Output Difference Between Resemble Enhance Web UI and Public Model

[chirag0320]

Hi team,

We're noticing a significant difference in output quality between the Resemble Enhance Web UI and the publicly available Resemble Enhance model (0.0.2.dev240104122303). Despite using identical settings, the web UI consistently produces clearer and louder speech, while our local version does not achieve the same quality.

What We Have Tried:

• Upgraded to the latest Resemble Enhance model.
• Used the same parameters as the Resemble Enhance Web UI (screenshot attached).
• Processed the same audio file both locally and on the web UI.
• Applied various configurations in our local setup, including:

  # Different solver and NFE configurations
  resemble-enhance {audio_folder} {output_folder} --denoise_only --device cpu --solver midpoint --nfe 128 --tau 0.1
  resemble-enhance {audio_folder} {output_folder} --denoise_only --device cpu --solver midpoint --nfe 64 --tau 0.5
  resemble-enhance {audio_folder} {output_folder} --device cpu --denoise_only --solver midpoint --nfe 64 --tau 0.1
  resemble-enhance {audio_folder} {output_folder} --denoise_only --device cpu --tau 0.0 --solver rk4 --nfe 264

• Further processed the output with FFmpeg loudness normalization, but the clarity still does not match.

Our Questions:

1. Is the public model identical to the one used in the Resemble Enhance Web UI and API?
2. Are there hidden processing steps or additional enhancements on the web UI that are not available in the open-source model?
3. What can we do to match the web UI’s output?

Any guidance would be greatly appreciated!

Thanks,
Chirag