CİK-SÖR: An Early Warning System for Respiratory Disease from Chicken Sounds
A bioacoustic early warning prototype that extracts chicken vocalisations from poultry house recordings, analyses them with deep learning, and explains a suspected respiratory disease by showing which frequency bands drove the decision.

Problem
Respiratory diseases in poultry houses are usually noticed once the whole flock shows symptoms, and by then the losses have grown. The sounds of sick chickens (wheezing, rales, distorted calls) change much earlier, but nobody can listen to them all day.
CİK-SÖR picks up this change from an ordinary phone recording and raises a "suspected respiratory disease" warning. It does not diagnose; it gives an early signal so a vet can step in sooner. The project is run by a team; I built its working prototype.
Approach
The field recordings were messy: long recordings held only a few seconds of chicken sounds, and some sat under constant hiss. I showed by measurement that the fixed-threshold event detection used in the literature fails on this data, and built a hysteresis-based detector that adapts to each recording.
The extracted sounds are split into 5-second windows and turned into embeddings with the Google Perch v2 bioacoustic model. Because labels exist only at recording level, I trained an attention-based Multiple Instance Learning head; the model learns on its own which moments of a recording drive the decision.
I was strict about leakage in evaluation: recording-level splits, a test set fixed before looking at the data, and grouped cross-validation. To check whether the model really learns chicken sounds rather than recording conditions, I built comparison models that look only at background noise.
Highlights
- Event detection that adapts to each recording and works on noisy, inconsistent field audio
- Frequency band plots that explain the decision, and biomarker deviations from a healthy reference (2–5 kHz band ratio, spectral flatness, HNR)
- Every biomarker tied to a literature source or clearly labelled as a "hypothesis"
- A Gradio interface that works from a browser recording or an uploaded file
- A demo that needs no GPU and runs on CPU on Hugging Face Spaces
- A TDD workflow whose tests guarantee that training and analysis share the same preprocessing functions
Next project
Back to all projects
Yerel-PDF-RAG: A PDF Question-Answering System That Cites Its Sources
A RAG application that answers questions in Turkish about the PDFs you upload, cites every answer with the document and page number, and runs entirely on your own computer.
- Python 3.11
- FastAPI
- Pydantic
- SQLite (FTS5/BM25)