birdnet

A Python library for identifying bird species by their sounds.

The library is geared towards providing a robust workflow for ecological data analysis in bioacoustic projects. While it covers essential functionalities, it doesn’t include all the features found in BirdNET-Analyzer, which is available here. Some features might only be available in the BirdNET Analyzer and not in this package.

The package is also available as an R package at: birdnetR.

Features

• 🐦 Extract classification scores and embeddings from audio recordings — 6,522 species (V2.4) or 11,000+ (V3.0 preview)
• 📍 Predict species presence for a given location and time — 6,522 species (V2.4) or 12,000+ (V3.0)
• 🧠 Utilize your custom-trained acoustic models from BirdNET-Analyzer
• ⚙️ Support for both CPU and GPU execution (including multiple GPUs at the same time)
• 🚀 Multiprocessing support for fast batch analysis of large datasets
• 💾 Low memory footprint and small disk space requirements
• 🎵 Support for various audio file formats (WAV, FLAC, OGG, MP3, etc.)
• 📊 Export results in various output file formats (CSV, Arrow table, Parquet, Numpy, etc.)
• 💻 Cross-platform: Windows, macOS, and Linux
• 🌍 Use multilingual species names (English, German, French, Spanish, etc.)
• ⬇️ Auto-download of all official models
• 🛜 Full offline usage using local (custom) model files

The library is optimized for a minimal memory footprint and maximum scalability, making it suitable for both edge devices and high-performance computing clusters.

Speed benchmarks

Device	Specs	Disk	OS	Real-time factor	Time to predict 1 h
Intel i7 8th Gen	4 cores	NVMe	Windows	50×	72 s
Ryzen 7 3800X	8 cores	NVMe	Linux	420×	8.5 s
Nvidia Titan RTX	24 GB VRAM	NVMe	Linux	2460×	1.5 s

For more detailed benchmarks, please refer to the documentation.

Installation

Instructions

TensorFlow is an optional dependency. The base package ships ONNX Runtime and LiteRT wherever they have wheels (everywhere but macOS x86_64 on Python 3.14 for ONNX Runtime; see the footnote for LiteRT), so it runs the V3.0 models out of the box; the TensorFlow-based backends come with the tf extra:

# Base: acoustic 3.0 + geo 3.0 via the onnx backend, 2.4 models via LiteRT
pip install birdnet --user

# TensorFlow: the tf (TFLite) and pb (ProtoBuf) backends, i.e. the 2.4 models with
# default settings, custom classifiers, Perch V2 and geo 3.0 tf
pip install birdnet[tf] --user

# TensorFlow with CUDA (Linux; NVIDIA driver and CUDA need to be installed in advance)
pip install birdnet[and-cuda] --user

# PyTorch backend for the 3.0 models (.pt models)
pip install birdnet[pt] --user

Install	Runtimes	Enables
birdnet	ONNX Runtime, LiteRT*	acoustic 3.0 and geo 3.0 (onnx); acoustic 2.4, custom 2.4 classifiers and geo 2.4 (tf with library="litert")
birdnet[tf]	+ TensorFlow	tf with the default TFLite interpreter and pb for all models: acoustic/geo 2.4, acoustic/geo 3.0, custom TFLite/ProtoBuf/Raven classifiers, Perch V2
birdnet[and-cuda]	+ TensorFlow with CUDA (Linux)	as [tf], plus GPU for the pb backends
birdnet[pt]	+ PyTorch	acoustic 3.0 and geo 3.0 (pt, CPU/GPU)
birdnet[onnx]	(compatibility alias)	ONNX Runtime is part of the base install
birdnet[repro]	exact pinned versions incl. TensorFlow	reproducible results (Python 3.12, CPU)

* LiteRT (ai-edge-litert) is installed with the base package on Linux, macOS ARM64 and Windows x86_64 for Python 3.11–3.13 (no wheels elsewhere; on 3.14 add it with pip install ai-edge-litert). It runs the 2.4 .tflite models (acoustic 2.4, custom 2.4 classifiers, geo 2.4; pass library="litert" to birdnet.load(..)) and does not need TensorFlow. Without the tf extra, birdnet.load("acoustic", "2.4", "tf") raises a ValueError naming both options.

import birdnet

# base install
model_onnx = birdnet.load("acoustic", "3.0", "onnx")
predictions_onnx = model_onnx.predict("example/soundscape.wav")

model_litert = birdnet.load("acoustic", "2.4", "tf", library="litert")
predictions_litert = model_litert.predict("example/soundscape.wav")

# birdnet[tf]: the default TFLite interpreter
model_tflite = birdnet.load("acoustic", "2.4", "tf")
predictions_tflite = model_tflite.predict("example/soundscape.wav")

Platform support and Python versions

The TensorFlow-based backends (birdnet[tf]) are limited to the platforms and Python versions TensorFlow ships wheels for:

Platform	Architecture	ProtoBuf-CPU	ProtoBuf-GPU	TFLite	LiteRT
Linux	x86_64	3.11, 3.12, 3.13	3.11, 3.12, 3.13	3.11, 3.12, 3.13	3.11, 3.12, 3.13
	ARM64	3.11, 3.12, 3.13	/	3.11, 3.12, 3.13	3.11, 3.12, 3.13
MacOS	x86_64	3.11, 3.12	/	3.11, 3.12	/
	ARM64	3.11, 3.12, 3.13	/	3.11, 3.12, 3.13	3.11, 3.12, 3.13
Windows	x86_64	3.11, 3.12, 3.13	/	3.11, 3.12, 3.13	3.11, 3.12, 3.13
	ARM64	/	/	/	/

For details see the official TensorFlow documentation. The base install (onnx backend, pt extra) has no such restriction beyond ONNX Runtime's and PyTorch's own wheels (ONNX Runtime's last macOS x86_64 build is 1.23.2, with no Python 3.14 wheel).

Python 3.14

TensorFlow does not yet publish wheels for Python 3.14, so birdnet[tf] installs nothing extra there and the table above stops at 3.13. The base install works as everywhere else: the acoustic 3.0 and geo 3.0 models via onnx (or pt), and — after pip install ai-edge-litert — the acoustic 2.4 model, custom 2.4 classifiers and the geo 2.4 model via the tf backend on LiteRT (library="litert"). The TensorFlow-only paths — the pb backend, the tf backend with the default tflite interpreter, and the Perch model — raise a clear error on Python 3.14; to use them, install birdnet[tf] on Python 3.11–3.13.

If you encounter issues with audio file reading, please ensure that libsndfile is installed on your system.

• Ubuntu/Debian: sudo apt-get install libsndfile1
• macOS (using Homebrew): brew install libsndfile
• Windows: Download and install the precompiled binaries from the official website, extract them and add the folder to path.

Supported operations, precisions and devices

V2.4

Model	Acoustic	Acoustic	Acoustic	Geo	Geo
Backend	TFLite/ LiteRT	ProtoBuf	ProtoBuf Raven*	TFLite/ LiteRT	ProtoBuf
predict(..)	✅	✅	✅	✅	✅
encode(..)	✅	✅	❌	❌	❌
INT8	CPU	❌	❌	❌	❌
FP16	CPU	❌	❌	❌	❌
FP32	CPU	CPU/GPU	CPU/GPU	CPU	CPU/GPU

✅ = Supported ❌ = Not supported
*ProtoBuf Raven is only available for custom acoustic models.

V3.0

The V3.0 acoustic and geo models are both available in four backends (TFLite/LiteRT, ProtoBuf, PyTorch and ONNX). ONNX Runtime is part of the base install; TFLite and ProtoBuf require birdnet[tf], and the PyTorch backend requires birdnet[pt].

Model	Acoustic	Acoustic	Acoustic	Acoustic	Geo	Geo	Geo	Geo
Backend	TFLite/ LiteRT	ProtoBuf	PyTorch	ONNX	TFLite/ LiteRT	ProtoBuf	PyTorch	ONNX
predict(..)	✅	✅	✅	✅	✅	✅	✅	✅
encode(..)	✅	✅	✅	✅	❌	❌	❌	❌
INT8	❌	❌	❌	❌	CPU	❌	❌	❌
FP16	CPU	❌	❌	CPU/GPU	CPU	❌	❌	CPU/GPU
FP32	CPU	CPU/GPU	CPU/GPU	CPU/GPU	CPU	CPU/GPU	CPU/GPU	CPU/GPU

✅ = Supported ❌ = Not supported

Load a V3.0 model by passing the version and backend, e.g. birdnet.load("acoustic", "3.0", "onnx") or birdnet.load("geo", "3.0", "tf").

Perch V2

Model	Acoustic	Geo
Backend	ProtoBuf	❌
predict(..)	✅	❌
encode(..)	✅	❌
INT8	❌	❌
FP16	❌	❌
FP32	CPU/GPU	❌

Note: Perch can not be run on Intel macOS systems due to compatibility issues with TensorFlow.

Example usage

The examples use the 2.4 models on the tf backend, which needs pip install birdnet[tf] (or library="litert" on the base install, see Installation).

Identify species within an audio file

import birdnet

model = birdnet.load("acoustic", "2.4", "tf")

predictions = model.predict(
  "example/soundscape.wav",
  # predict only the species from the file
  custom_species_list="example/species_list.txt",
)

predictions.to_csv("example/predictions.csv")

The resulting predictions look like this (excerpt, scores may vary):

input	start_time	end_time	species_name	confidence
/home/.../example/soundscape.wav	00:00:00.00	00:00:03.00	Poecile atricapillus_Black-capped Chickadee	0.814
/home/.../example/soundscape.wav	00:00:03.00	00:00:06.00	Poecile atricapillus_Black-capped Chickadee	0.3084
/home/.../example/soundscape.wav	00:00:06.00	00:00:09.00	Baeolophus bicolor_Tufted Titmouse	0.1864
/home/.../example/soundscape.wav	00:00:09.00	00:00:12.00	Haemorhous mexicanus_House Finch	0.6392
/home/.../example/soundscape.wav	00:00:18.00	00:00:21.00	Cyanocitta cristata_Blue Jay	0.4353
/home/.../example/soundscape.wav	00:00:21.00	00:00:24.00	Cyanocitta cristata_Blue Jay	0.3291
/home/.../example/soundscape.wav	00:00:21.00	00:00:24.00	Haemorhous mexicanus_House Finch	0.1866
...	...	...	...	...

For a more detailed prediction you can take a look at example/predictions.csv.

Predict species for a given location and time

import birdnet

model = birdnet.load("geo", "2.4", "tf")

predictions = model.predict(42.5, -76.45, week=4)

predictions.to_csv("example/location.csv")

The resulting predictions look like this (excerpt, scores may vary; sorted alphabetically):

species_name	confidence
Acanthis flammea_Common Redpoll	0.0442
Accipiter cooperii_Cooper's Hawk	0.0812
Agelaius phoeniceus_Red-winged Blackbird	0.0996
Anas platyrhynchos_Mallard	0.4468
Anas rubripes_American Black Duck	0.11
...	...

The full result is at example/location.csv.

Location of Log File

Each prediction or encoding session writes its own log file to the system temp directory:

• Windows: C:\Users\{user}\AppData\Local\Temp\birdnet_session_*.log
• Linux/MacOS: /tmp/birdnet_session_*.log

Environment variables

Variable	Default	Effect
BIRDNET_APP_DATA	platform app data dir	Where downloaded models, labels and the taxonomy are cached. Set it to a persistent path in ephemeral environments.
BIRDNET_TF_VERBOSE	0	Set to 1 to let TensorFlow's native startup output through. It is hidden by default, because TensorFlow prints its absl::InitializeLog warning and oneDNN notice on every import — and every worker process imports it, so one prediction emits dozens of lines. Diagnostics from a failed import are always shown, whatever this is set to.
BIRDNET_START_METHOD	spawn	Multiprocessing start method: spawn, forkserver or fork. See the documentation before choosing fork.

File formats

The audio models support all formats compatible with the SoundFile library (see here). This includes, but is not limited to, WAV, FLAC, OGG, and AIFF. The flexibility of supported formats ensures that the models can handle a wide variety of audio input types, making them adaptable to different use cases and environments.

• Supported: AIFC, AIFF, AU, AVR, CAF, FLAC, HTK, IRCAM, MAT4, MAT5, MP3, MPC2K, NIST, OGG, OPUS, PAF, PVF, RAW, RF64, SD2, SDS, SVX, VOC, W64, WAV, WAVEX, WVE, XI
• Not supported at the moment: AAC, M4A, WMA

Model formats and execution details

Depending on the model version, this project provides up to four model formats: Protobuf/Raven, TFLite, PyTorch and ONNX. All formats are designed to have identical precision up to 2 decimal places, with differences only appearing from the third decimal place onward.

• TFLite Model (tf): Limited to CPU execution only. Requires birdnet[tf] for the default TFLite interpreter; the LiteRT interpreter (library="litert") is part of the base install where wheels exist.
• ProtoBuf Model (pb): Can be executed on both GPU and CPU. Requires birdnet[tf] (birdnet[and-cuda] for GPU).
• PyTorch Model (pt, V3.0 only): Can be executed on both GPU and CPU. Requires birdnet[pt].
• ONNX Model (onnx, V3.0 only): Can be executed on both GPU and CPU. Part of the base install (CPU build); for GPU replace it with the GPU build after installing, since only one of the two may be installed: pip uninstall onnxruntime && pip install onnxruntime-gpu (pip check will then report the missing onnxruntime, and every later pip install/upgrade of birdnet puts the CPU build back — redo the swap).

Ensure your environment is configured to utilize the appropriate model and available hardware optimally.

License

• Source Code: The source code for this project is licensed under the MIT License.
• Models: The models used in this project are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Please ensure you review and adhere to the specific license terms provided with each model. Note that educational and research purposes are considered non-commercial use cases.

Citation

Feel free to use birdnet for your acoustic analyses and research. If you do, please cite as:

@article{kahl2021birdnet,
  title={BirdNET: A deep learning solution for avian diversity monitoring},
  author={Kahl, Stefan and Wood, Connor M and Eibl, Maximilian and Klinck, Holger},
  journal={Ecological Informatics},
  volume={61},
  pages={101236},
  year={2021},
  publisher={Elsevier}
}

Funding

Our work in the K. Lisa Yang Center for Conservation Bioacoustics is made possible by the generosity of K. Lisa Yang to advance innovative conservation technologies to inspire and inform the conservation of wildlife and habitats.

The development of BirdNET is supported by the German Federal Ministry of Research, Technology and Space (FKZ 01|S22072), the German Federal Ministry for the Environment, Climate Action, Nature Conservation and Nuclear Safety (FKZ 67KI31040E), the German Federal Ministry of Economic Affairs and Energy (FKZ 16KN095550), the Deutsche Bundesstiftung Umwelt (project 39263/01) and the European Social Fund.

Partners

BirdNET is a joint effort of partners from academia and industry. Without these partnerships, this project would not have been possible. Thank you!