Improve Installation steps for MacOS users

[yyovil]

On Apple Silicon macOS, the current requirements.txt blocks a native install because it pins:

kaldi_native_fbank==1.15

That version does not provide an arm64 macOS wheel, so pip/uv pip resolution fails on native Apple Silicon environments.

Repro

Machine:

• macOS on Apple Silicon (arm64)
• Python 3.11

Using the current requirements file:

pip install -r requirements.txt

or

uv pip install --python .venv/bin/python -r requirements.txt

Resolution fails because kaldi-native-fbank==1.15 has no matching macosx_arm64 wheel.

What worked

Keeping the rest of the stack effectively the same, a native install worked after changing only the Kaldi fbank package to a newer release with Apple Silicon wheels:

torch==2.1.0
numpy==1.26.1
kaldiio==2.18.0
kaldi_native_fbank==1.22.3
soundfile==0.12.1
textgrid

Then FireRedVAD loaded and ran natively on Apple Silicon, including:

• importing fireredvad
• constructing FireRedStreamVad
• loading Stream-VAD/model.pth.tar
• running detect_frame(...)

Suggested fix

Please consider updating or loosening the pin in requirements.txt, for example:

kaldi_native_fbank==1.22.3

or a compatible bounded range if there is a minimum known-good version.

Related note

I also hit a separate packaging issue where kaldiio==2.18.0 still imports pkg_resources, so environments may need setuptools<81 until kaldiio is updated. But the main Apple Silicon installation blocker was the old kaldi_native_fbank==1.15 pin.

[yyovil]

this could be useful as well for people looking for CoreML model.

[kaituoxu]

You can use any version of kaldi_native_fbank >=1.15 :)

[eschmidbauer]

try the onnx implementation

[yyovil]

this could be useful as well for people looking for CoreML model.

This worked

[kaituoxu]

@yyovil Hi bro, would like to PR CoreML related runtime code to this repo? Need your skill!