README.md

metadata

title: Inference Half Duplex
emoji: 🚀
colorFrom: gray
colorTo: purple
sdk: gradio
python_version: '3.11'
app_file: app.py
license: apache-2.0

Inference Half Duplex

This folder is for the minimized ONNX inference demo.

Hugging Face Spaces

This repo can run as a Gradio Space. Create a new Space with the Gradio SDK, then push this repository, including app.py, requirements.txt, packages.txt, runtime.txt, and the models/ directory.

The app downloads anyreach-ai/dualturn-qwen2.5-mimi-0.5B on first use, so the first run can take a while. If you use a CPU Space, leave the device set to cpu; on GPU Spaces the UI defaults to cuda.

Models

• models/pvad_core.onnx: exported Personal VAD core from vad_lstm_personal_light_lstm_fixed.safetensors.
• models/silero_vad.jit: Silero VAD frontend used to compute vad_score for pvad_core.onnx.

pvad_core.onnx expects the Silero VAD score to be computed outside the graph. check_pipeline.py does this with models/silero_vad.jit. Inputs:

• frame_pcm: float32 [batch, 512]
• target_vector: float32 [batch, 16]
• vad_score: float32 [batch, 1]
• h0: float32 [1, batch, 64]
• c0: float32 [1, batch, 64]

Outputs:

• final_prob: target probability after multiplying by vad_score
• raw_prob: raw target/non-target probability
• frame_embed: 16-dim frame embedding
• hn, cn: next LSTM state

Re-export from the source repo:

cd /Utilisateurs/tnguye28/vad-lstm
UV_CACHE_DIR=/tmp/uv-cache NUMBA_CACHE_DIR=/tmp/numba-cache uv run python scripts/export_pvad_core_onnx.py --output /Utilisateurs/tnguye28/inference-half-duplex/models/pvad_core.onnx

Run the pipeline smoke test:

cd /Utilisateurs/tnguye28/inference-half-duplex
UV_CACHE_DIR=/tmp/uv-cache NUMBA_CACHE_DIR=/tmp/numba-cache uv run python check_pipeline.py

Run the Gradio demo:

cd /Utilisateurs/tnguye28/vad-lstm
GRADIO_SERVER_PORT=8787 UV_CACHE_DIR=/tmp/uv-cache NUMBA_CACHE_DIR=/tmp/numba-cache uv run python /Utilisateurs/tnguye28/inference-half-duplex/app.py