PersonaPlex-MLX

MLX inference for NVIDIA PersonaPlex on Apple Silicon.

This package supports:

• Realtime local mode (personaplex_mlx.local)
• Realtime web mode (personaplex_mlx.local_web)
• Offline WAV-to-WAV mode (personaplex_mlx.offline)

Console entrypoints are also installed: personaplex-local, personaplex-local-web, personaplex-offline.

Requirements

• Apple Silicon Mac
• Python 3.12
• Hugging Face access to nvidia/personaplex-7b-v1

Install:

pip install -e .

Model Access

Accept the model license: https://huggingface.co/nvidia/personaplex-7b-v1

Set your token:

export HF_TOKEN=<your_token>

Quickstart

Launch realtime web mode (recommended first):

python -m personaplex_mlx.local_web \
  -q 4 \
  --voice NATF2 \
  --text-prompt "You enjoy having a good conversation."

Open http://localhost:8998 in your browser.

Realtime local terminal mode:

python -m personaplex_mlx.local \
  -q 4 \
  --voice NATF2 \
  --text-prompt "You enjoy having a good conversation."

Offline inference:

python -m personaplex_mlx.offline \
  --voice NATF2 \
  --text-prompt "You are a wise and friendly teacher. Answer questions or provide advice in a clear and engaging way." \
  --input-wav input.wav \
  --output-wav output.wav \
  --output-text output.json \
  --seed 42424242

Voices

Built-in voice IDs:

• NATF0 NATF1 NATF2 NATF3
• NATM0 NATM1 NATM2 NATM3
• VARF0 VARF1 VARF2 VARF3 VARF4
• VARM0 VARM1 VARM2 VARM3 VARM4

--voice NATF2 resolves to NATF2.pt from the downloaded voices/ bundle.

Notes

• First run downloads model assets from Hugging Face.
• Local and web clients are barebone and do not include echo cancellation. Use headphones to avoid feedback.

Attribution

This project is an MLX port of NVIDIA PersonaPlex for Apple Silicon.

• NVIDIA PersonaPlex repo: https://github.com/NVIDIA/personaplex
• PersonaPlex model card: https://huggingface.co/nvidia/personaplex-7b-v1

Citation

If you use PersonaPlex in research, cite:

@misc{roy2026personaplexvoicerolecontrol,
  title={PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models},
  author={Rajarshi Roy and Jonathan Raiman and Sang-gil Lee and Teodor-Dumitru Ene and Robert Kirby and Sungwon Kim and Jaehyeon Kim and Bryan Catanzaro},
  year={2026},
  eprint={2602.06053},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2602.06053}
}

License

• Code: MIT (LICENSE)
• Model weights: NVIDIA Open Model License (via Hugging Face model card)