El1as42 · Oct 26, 2026 · Oct 23, 2026 · Oct 23, 2026 · Oct 23, 2026 · Oct 23, 2026
diff --git a/.gitignore b/.gitignore
@@ -22,7 +22,7 @@ parity_results*/
 *.engine
 !docs/models/*.onnx
 
-# currently online streamlit app
+# local deployment files, not part of the repository
 online/
 
 # mkdocs

diff --git a/.streamlit/config.toml b/.streamlit/config.toml
diff --git a/NOTICE.md b/NOTICE.md
@@ -18,8 +18,8 @@ authoritative.
   treats models trained or fine-tuned with it as AGPL-3.0, so the BirdBox model
   weights are AGPL-3.0 in every format (`.pt`, `.onnx`, `.tflite`, `.engine`).
   Running BirdBox together with Ultralytics or the BirdBox weights, distributing
-  it in that form, or offering it to users over a network (for example the
-  Streamlit app) must therefore comply with AGPL-3.0. In particular, if you
+  it in that form, or offering it to users over a network must therefore
+  comply with AGPL-3.0. In particular, if you
   modify that combined work and let users interact with it over a network, you
   must offer them its corresponding source code (AGPL-3.0, section 13). The
   weights stay AGPL-3.0 even when they run without the `ultralytics` package,
@@ -40,10 +40,11 @@ them.
 |---|---|---|
 | Everything authored by BirdNET-Team and contributors not listed below (`src/`, `tests/`, `taxonomies/*.py`, `install.py`, `run_*.sh`/`.ps1`/`.bat`, `docs/`, configuration) | MIT | Source files carry `SPDX-License-Identifier: MIT` headers. |
 | `mel_filterbank()` in `src/inference/onnx_export.py` | ISC | Adapted from `librosa.filters.mel` (librosa 0.11.0), marked with `SPDX-Snippet` tags. Rest of the file is MIT. Notice [below](#librosa). |
-| Inferno colormap table in `docs/javascripts/onnx_spectrogram.js` | CC0-1.0 | 8-bit samples of the inferno colormap by Nathaniel J. Smith and Stéfan van der Walt, released under CC0 and shipped with matplotlib. |
+| Inferno, magma and viridis colormap tables in `docs/javascripts/demo/colormaps.js` | CC0-1.0 | 8-bit samples of the matplotlib colormaps by Nathaniel J. Smith and Stéfan van der Walt, released under CC0. |
 | `docs/models/Just-Bird_fp16.onnx` | AGPL-3.0 | YOLO11 weights trained with Ultralytics, exported to ONNX. |
-| Model weights distributed outside this repository ([TUC-Cloud](https://tuc.cloud/index.php/s/ET4KE4LdSaysSSL), default model of the Streamlit app) | AGPL-3.0 | YOLO11 weights trained with Ultralytics. |
-| `tests/test.wav`, `docs/audio/sample_15s.wav`, and the spectrogram images `docs/img/example_viz.png`, `docs/img/getting-started/streamlit_ui_screenshot.png`, `docs/img/getting-started/detect_and_merge_transparent.png` | CC-BY-4.0 | Soundscape `UHH_001_S01_20161121_150000` from the BirdNET Hawaii soundscape collection, and spectrograms of recordings from that collection. Attribution [below](#audio-samples). |
+| Model weights distributed outside this repository ([TUC-Cloud](https://tuc.cloud/index.php/s/ET4KE4LdSaysSSL)) | AGPL-3.0 | YOLO11 weights trained with Ultralytics. |
+| `tests/test.wav` and the spectrogram images `docs/img/example_viz.png`, `docs/img/how-it-works/detect_and_merge_transparent.png` | CC-BY-4.0 | Soundscape `UHH_001_S01_20161121_150000` from the BirdNET Hawaii soundscape collection, and spectrograms of recordings from that collection. Attribution [below](#audio-samples). |
+| `docs/audio/BirdNET_soundscape_32kHz.flac` and the screenshots `docs/img/demo_screenshot_light.png`, `docs/img/demo_screenshot_dark.png` | MIT | Example soundscape `birdnet_analyzer/example/soundscape.wav` of [BirdNET-Analyzer](https://github.com/birdnet-team/BirdNET-Analyzer) (Copyright (c) 2024 birdnet-team, MIT License), resampled from 48 kHz to 32 kHz. The same file is the sample of the BirdNET website demo. The screenshots show it in the Demo tab. |
 | `taxonomies/Cornell-to-AviList-mapping.json` | CC-BY-4.0 | Data derived from the AviList v2025 checklist. Attribution [below](#avilist). |
 | Soundscapes fetched by `tests/parity_audio/download.py` | Not MIT | Not part of the repository. xeno-canto recordings under the Creative Commons licenses listed in `tests/parity_audio/recordings.csv`. |
 
@@ -53,12 +54,13 @@ BirdBox installs or loads third-party packages that are not part of this
 repository and keep their own licenses: the Python packages in
 `requirements.txt`, `model_format_requirements/` and `install.py` (notably
 `ultralytics`, AGPL-3.0), and the scripts the documentation site loads from
-CDNs (see `extra_javascript` in `mkdocs.yml`).
+CDNs (see `extra_javascript` in `mkdocs.yml`, and ONNX Runtime Web in
+`docs/javascripts/demo/ort-worker.js` for the Demo tab).
 
 ## Audio samples
 
-`tests/test.wav`, `docs/audio/sample_15s.wav` and the spectrogram images listed
-above are taken from or made from:
+`tests/test.wav` and the spectrogram images listed above are taken from or made
+from:
 
 > Amanda Navine, Stefan Kahl, Ann Tanimoto-Johnson, Holger Klinck, Patrick Hart
 > (2022). *A collection of fully-annotated soundscape recordings from the Island
@@ -69,10 +71,9 @@ Licensed under [Creative Commons Attribution 4.0 International (CC BY 4.0)](http
 
 Changes: `tests/test.wav` is recording `UHH_001_S01_20161121_150000`, converted
 from FLAC to 16-bit mono WAV at 32 kHz. The samples are identical to the original.
-`docs/audio/sample_15s.wav` is the first 15 seconds of `tests/test.wav`.
-`docs/img/example_viz.png` and `docs/img/getting-started/streamlit_ui_screenshot.png`
-show spectrograms of `tests/test.wav` with BirdBox detections drawn on top.
-`docs/img/getting-started/detect_and_merge_transparent.png` shows spectrogram
+`docs/img/example_viz.png` shows a spectrogram of `tests/test.wav` with BirdBox
+detections drawn on top.
+`docs/img/how-it-works/detect_and_merge_transparent.png` shows spectrogram
 excerpts of a recording from the collection, annotated to illustrate the
 detect-and-merge policy.
 

diff --git a/README.md b/README.md
@@ -21,18 +21,20 @@ Everything you need, including installation instructions, an interactive demo, m
 
 ## Quick Links
 
-- [Installation](https://birdnet-team.github.io/BirdBox/getting-started/installation/) - set up the environment
+- [Installation](https://birdnet-team.github.io/BirdBox/installation/) - set up the environment
 - [Data Flow](https://birdnet-team.github.io/BirdBox/data/fileflows/) - pipeline description
 - [Models and Metrics](https://birdnet-team.github.io/BirdBox/models-and-metrics/overview/) - list of models with corresponding metrics
 - [CLI Reference](https://birdnet-team.github.io/BirdBox/cli/workflows/) - command line interface
 - [API Reference](https://birdnet-team.github.io/BirdBox/api/api-type/) - application programming interface
 
-## Interactive Demo
+## Demo
 
-Try out the [Interactive Demo](https://birdnet.cornell.edu/birdbox/) or browse the [Demo Documentation](https://birdnet-team.github.io/BirdBox/getting-started/demo_streamlit/).
-If everything is working as expected, the web interface will look like this:
+Try the [Demo](https://birdnet-team.github.io/BirdBox/demo/) in your browser: it runs the Just-Bird model on your device, with no installation and no upload. Load a recording, tune the confidence and phrase gap, review and edit the boxes, and export them in the BirdBox output formats.
 
-![Streamlit app screenshot](docs/img/getting-started/streamlit_ui_screenshot.png)
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="docs/img/demo_screenshot_dark.png">
+  <img src="docs/img/demo_screenshot_light.png" alt="The BirdBox Demo with the sample soundscape">
+</picture>
 
 ## Citation
 
@@ -41,7 +43,7 @@ Feel free to use BirdBox for your acoustic analyses and research. If you do, ple
 ```bibtex
 @software{Schlosser_BirdBox,
     author = {Schlosser, Elias and Kahl, Stefan and Eibl, Maximilian},
-    license = {MIT},
+    license = {MIT (code), AGPL-3.0 (project as a whole)},
     title = {{BirdBox}},
     url = {https://github.com/birdnet-team/BirdBox}
 }
@@ -53,7 +55,7 @@ BirdBox's own source code is licensed under the [MIT License](LICENSE).
 
 BirdBox is built on [Ultralytics YOLO11](https://github.com/ultralytics/ultralytics), which is licensed under the [AGPL-3.0](LICENSE-AGPL-3.0). Ultralytics also treats models trained or fine-tuned with it as AGPL-3.0, so the BirdBox model weights (`.pt` and every exported format) are AGPL-3.0 as well. As a result, the project as a whole is subject to the AGPL-3.0:
 
-- **Using BirdBox with Ultralytics or the BirdBox models** (running, redistributing, or offering it over a network, such as the Streamlit app) has to comply with the AGPL-3.0. If you modify it and let users interact with it over a network, you must offer them its source code.
+- **Using BirdBox with Ultralytics or the BirdBox models** (running, redistributing, or offering it over a network) has to comply with the AGPL-3.0. If you modify it and let users interact with it over a network, you must offer them its source code.
 - **Reusing BirdBox's own code** on its own, in a project that uses neither Ultralytics nor the BirdBox models, is possible under the MIT terms.
 
 A few files are under other licenses (a function adapted from librosa, xeno-canto audio samples, AviList-derived taxonomy data). See [NOTICE.md](NOTICE.md) for the full breakdown.

diff --git a/docs/api/api-type.md b/docs/api/api-type.md
@@ -5,7 +5,7 @@ BirdBox is a **Python library**. There is no server, no HTTP endpoint, and no RE
 from inference.detect_birds import BirdCallDetector   # ← local Python import
 ```
 
-If you are looking for a point-and-click interface or want to run detection without writing code, use the [CLI Reference](../cli/workflows.md) or the [Streamlit demo](../getting-started/demo_streamlit.md) instead.
+If you are looking for a point-and-click interface or want to run detection without writing code, use the [CLI Reference](../cli/workflows.md) or the [Demo](../demo.md) instead.
 
 ---
 
@@ -31,4 +31,4 @@ sys.path.insert(0, "src")   # only needed when running outside the installed pac
 from inference.detect_birds import BirdCallDetector
 ```
 
-See [Installation](../getting-started/installation.md) for the full setup instructions.
+See [Installation](../installation.md) for the full setup instructions.
diff --git a/docs/audio/BirdNET_soundscape_32kHz.flac b/docs/audio/BirdNET_soundscape_32kHz.flac
diff --git a/docs/audio/sample_15s.wav b/docs/audio/sample_15s.wav
diff --git a/docs/cli/detect-birds.md b/docs/cli/detect-birds.md
@@ -47,7 +47,7 @@ Detect bird calls in arbitrary-length audio files using a trained YOLO model. Pr
 | `--verbose` | flag / off | No | Print per-file processing details, clip-level progress bars, and a detection summary. Default is a single file-level progress bar. |
 
 !!! danger "Environment must match the model format"
-    Each model format (`.pt`, `.onnx`, `.tflite`, `.engine`) requires a different Python environment. Using the wrong environment raises an import error or silently degrades results. Run `python install.py --model-format <FORMAT>` once for each format you want to use. See [Install Parameters](../getting-started/installation.md#install-parameters) for the full table.
+    Each model format (`.pt`, `.onnx`, `.tflite`, `.engine`) requires a different Python environment. Using the wrong environment raises an import error or silently degrades results. Run `python install.py --model-format <FORMAT>` once for each format you want to use. See [Install Parameters](../installation.md#install-parameters) for the full table.
 
 ## Parameter Deep-Dives
 

diff --git a/docs/data/audio-and-model.md b/docs/data/audio-and-model.md
@@ -46,7 +46,7 @@ Pretrained releases on [TUC-Cloud](https://tuc.cloud/index.php/s/ET4KE4LdSaysSSL
 Each format requires its own Python environment. Install the correct runtime before switching formats.
 
 !!! danger "Environment must match the model format"
-    Loading a `.tflite`, `.onnx`, or `.engine` model from a `.pt` environment will fail with an import error or silently produce incorrect results. Run `python install.py --model-format <FORMAT>` to set up the right environment. See [Install Parameters](../getting-started/installation.md#install-parameters) for the full format-to-runtime table.
+    Loading a `.tflite`, `.onnx`, or `.engine` model from a `.pt` environment will fail with an import error or silently produce incorrect results. Run `python install.py --model-format <FORMAT>` to set up the right environment. See [Install Parameters](../installation.md#install-parameters) for the full format-to-runtime table.
 
 !!! info "Format parity"
     Detection quality is validated across all formats by running inference on the same audio file and comparing results against the `.pt` baseline. See [Just-Bird format parity](../models-and-metrics/just-bird-model-types.md) and [All-In-One format parity](../models-and-metrics/all-in-one-model-types.md) for the current scores.
@@ -59,7 +59,7 @@ Each format requires its own Python environment. Install the correct runtime bef
 ## Species mapping (`--species-mapping`)
 
 The mapping name must match the label space the model was trained on. It is **not** inferred from the weights filename. You pass it explicitly.
-For specific mappings see [Species Mapping](https://github.com/birdnet-team/BirdBox/blob/main/src/config.py#L19){ target="_blank" rel="noopener noreferrer" }.
+For specific mappings see [Species Mapping](https://github.com/birdnet-team/BirdBox/blob/main/src/config.py#L17){ target="_blank" rel="noopener noreferrer" }.
 
 Allowed values for the species-mapping parameter:
 

diff --git a/docs/demo.md b/docs/demo.md
@@ -0,0 +1,96 @@
+---
+title: Demo
+template: demo.html
+hide:
+  - navigation
+  - toc
+  - footer
+---
+
+## About this demo
+
+The demo runs the self-contained Just-Bird graph `docs/models/Just-Bird_fp16.onnx` with [ONNX Runtime Web](https://onnxruntime.ai/docs/tutorials/web/){ target="_blank" rel="noopener noreferrer" } in your browser. The graph computes the PCEN spectrogram, runs the YOLO detector on 3 s windows every 1.5 s and applies NMS, just like [`detect_birds.py`](cli/detect-birds.md) with the Just-Bird model. There are no species labels: every box is a bird vocalization (`bird`).
+
+**Instant sliders.** The model runs once per recording, at the lowest confidence of the slider (0.05). The demo asks the graph for its per-window boxes (`raw_all`) and repeats the graph's window fusion in JavaScript: window ownership and weighted box fusion, joining calls into phrases, removing nested boxes. NMS only lets higher-scoring boxes suppress lower ones, so the boxes for any confidence and phrase gap are the same as a graph run with those values.
+
+**Spectrogram.** The viewer shows an STFT spectrogram by default. In its display settings, *PCEN* switches to the PCEN mel spectrogram BirdBox detects on.
+
+**Review.** Boxes can be verified, moved, resized, removed and drawn. The exports use the BirdBox [output formats](data/outputs.md): the JSON of `--output-format json-with-algorithm-metadata`, the simplified CSV and the Raven selection table, plus a CSV with every column of the table.
+
+!!! warning "Browser limits"
+    - The model is downloaded once (about 35 MB) and then comes from the browser cache.
+    - Audio is decoded locally, resampled to 32 kHz mono and truncated to **10 minutes**.
+    - The browser's resampler is not librosa's, so files that are not recorded at 32 kHz can give slightly different boxes than the CLI.
+    - ONNX Runtime runs single-threaded here: about 10 s for 2 minutes of audio on a laptop. Safari can be slower than Chromium or Firefox.
+    - Use the [CLI](cli/detect-birds.md) for long files and batch jobs.
+
+The sample soundscape is the example recording of [BirdNET-Analyzer](https://github.com/birdnet-team/BirdNET-Analyzer){ target="_blank" rel="noopener noreferrer" } (`birdnet_analyzer/example/soundscape.wav`, MIT License), resampled to 32 kHz. The [BirdNET website demo](https://birdnet.cornell.edu/demo/){ target="_blank" rel="noopener noreferrer" } uses the same file. Recordings loaded by number come from [Xeno-canto](https://xeno-canto.org){ target="_blank" rel="noopener noreferrer" } under the license of each recording.
+
+## Adding models
+
+The demo runs the first model in `docs/models/manifest.json` with `"merge_method": "fusion"`. Only window-fusion exports work: the instant sliders need the graph's per-window boxes (`raw_all`) and the JavaScript port of its window fusion (`docs/javascripts/demo/fusion.js`). The export uses window fusion for Just-Bird, Hawaii and Southern-Sierra-Nevada (`--merge auto`); `--merge fusion` forces it for the other models.
+
+1. Export the model.
+
+    === "Linux / macOS"
+        ```bash
+        python src/inference/onnx_export.py \
+            --output-path docs/models/Just-Bird_fp16.onnx \
+            --precision fp16
+        ```
+
+    === "Windows (PowerShell)"
+        ```powershell
+        python src/inference/onnx_export.py `
+            --output-path docs/models/Just-Bird_fp16.onnx `
+            --precision fp16
+        ```
+
+    === "Windows (CMD)"
+        ```cmd
+        python src/inference/onnx_export.py ^
+            --output-path docs/models/Just-Bird_fp16.onnx ^
+            --precision fp16
+        ```
+
+    `--precision fp16` stores the network weights in fp16, which makes the file about a third smaller (35 MB instead of 53 MB for Just-Bird). Feature extraction and merging stay in fp32. On 13 test recordings, 99.4% of the fp32 graph's boxes were reproduced (mean IoU 0.996). The export prints the merge it baked in (`Merge (baked in)`) and the suggested defaults (`Default conf`, `Default song_gap`), which it also stores in the ONNX metadata.
+
+2. Place the `.onnx` file under `docs/models/` if the export path was elsewhere.
+
+3. Add an entry to `docs/models/manifest.json`, first in the list if the demo should run it.
+
+    ```json
+    {
+      "models": [
+        {
+          "file": "Just-Bird_fp16.onnx",
+          "label": "Just-Bird_fp16.onnx",
+          "merge_method": "fusion",
+          "conf": 0.15,
+          "phrase_gap": 0.05,
+          "species_mapping": "Just-Bird",
+          "names": { "0": "bird" }
+        }
+      ]
+    }
+    ```
+
+    | Field | Meaning |
+    | :--- | :--- |
+    | `file` | The ONNX file, relative to `docs/models/` |
+    | `label` | Name shown at the top of the demo |
+    | `merge_method` | `fusion`, as printed by the export. Other entries are skipped. |
+    | `conf` | Where the confidence slider starts: the printed `Default conf` (0.15 for Just-Bird, 0.2 for the other models) |
+    | `phrase_gap` | Where the phrase gap slider starts: the printed `Default song_gap` (0.05 for Just-Bird, 0 for Hawaii and Southern-Sierra-Nevada) |
+    | `species_mapping` | Written to `model_config` of the exported JSON |
+    | `names` | Class labels for the exports, by class index (ONNX Runtime Web does not expose ONNX metadata) |
+
+4. Check that the demo's fusion still matches the graph:
+
+    ```bash
+    cd tests/js && npm install
+    node demo_fusion_parity.mjs ../../docs/models/Just-Bird_fp16.onnx
+    ```
+
+    The script compares `fusion.js` with the graph's own detections for 16 slider settings on `tests/test.wav` and fails on any difference. Run it after exporting a new model or changing the fusion in `src/inference/onnx_export.py`.
+