Resemble Enhance

Resemble Enhance

Use this skill for AI speech denoising and enhancement with the local project at:

E:\resemble-enhance-main

Resemble Enhance improves speech recordings by separating voice from noise and optionally restoring perceived audio quality/bandwidth. It can run either as a Gradio web demo or as a command-line batch processor.

Default Workflow

1. Confirm the input audio path or folder from the user request.
2. Check whether E:\resemble-enhance-main exists.
3. Prefer a project-local virtual environment if one exists; otherwise use the active Python environment.
4. For one-off interactive use, launch the Gradio app:

Set-Location -LiteralPath 'E:\resemble-enhance-main'
python app.py

5. For batch processing, use the installed console command when available:

resemble-enhance in_dir out_dir

Use --denoise_only when the user asks only for noise removal:

resemble-enhance in_dir out_dir --denoise_only

Environment Checks

Before running processing, inspect:

Set-Location -LiteralPath 'E:\resemble-enhance-main'
python --version
python -c "import torch, torchaudio, gradio; print('ok')"

If dependencies are missing, install from the project:

Set-Location -LiteralPath 'E:\resemble-enhance-main'
python -m pip install -r requirements.txt
python -m pip install -e .

Ask before installing large dependencies or downloading model weights. Torch, torchaudio, and model checkpoints may be large.

Input And Output Conventions

• Keep user originals untouched.
• For batch work, create an output folder with a clear name such as enhanced, denoised, or a timestamped directory.
• If the user does not specify an output location, place deliverables under the active workspace outputs directory when available.
• For Windows paths, use PowerShell -LiteralPath for paths containing spaces or non-ASCII characters.

Gradio App Notes

app.py exposes:

• input audio upload
• CFM ODE solver: Midpoint, RK4, Euler
• number of function evaluations
• prior temperature
• optional denoise-before-enhancement checkbox
• outputs for denoised audio and enhanced audio

Use the local web app when the user wants to compare results manually or adjust settings.

Quality Guidance

• For noisy speech, try enhancement with denoising enabled.
• For simple background noise removal, use denoise-only mode first.
• If processing is slow on CPU, check whether CUDA is available; the app automatically uses cuda when torch.cuda.is_available() is true.
• If the output sounds over-processed, reduce enhancement intensity through the Gradio controls or try denoise-only.

Safety

Do not delete source audio. Do not overwrite an existing output directory unless the user explicitly asks.