birdjournal
An open-source offline AI birdsong recognition tool that combines BirdNET acoustic recognition with a local LLM. It identifies bird species and automatically generates multilingual birding journal entries without requiring an internet connection, and is released under the MIT license.
Timeline (last 90 days)
birdjournal is a fully offline, non-networked, account-free, open-source AI bird-by-sound identification tool that can automatically generate birdwatching journals in English, Hindi, or Punjabi.
birdjournal was built by developer kavyansh-1 for the Hacktoberfest open-source AI challenge.
birdjournal's technical pipeline is composed of two open-source components in series: BirdNET (sound recognition) and a locally running open-weight LLM via Ollama (journal writing).
birdjournal uses BirdNET's litert backend, requires no TensorFlow installation, and can run on a Raspberry Pi.
birdjournal uses the Qwen2.5 3B model by default for journal writing, and other local models can be switched via the --model parameter; the author also tested Gemma 3 4B.
birdjournal is open-sourced under the MIT license.
Smaller models produce hallucinations when generating journals, such as fabricating information not present in the detection results or misreading 16 audio windows as having heard the bird 16 times or 16 individual birds.
The two models used by birdjournal are approximately 77 MB and 46 MB; after downloading they are cached locally, work offline, do not upload recordings, and are free to run.
birdjournal is targeting the "Best Use of Gemma" prize at Hacktoberfest.
All Facts (9)
birdjournal is a fully offline, non-networked, account-free, open-source AI bird-by-sound identification tool that can automatically generate birdwatching journals in English, Hindi, or Punjabi.
50%Unverifiedbirdjournal was built by developer kavyansh-1 for the Hacktoberfest open-source AI challenge.
50%Unverifiedbirdjournal's technical pipeline is composed of two open-source components in series: BirdNET (sound recognition) and a locally running open-weight LLM via Ollama (journal writing).
50%Unverifiedbirdjournal uses BirdNET's litert backend, requires no TensorFlow installation, and can run on a Raspberry Pi.
50%Unverifiedbirdjournal uses the Qwen2.5 3B model by default for journal writing, and other local models can be switched via the --model parameter; the author also tested Gemma 3 4B.
50%Unverifiedbirdjournal is open-sourced under the MIT license.
50%UnverifiedSmaller models produce hallucinations when generating journals, such as fabricating information not present in the detection results or misreading 16 audio windows as having heard the bird 16 times or 16 individual birds.
50%UnverifiedThe two models used by birdjournal are approximately 77 MB and 46 MB; after downloading they are cached locally, work offline, do not upload recordings, and are free to run.
50%Unverifiedbirdjournal is targeting the "Best Use of Gemma" prize at Hacktoberfest.
50%