install.sh expects S16_LE capable recording

[jkrauska]

Issue Type

• Bug report
• Feature request
• Question / Support

Description

[A clear and concise description of the issue or feature request]
install.sh script expects audio device to support S16_LE.
Not all audio inputs support this. eg. BOYA BY-LM40
It only support Signed 24 bit Little Endian in 3bytes. S24_3LE mode.

Steps to Reproduce (for bugs)

1. use install.sh on a pi with a BOYA BY-LM40

Expected behavior

USB audio device works

Actual behavior

install fails

Screenshots

[If applicable, add screenshots to help explain your problem]

Environment

Version Information (build date)

Provide the following line from your system:
current dev release

$ ./birdnet-go
[Paste the output line here, which should look like this:]
BirdNET-Go build date: YYYY-MM-DDTHH:MM:SSZ, using config file: /path/to/config.yaml

System Information

• Device:

  • [x ] Raspberry Pi 5

• OS: Raspberry Pi OS

• OS Version: Raspberry Pi OS Bullseye

• Hardware specs:

  • CPU: [e.g. Raspberry Pi 5]

  • RAM: [e.g. 4GB]

  • Storage type and size: [e.g. 32GB microSD card, 128GB SSD]

• Additional hardware (if applicable):

  • USB Audio device: BOYA LM-40
  • Other peripherals: [e.g. Specific microphone, GPS module]

For Non-Raspberry Pi Linux Users

Have you tested with the latest dev branch container image?

• [x ] Yes
• No
• Not applicable (Windows/macOS user)

If not, first test with the latest dev branch container image and see if the issue is already fixed.

$ docker pull ghcr.io/tphakala/birdnet-go:dev

Additional context

unclear if later code will require SE16_LE.. haven't see arecord calls in go.

Possible Solution

try other formats?

[jkrauska]

on deeper inspection, we assume its S16 in two places

https://github.com/search?q=repo%3Atphakala%2Fbirdnet-go%20malgo.FormatS16&type=code

lets see if the malgo package has the ability to DETECT the encoding.

[tphakala]

BirdNET model expects input audio to be in 16-bit 48khz mono format

[jkrauska]

We should include that somewhere I guess. If I bypass the check, the app works, but it thinks all the birds sounds are humans. I'm going to benchmark how much load dynamically re-encoding takes and try to add some additional error handling if the hardware doesn't support 16-bit.