Demo ONNX

Run BirdBox detection entirely in your browser. The graphs under docs/models/ bake in PCEN preprocessing, YOLO inference, NMS, and song reconstruction. No Python install and no server-side GPU queue.

Browser limits

• First model download is large (about 50 MB for Just-Bird fp32). The file is cached by your browser afterward.
• Audio is decoded locally, resampled to 32 kHz, and truncated to 60 seconds.
• Requires a recent browser with WebAssembly. Safari can be slower than Chromium or Firefox.
• This showcase is for quick checks. Use the Streamlit demo or the CLI for long files and batch jobs.

---

Interactive Demo

Select Model

Models are loaded from docs/models/ via manifest.json.

Audio file

Choose file No file chosen

WAV or FLAC recommended. Minimum 3 seconds. Maximum 60 seconds in this demo.

Confidence Threshold

0.18

Song Gap Threshold (seconds)

0.10

Detect Bird Vocalizations Load 15 s sample Download CSV

Loading demo…

Songs 0

Species 0

Avg Confidence —

Start (s)	End (s)	Low (Hz)	High (Hz)	Avg Conf	Max Conf	Merged	Species

---

Adding models

1. Export for the docs demo.

   Linux / macOSWindows (PowerShell)Windows (CMD)

   python src/inference/onnx_export.py \
       --output-path docs/models/Just-Bird_fp32.onnx \
   

   python src/inference/onnx_export.py `
       --output-path docs/models/Just-Bird_fp32.onnx `
   

   python src/inference/onnx_export.py ^
       --output-path docs/models/Just-Bird_fp32.onnx ^
   

   conf and song_gap are required graph inputs. The demo sliders feed them on every run. Suggested defaults are stored in ONNX metadata (default_conf, default_song_gap).

2. Place the .onnx file under docs/models/ if the export path was elsewhere.

3. Add an entry to docs/models/manifest.json.

   {
     "models": [
       {
         "file": "Just-Bird_fp32.onnx",
         "label": "Just-Bird_fp32.onnx",
         "names": { "0": "bird" }
       },
       {
         "file": "Hawaii_fp32.onnx",
         "label": "Hawaii_fp32.onnx",
         "names": { "0": "hawama", "1": "ercfra" }
       }
     ]
   }
   

The dropdown reads that manifest on page load. Put class labels in the optional names map (ONNX Runtime Web does not expose ONNX metadata props).

---

Output columns

Column	Meaning
Start / End	Song bounds in seconds from the start of the (possibly truncated) audio
Low / High	Frequency bounds in Hz
Avg Conf / Max Conf	Mean and peak confidence over merged clip detections
Merged	Number of clip-level boxes folded into the song
Species	Label from model metadata (names)

For the full set of BirdBox file formats outside the browser, see Detection Output Formats.