agent-translator

Local real-time speech-to-speech translation for Intel AI PCs. Microphone in, translated speech and live captions out — every model runs locally via OpenVINO and PyTorch CPU. No cloud calls.

The pipeline is: audio capture → FSMN-VAD sentence segmentation → streaming Paraformer ASR (with an optional Qwen3-ASR pass on the iGPU) → Hunyuan-1.8B translation on the iGPU, falling back to Opus-MT on CPU → MeloTTS synthesis, pushed to a bundled WebSocket + HTTP web UI.

This package is the reusable core of the local-realtime-translator AI skill.

Install

pip install agent-translator

On Linux, add --extra-index-url https://download.pytorch.org/whl/cpu unless you want the CUDA build of torch — it is several GB and the pipeline does not use it. On Windows this is unnecessary: PyPI's torch wheels are CPU-only there.

torch is pinned to 2.4.1 (newer builds fail to initialize c10.dll on a clean Intel AI PC) and transformers to >=4.57,<4.58 (5.x silently makes Hunyuan return empty translations). Those pins cap the package at Python 3.12.

Use

from agent_translator import config
from agent_translator.live_pipeline import LivePipeline

config.configure(
    assets_root=r"C:\Users\me\.openvino\assets\my-app",
    output_dir=r"C:\Users\me\.openvino\temp\my-app\output",
)

pipeline = LivePipeline(src_lang="zh", tgt_lang="en")
pipeline.start()
try:
    ...  # the web UI is served while the pipeline runs
finally:
    pipeline.stop()

Models are expected to already be on disk under config.MODELS_ROOT; the skill downloads them from ModelScope before starting the pipeline.

Configuration

Runtime locations are host-owned. Set them with config.configure(...) in-process, or with environment variables when the pipeline runs in a child process:

SKILL_ROOT, when set and carrying an nltk_data/ pack, makes that copy win over the downloaded one — that is how a source checkout keeps working offline.

Everything else in agent_translator.config is a tuning constant (sample rates, VAD thresholds, model ids) you can override by assignment before loading.

License

Apache-2.0.