This page is generated. It is published automatically from doc/wiki/training-a-custom-classifier.md in the BirdNET-Go repository. Do not edit it here; changes are overwritten on the next sync. To propose a change, open a pull request against the source file.

This guide explains how to train a custom BirdNET classifier for detecting specific sounds using the BirdNET Analyzer GUI. In this example, we'll demonstrate training a classifier specifically for dog bark sounds. By improving dog bark detection, we reduce the possibility of false positive matches where dog barks might be incorrectly identified as bird sounds (such as owl hoots or other similar vocalizations). The same process applies to any custom sound classification.

• BirdNET Analyzer GUI (screenshots in this guide are from v1.5.1; v2.4.0 has also been verified to work correctly, though some intermediate versions had serious bugs in the training code that produced broken models)
• Audio editing software (Audacity, iZotope RX, or similar)
• Audio samples containing target sounds (positive matches)
• Audio samples without target sounds (negative matches)

Training a custom BirdNET classifier requires two types of audio data:

• Positive matches: Audio segments containing the target sound you want to detect
• Negative matches: Audio segments that do not contain the target sound

The quality of your classifier depends heavily on having diverse and representative samples of both positive and negative matches.

BirdNET requires a specific folder structure for training data. Each folder represents a class that the AI model will learn to recognize. Folder names follow the pattern: ScientificName_CommonName

The most important folder for a well-performing classifier is the Background (or Noise) folder. This special folder contains ambient audio samples that should NOT be matched to any specific class. It teaches the model what constitutes "nothing of interest" in your environment.

The Background folder should contain:

• Background hiss and static
• Wind sounds
• Rain and weather sounds
• Traffic noise
• General environmental ambience
• Any other sounds that you want BirdNET to ignore

This Background class is essential - without it, the classifier will struggle to distinguish between actual target sounds and random environmental noise, leading to many false positives.

Here's an example of a complete training data folder structure with multiple classes:

The structure includes:

• Background: Ambient sounds that should not trigger any detection
• Target classes: Each folder named as ScientificName_CommonName
  • Dog_Dog - Dog bark sounds
  • Garrulus glandarius_Eurasian Jay - Eurasian Jay calls
  • Dendrocopos major_Great Spotted Woodpecker - Woodpecker sounds
  • Certhia familiaris_Eurasian Treecreeper - Treecreeper calls
  • Engine_Engine - Engine sounds
  • Human vocal_Human vocal - Human vocalizations
  • Loxia curvirostra_Common Crossbill - Crossbill calls
  • Ovis aries_Sheep - Sheep sounds
  • Picus Canus_Grey-headed Woodpecker - Grey-headed Woodpecker calls
  • Pyrrhula pyrrhula_Eurasian Bullfinch - Bullfinch calls
  • Tetrastes bonasia_Hazel Grouse - Hazel Grouse calls
  • Vulpes vulpes_Red Fox - Fox vocalizations
  • And many more...

Note: You can also create specific negative match folders using the minus prefix (e.g., -Dog_Dog, -Ovis aries_Sheep) for targeted negative samples of a particular class, though the Background folder serves as the universal negative match repository.

For each class folder:

• Save audio clips containing the target sound in the appropriately named folder
• Critical: Populate the Background folder with diverse ambient sounds from your recording environment
• BirdNET Analyzer supports a wide range of audio formats - you can mix formats like AAC, MP3, WAV, FLAC, etc. within and across folders
• Ensure each folder has sufficient samples (minimum 10-20 clips recommended)

Personal preference note: I use uncompressed FLAC files as they provide the best flexibility for audio post-processing (noise suppression, filtering, etc.) if needed later. However, compressed formats like MP3 and AAC work perfectly fine for training.

Most audio editors display audio as waveforms by default:

However, for this task, you should use spectrogram view instead. The spectrogram provides a visual representation of frequency content over time, making it easier to identify and isolate specific sounds:

In the spectrogram view, different sounds appear as distinct patterns:

• Dog barks appear as vertical bands with specific frequency characteristics
• Background noise has different visual patterns
• This visual distinction helps in accurately separating positive and negative samples

For this example, we'll process 15-second audio clips containing dog barks:

1. Load your audio file into your audio editor (Audacity or iZotope RX)
2. Switch to spectrogram view to better visualize the sound patterns
3. Identify the target sounds (dog barks in this example)

The negative matches are all portions of the audio that do NOT contain your target sound:

1. Identify dog barks in the spectrogram view

2. Delete the dog bark segments to isolate only the background/negative audio

3. Combine the remaining segments into a continuous audio track without any dog barks - in this example selected part will be exported as negative sample

4. Export the audio using your audio editor's export function

5. Save to the negative folder (e.g., -Dog_Dog or Background folder)

Launch the BirdNET Analyzer GUI and navigate to the Train tab:

Configure the following settings:

Parameter	Value	Description
Input folder	Select your training data folder	Contains both positive and negative sample folders
Output folder	Choose destination for classifier	Where the trained model will be saved
Classifier name	e.g., "my_custom_classifier"	Name for your custom classifier file
Upsampling ratio	0.4	Balances class distribution
Hidden units	2048	Network complexity parameter
Use mixup	Enabled	Data augmentation technique
Crop mode	Segments	Allows training with continuous audio
Overlap	1.0	Increases training data through overlapping
Model save	Append	Adds to existing model

• Segments mode: Allows training with long continuous audio clips. BirdNET automatically segments input audio into 3-second chunks for learning
• Overlap (1.0): Creates slightly overlapping segments from the audio, effectively increasing training data and potentially improving the learning process
• Upsampling ratio (0.4): Helps balance the dataset when you have unequal amounts of positive and negative samples
• Hidden units (2048): Determines the complexity of the neural network; higher values may capture more complex patterns but require more training data
• Model save - Append: This critical setting means your training data will augment (be appended to) the default BirdNET classifier rather than replacing it. This creates new classes/labels in the model while preserving all existing bird species detection capabilities

Important Note about Append Mode: When using append mode, your custom classes (like Dog_Dog) are added as new labels to the existing BirdNET model. If a label already exists in the base model, BirdNET-Go will intelligently handle matching labels as a single species, preventing duplicate detections.

Note: These values were determined through trial and error and I have found that they work well for classifiers used with BirdNET-Go in non-scientific setting

Once all settings are configured:

1. Click the Train button
2. Monitor the training progress in the output window
3. Wait for training to complete

The training process will:

• Load and preprocess your audio data
• Segment the audio into 3-second chunks
• Train the neural network on these segments
• Save the resulting classifier model

After training completes, you'll have a custom classifier file that can be used with BirdNET to detect your target sounds.

1. Load the newly created classifier in BirdNET
2. Test it with new audio samples
3. Adjust training parameters and retrain if needed

1. Diverse samples: Include various examples of your target sound in different conditions
2. Quality negative samples: Ensure negative samples represent the typical environment where detection will occur
3. Balanced dataset: Try to have roughly equal amounts of positive and negative training data
4. Iterative refinement: Test your classifier and retrain with adjusted parameters or additional data as needed
5. Audio quality: Use high-quality recordings when possible (48kHz sample rate is ideal since BirdNET resamples all audio to 48kHz internally)
6. Background folder importance: Always include a well-populated Background folder with diverse ambient sounds

• Poor detection accuracy: Add more diverse training samples or adjust the upsampling ratio
• Too many false positives: Include more representative negative samples in the Background folder
• Training fails: Check audio file formats (BirdNET supports WAV, MP3, AAC, FLAC, and many others) and folder structure
• Overfitting: Reduce hidden units or add more training data variety

Training a custom BirdNET classifier requires careful preparation of audio data and thoughtful configuration of training parameters. The key to success is having representative samples of both what you want to detect (positive matches) and what you want to ignore (negative matches). The Background folder plays a crucial role in teaching the model to distinguish between target sounds and ambient noise. Through iterative training and testing, you can create effective classifiers for various sound detection tasks beyond bird identification.