Installation¶

Install in 60 seconds. Pick the PyTorch flavour that matches your hardware and you're done.

TL;DR¶

pip install strands-omnivoice

This pulls in omnivoice>=0.1.5, strands-agents, soundfile, and numpy. It does not force a particular PyTorch flavour — pick yours below.

PyTorch by Hardware¶

NVIDIA (CUDA)Apple Silicon (MPS)CPU only

pip install torch==2.8.0+cu128 torchaudio==2.8.0+cu128 \
    --extra-index-url https://download.pytorch.org/whl/cu128

pip install torch==2.8.0 torchaudio==2.8.0

pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu

Generation is slow on CPU. Use only for development or low-volume jobs.

Verifying¶

python -c "
from strands_omnivoice import omnivoice_sysinfo
import json
print(json.dumps(omnivoice_sysinfo(), indent=2, default=str))
"

You should see your detected device, torch version, and omnivoice_version.

Optional Extras¶

pip install 'strands-omnivoice[demo]'   # gradio (omnivoice_demo_serve)
pip install 'strands-omnivoice[dev]'    # pytest, ruff
pip install 'strands-omnivoice[all]'    # everything

Troubleshooting¶

`ImportError: cannot import name 'OmniVoice'`¶

omnivoice failed to install. Check the upstream's installation guide for your platform.

`RuntimeError: MPS device not available`¶

You're on macOS without Apple Silicon, or Python wasn't built for arm64. Pin device explicitly:

export STRANDS_OMNIVOICE_DEVICE=cpu

Out of memory on first generation¶

OmniVoice loads ~1 GB of weights. Free GPU memory before calling omnivoice_load_model, or pass device="cpu".