOsd handles orientation
var ocr = new IronTesseract();
ocr.UseCustomTesseractLanguageFile("path/to/AMGDT.traineddata");
ocr.Configuration.PageSegmentationMode = TesseractPageSegmentationMode.AutoOsd;

// Preprocess so the model sees clean glyphs
using var input = new OcrInput();
input.LoadImage("test-image-with-amgdt-font.png");
input.EnhanceResolution(300);
input.DeNoise();

// Confidence reflects training quality
var result = ocr.Read(input);
Console.WriteLine($"Text: {result.Text}");
Console.WriteLine($"Confidence: {result.Confidence}%");

Imports IronOcr

' Load the trained model; AutoOsd handles orientation
Dim ocr As New IronTesseract()
ocr.UseCustomTesseractLanguageFile("path/to/AMGDT.traineddata")
ocr.Configuration.PageSegmentationMode = TesseractPageSegmentationMode.AutoOsd

' Preprocess so the model sees clean glyphs
Using input As New OcrInput()
    input.LoadImage("test-image-with-amgdt-font.png")
    input.EnhanceResolution(300)
    input.DeNoise()

    ' Confidence reflects training quality
    Dim result = ocr.Read(input)
    Console.WriteLine($"Text: {result.Text}")
    Console.WriteLine($"Confidence: {result.Confidence}%")
End Using

The Confidence property is the per-document score; if it stays low even on clean inputs, the most common causes are too few training samples or a base model that doesn't match the script. Once your .traineddata is verified, see our custom language guide for the general workflow of loading any custom language file.

What Are the Key Takeaways for Custom Font Training?

Training a custom font is a one-time setup: generate .box / .tif pairs from your target font, build a .traineddata model with tesstrain, then load it through UseCustomTesseractLanguageFile. From there IronOCR reads images with the new model exactly the same way it reads stock English.

Key advantages of using IronOCR with a custom Tesseract model:

• Reuses standard Tesseract artifacts: any .traineddata file you can build with tesstrain works in IronOCR without conversion.
• Cross-platform output: training requires Linux (or WSL2), but the trained file ships with your application on Windows, macOS, Linux, and Docker.
• Drop-in with the rest of the API: combine custom fonts with multiple secondary languages, image quality correction, and DPI tuning without changing the recognition path.
• Tunable accuracy: error rate is a function of training samples times iterations. Both knobs are exposed (the script's sample count plus MAX_ITERATIONS) so you can dial in the trade-off between training time and BCER without leaving Tesseract.

For larger pipelines, consider progress tracking and async processing when applying your trained model across many documents.