# Lyrebird for M5Stack

> A pocket dawn chorus. This page flashes an M5Stack Fire or CoreS3 with firmware that
> synthesizes 2423 songbird species on the device, from a physical model of the avian
> syrinx. No samples, no SD card, no network: the corpus is compiled into the binary as
> tables and every note is integrated on a 240 MHz microcontroller.

Live: https://m5lyrebird.variable.gallery/
Source: https://github.com/sha5b/M5Stack-Lyrebird
Parent project: https://lyrebird.variable.gallery/ (the browser instrument this is a port of)
Artist: Shahab Nedaei — https://shahabnedaei.variable.gallery/ · https://variable.gallery/
Licence: CC-BY-NC-SA-4.0

## The three things most often got wrong

1. **It plays no samples.** There is no audio in the firmware. Each note is integrated from
   the Mindlin–Laje syrinx oscillator, sample by sample, driven by the two parameters a real
   bird controls: air-sac pressure and syringeal tension. What is compiled in is the
   *description* of a syllable — pitch contour, duration, timbre class, envelope — and the
   offline calibration tables the model needs.
2. **It is not a streaming device.** No Wi-Fi, no SD card, no phone. The board is a
   standalone instrument once flashed; the browser is only the flasher.
3. **The picture on the screen is not a spectrogram.** It is a camera inside the corpus,
   following whoever is singing. A spectrogram is still in the source behind a build flag,
   because it answers a different question — what pitch is sounding now.

## The clip

https://m5lyrebird.variable.gallery/lyrebird-boards.mp4 — 28 seconds of a Fire and a CoreS3
running this firmware side by side, with sound. Recorded on a desk; nothing in it is
rendered or simulated.

## What you get

Two boards, two binaries, one behaviour:

- **M5Stack Fire** — ESP32, audio out of the built-in DAC on GPIO25, DMA-paced at 22050 Hz,
  three physical buttons.
- **M5Stack CoreS3** — ESP32-S3, 16-bit PCM into the AW88298 amplifier, three touch zones
  drawn along the bottom of the screen instead of buttons.

It boots into **all birds**: twelve species at a time, two individuals of each, arriving as a
Poisson process at about 34 songs a minute, with a 25 % chance a conspecific answers. The
twelve turn over every few minutes, so the slot is a tour of the corpus rather than a fixed
dozen. Stepping the dial goes down to one species at a time — 2424 positions in all.

## The corpus, in numbers

- **2423** species
- **12724** syllables, each with its own pitch contour
- **2442** songs — authored syllable sequences with their gaps
- **4.0 MB** of generated C, which lands as **1.79 MB** of flash and no RAM: the tables are
  read where they lie
- **8** voices maximum, **16–64** Euler substeps per sample

The data comes from the parent project's inventory: twelve species authored by hand, the rest
fitted by its pipeline. The recordings behind it (xeno-canto, iNaturalist) are described as
numbers and deleted there; no audio is redistributed by either project.

## Flashing it

Web Serial, so Chrome, Edge or Opera. The page reads the chip off the board and picks the
matching image. A Fire resets itself into its bootloader over the cable. A CoreS3 needs its
side button held for 2–3 seconds the first time, because it has no separate USB-UART bridge —
after that its own USB-Serial-JTAG keeps its identity across a reset and it reflashes like any
other board.

There is also a command line: `pio run -e m5stack-fire -t upload`.

## What it does not claim

Nothing here is a claim about what birdsong means. The parent project states that position at
length and labels every phrase with the evidence behind it; this firmware sings the syllables
and their documented sequences, and does not translate anything.