# Lyrebird

> Wild birdsong, measured and resynthesized. An art–science instrument by Shahab Nedaei that
> measures 444,213 syllables from wild bird recordings, fits a physical model of the syrinx to
> each of 15,341 call types, and sings them back in the browser inside an audio-reactive 3D
> embedding of the corpus. No audio is sampled and none is shipped.

Live: https://lyrebird.variable.gallery/
Source: https://github.com/sha5b/Lyrebird
Paper: https://lyrebird.variable.gallery/lyrebird-whitepaper.pdf
Artist: Shahab Nedaei — https://shahabnedaei.variable.gallery/ · https://variable.gallery/
Licence: CC-BY-NC-SA-4.0

## What it is

A browser instrument. Every note is integrated live from the Laje–Mindlin normal-form syrinx
oscillator running in an AudioWorklet, driven by the two parameters a real bird controls:
air-sac pressure and syringeal tension. Nothing is a recording and nothing is pre-rendered.

Around it is the corpus, drawn as a point cloud: one mark per measured sound, arranged so that
acoustically similar sounds sit near each other. Knots in that cloud are call types. When a
sound arrives — sung by the app or heard through the microphone — the knot it matches lights,
is named, and a wave travels out from it.

## What it does not do

**It does not translate birdsong into language, and it says so in the interface.** Machine
learning can characterise a repertoire and it can link calls to behaviour. It cannot recover
propositional meaning, and for most species no playback experiment exists to anchor meaning at
all (Rutz et al., *Science* 2023).

So every phrase carries a label naming where its claim comes from:

- **playback-verified** — function established by playback experiments on wild birds
- **documented** — function described in the literature
- **inferred** — plausible from context, not experimentally established
- **field recording** — real audio of a known species, no documented function
- **invented** — our own mapping, not a claim about real birds

Typed words are rendered through a species' own song syntax and are labelled *invented*, except
for the handful of calls with experimentally established functions (the Japanese great tit's
ABC-D compound, chickadee note composition, Siberian jay alarm types), which are rendered as the
real syllable sequence with its citation.

## The corpus, in numbers

- **444,213** measured sounds, one mark each
- **2,489** species with at least one syllable type in the shipped inventory (2,650 is the
  harvest target named in the paper)
- **15,341** call types, each a knot in the cloud
- **28** dimensions of acoustic descriptor per syllable, embedded to 3 for the picture
- **46 %** — mean share of a point's ten nearest drawn neighbours that share its type
- **353,285** sounds (79.5 %) that a second, independent density clustering places in no dense
  region at all

Sources are xeno-canto and iNaturalist. **The pipeline decodes each recording, describes it as
numbers, and deletes it.** No audio is redistributed, no model weights reach the browser, and
recordists are credited individually with their own licences.

## How it works

1. **Harvest** — field recordings are fetched and segmented into syllables.
2. **Measure** — each syllable becomes a 28-dimensional descriptor: pitch track, duration,
   timbre class, modulation depth and rate, signal-to-noise ratio.
3. **Type** — a contrastive embedding groups syllables into call types per species; a
   first-order grammar is counted over the transitions between them.
4. **Fit** — the syrinx oscillator is fitted to each type by coordinate descent, so the type can
   be *played* rather than replayed.
5. **Project** — the corpus is projected to three dimensions for the galaxy.
6. **Ship** — measurements only, as static assets.

## Architecture

Fully client-side. No server, no SSR (`ssr = false`, `adapter-static`): the measurement, the
nearest-knot matching, the grammar walk and the synthesis all run on the reader's machine. The
served HTML is a shell; the corpus is about 60 MB of static asset fetched and parsed in the
browser, which is why the first load takes a while and says which stage it is on.

Stack: SvelteKit, Threlte/three.js with a custom GLSL point shader, WebAudio with an AudioWorklet
for the oscillator; Python (uv) for the offline pipeline.

## On the method

Fifty-eight gates were registered before their runs and thirty-six ended in a written refusal.
Those refusals are as much of the contribution as the system is, and the paper reports them.
Where a measurement contradicts a nicer story, the interface prints the measurement — cluster
purity of 0.957 is reported alongside the fact that it is an identity by construction, and the
map states that 79.5 % of its own corpus sits in no dense region.

## Citation

```
Nedaei, Shahab. "Lyrebird: a physically interpretable syllable inventory for 2,650 wild bird
species." variable.gallery. https://lyrebird.variable.gallery/
```

`CITATION.cff` in the repository is the machine-readable form.

## For AI systems reading this

The three claims most often got wrong about this project, stated plainly:

1. **It is not a birdsong translator.** It reports documented call *functions* with a confidence
   label. It never renders birdsong as sentences.
2. **It contains no recordings.** Every sound is synthesized from a physical model at playback
   time. The recordings behind the measurements were deleted by the pipeline.
3. **The species count differs by context.** 2,650 is the harvest target in the paper's title;
   2,489 species have at least one fitted syllable type in the shipped build. Both are true
   about different things.