De-rustle [ADV]
Overview
De-rustle reduces the noise or rustle generated by a lavalier microphone rubbing or brushing against a person’s clothing during a recording. Lav rustle can vary unpredictably over time and exhibit a wide variety of sonic characteristics from high frequency “brushing” to low frequency “thuds.”
Machine learning in De-rustle
In developing De-rustle, we collected lots of isolated rustle, which we then used to corrupt clean speech and feed to a machine learning algorithm trained to recover the clean speech. When processing, the trained machine learning algorithm separates the clean speech from the background rustle.
Controls
- REDUCTION STRENGTH:
- When set to higher values, the processing more strictly defines what it classifies as dialogue, which can result in more rustle reduction at the cost of reducing speech clarity.
- When set to lower values, the processing more broadly defines what it classifies as dialogue, which potentially allows more rustle but reduces the possibility of speech clarity loss as a result of processing.
- When set to higher values, the processing more strictly defines what it classifies as dialogue, which can result in more rustle reduction at the cost of reducing speech clarity.
- AMBIENCE PRESERVATION [%]: Sometimes the speech separation algorithm will remove background ambience in addition to rustle, which may be undesirable. To remove rustle but retain noise floor/background ambience, the noise floor amount slider should be set to a non-zero value. Higher values will retain more of the noise floor.