Cleaning Up Noisy Recordings
Trevor Wishart (July 2007)
The 'Cleaning Kit' was originally developed to clean up vocal recordings made in noisy environments.
Typical sources of noise might be ...
The Cleaning Kit uses some existing and some new CDP programs to tackle all of these situations. However,
it is not possible to deal with all possible noise intrusions.
- Relatively Constant (e.g. extractor fan or air-conditioning hum)
- Sustained but Changing (e.g. a passing motor vehicle)
- Intermittent (creaks of furniture, clothing noise)
- Intermittent but Complex (e.g. other people talking)
At present, the Cleaning Kit handles only mono files. To clean stereo or multichannel files you will need to split
the source into its constituent channels and clean each channel separately.
The Cleaning Kit Framework
The Main Cleaning Kit window presents
There are currently 13 options for cleaning sources.
- A list of Cleaning Tools to apply to your source.
- A soundfiles window for listing soundfiles, where your file to be cleaned will appear.
Typically, cleaning a source will involved several passes.
As you proceed, the soundfiles generated at each cleaning stage will be listed in the window.
You can select (by clicking on a file in the window) which of these sounds you wish to clean (e.g. you can go back to the output of
the penultiamte cleaning procedure, if the last cleaning operation proved unsatisfactory).
The window defaults to chosing the output of the last cleaning operation.
- An output filename box, where you can choose a name for the final, completely cleaned sound.
Often you will wish to overwrite the original source, so there is a button to put the name of the source file
in the output-filename box.
You can call any one of these options br pressing the button adjacent to its name.
When a tool is called, a new window will appear and you will be asked to enter relevant parameters for that Cleaning process.
TheCleaning options are
- Clean General Noise
This would typically be used to remove constant background sounds like air-conditioning or extractor-fan hum.
- Gate out Bad Signal
This replaces unwanted sound by silence of the same duration. This can be used, for example
- To remove intakes of breath, or unwanted glossalalia.
- To remove any residual 'flutterings' from the Clean General Noise process.
- To remove low level, but complex, background sounds, like speech, where the principal source pauses (even momentarily).
Particularly useful in situations where the foreground sound you want is loud enough to mask the background sounds most
of the time, except when it pauses.
- Cut away Bad Signal
This can be used, for example
- To remove loud, short, intrusive sounds (from door-slams to clicks and clinks).
- To remove hesitations.
- Reinsert Original Signal
- Remove Pitch in Signal
This is a spectral subtraction process, and can be used, for example
- To remove clearly pitched electronic tones from gadgets used in the environment.
- To remove residual pitches left behind by the Subtract Spectrum or Clean General Noise options.
- To remove an interrupting voice, where both the principal speaker, and the interrupting speaker, are speaking vowels,
and each has a clear but different pitch. This sometimes works perfectly and at other times fails completely!
- Remove Sound under Sibilant
This is a filtering process, and can be used to remove intrusive sounds where the foreground speaker is using a sibilant
Typical intrusions might be
- the hum of passing traffic.
- a 2nd voice using a vowel.
- Remove Noise Above Signal
This is a filtering process, and can be used to remove noise type intrusions
Typical intrusions might be
This often causes some degradation of the principal source at the point where the filtering is done.
- clothing noise.
- a 2nd voice using a sibilant.
It is usually better to use spectral subtraction if this is possible at the particular time in the source.
- Low Frequency Problem
This is a filtering process, and can be used to remove low frequncy intrusions
Typical intrusions might be
- Footfalls or falling objects on carpeted floors.
- physical knocks on microphones or stands.
- wind/breath noise on a microphone.
- Subtract Spectrum
This is typically useful where a noise intrusion begins, ends, or overlays a time where the foreground source is
silent, but then continues over the foreground source.
You are asked to specify the place where the intrusion is not overlaying the principal source as well as the whole area
to be cleaned. The process works by subtracting the former from the latter. It can sometimes be used instead of
filtering or gating.
- Graft segment Elsewhere
This is a last resort process, where nothing can be done to remove a noise intrusion. It allows you, for example, to cut
a sliver of a sybillant from one part of the recording, and use that to replace a noise-intruded sibilant elsewhere.
- Insert Silence
Cutting away bad signal can sometimes result in unnatural speech (or other) rhythms. These can sometimes be mitigated by
inserting an appropriate durations of silence.
- Attenuate Level
This can be used, for example, to compensate for the source's back and forth movement in front of the microphone.
- Dovetail the Ends
This can be useful where a section of recording has been cut from a longer source.
For example, a single voice is extracted from a conversation and the 2 speakers overlap at the boundary but,
to preserve the sense of the speech, the overlapped segment has to be retained.
Using other processes (pitch or sibilant removal, see above) may still leave some residual unwanted signal.
This can sometimes be hidden by subtlely dovetailing the start (or end) of the recording.