README.md

metadata

title: LocalVQE Demo
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
python_version: '3.12'
pinned: false
license: apache-2.0
models:
  - LocalAI-io/LocalVQE

LocalVQE Demo

Interactive demo for LocalAI-io/LocalVQE — compact open-source models that clean up a microphone signal on CPU in a single causal pass. The selector offers the joint models (acoustic echo cancellation + noise suppression + dereverberation: v1.3 at 4.8 M parameters, v1.2 at 1.3 M, plus older releases for A/B), and v1.4-AEC (203 K parameters) which removes only the echo — near-end speech, background noise, and room acoustics are kept intact by design.

Pick one of the bundled examples or upload your own (mic, far-end-reference) 16 kHz wav pair. For a pure noise-suppression test, upload silence (or nothing) as the reference. With v1.4-AEC selected, noise-only clips should come back nearly unchanged — that's the intended behaviour.

Built with Gradio. Inference runs through the released GGML C++ engine (the same liblocalvqe.so production users build from the repo) loading the published .gguf files — what you hear is the deployed artifact, not a Python re-implementation. CPU-only; a 10-second clip takes well under a second on the Space's default tier.

Code: github.com/localai-org/LocalVQE · Training: github.com/localai-org/LocalVQE-train.