BirdNET to BirdNET-Plus Vision and Scope
📌 The overarching goal of BirdNET-Plus is to evolve from simple bird species identification to a suite of tools that enhance applied ecological research using AI-powered acoustic monitoring.
🦉 The vision is for BirdNET to become the most sophisticated tool for terrestrial audio data analysis and the go-to resource for bioacoustics researchers.
➕ The "Plus" signifies improvements in models (faster, better performance), taxonomic agnosticism (moving beyond just birds to terrestrial, potentially marine bioacoustics), and supporting more devices.
⌚ A key focus is implementing real-time acoustic monitoring, demonstrated by an example where immediate detection of Western Capercaillie breeding grounds can inform visitor path closures to prevent disturbance.

Feature Embeddings and Transfer Learning
🧠 The solution for extending functionality beyond pre-trained species (like hyenas) is using feature embeddings, which are high-level numerical representations of the sound derived from the model's deep convolutional neural network (CNN) backbone.
📊 Embeddings allow for measuring similarity between audio inputs using metrics like Euclidean distance or cosine similarity in the vector space.
🛠️ Transfer learning involves freezing the main BirdNET model, removing the final classification layer, and replacing it with a custom-trained fully connected unit for new downstream tasks (e.g., identifying bat species or call types).
🔬 This transfer learning approach demonstrated success in distinguishing subtle differences like bird dialects or classifying non-bird species (like bats and marine mammals) with surprisingly few samples (e.g., achieving an AUC $\ge 0.9$ with only 8-32 examples per species).

Practical Model Training with BirdNET Analyzer GUI
💻 BirdNET-Plus development emphasizes open-sourcing the analyzer repository to facilitate adoption by conservation practitioners, with a dedicated GUI available (at birdnet.cornell.edu/analyzer) for no-code model training.
📂 Training requires organizing audio files into folders, where folder names become the class labels; for instance, spotted hyena call types (giggles, grunts, rumbles, etc.) were used as classes.
⚠️ A crucial insight for improving model performance is to include a "noise" or "background" class in the training data to teach the model what *not* to detect, significantly reducing false positives (e.g., false grunts in hyena analysis).
📏 Input constraints mandate that audio must be processed as 3-second snippets sampled at 48 kHz; inputs outside this range (like ultrasonic bat recordings) must be pre-processed, often by frequency shifting to stretch the signal into the supported 0–15 kHz range.

Key Points & Insights
➡️ The BirdNET-Plus roadmap targets a Version 1.0 release later this year, promising multi-OS support (including Mac/Linux builds) and feature freeze for reproducible results.
➡️ Users should leverage the open-source repository for feedback via the GitHub "Issues" tab for bug reports or feature requests, as developer understanding of real-world use cases is vital.
➡️ When training custom models, starting with a small, representative dataset and incrementally adding data is recommended over theoretically estimating the required sample size, emphasizing that including a negative/noise class dramatically improves classification precision.
➡️ Users interested in analyzing ultrasonic data (like bats) can use the current system by ensuring the signal frequency is shifted or contained within the 0–15 kHz range after resampling to the model's required 48 kHz input.

📸 Video summarized with SummaryTube.com on Nov 16, 2025, 12:39 UTC