Sound Cleanup on Difficult Source Recordings

What can genuinely be repaired in a bad recording, what cannot, and the capture decisions that prevent the problem entirely.

Difficult audio arrives in every production: an interview in a room with air conditioning, a factory floor where the machine cannot be stopped, an event recorded from the back, a client supplied file recorded on a phone. Some of these can be rescued and some cannot, and knowing which before promising a result is the difference between a manageable finishing task and an unpleasant conversation.

The repairs that work well are the ones where the problem is separable from the voice. Steady broadband noise, air conditioning, traffic hum, fan noise, has a consistent profile and can be substantially reduced without damaging speech. Electrical hum at a fixed frequency can be notched out precisely. Intermittent clicks and pops can be removed individually. These are routine and the results are frequently better than clients expect.

The repairs that work partially are those where the problem overlaps the voice. Reverberation in a hard room can be reduced but not eliminated, because the reflections contain the speech itself. Overlapping speech from another person can sometimes be attenuated. Wind noise can be reduced where it sits below the voice and not where it saturates. In these cases the honest description is improvement rather than repair.

The repairs that do not work are worth naming plainly so nobody promises them. Distortion from a signal recorded too loud has lost the information permanently. Speech recorded so quietly that it sits in the noise floor cannot be raised without raising the noise equally. Words obscured entirely by a louder sound are gone. In each case the remaining options are to re-record, to replace the line with narration, or to cut around it.

Aggressive processing has a cost that is easy to underestimate. Heavy noise reduction produces artefacts, a watery, metallic quality on the voice, that many listeners find more objectionable than the original noise. The practical rule is to reduce noise until it stops being distracting rather than until it is gone, and to accept a floor of residual noise as the more natural outcome.

The reason this matters commercially rather than only technically is that voice quality affects belief. Peng et al. (2025) found that vocal cues shape dynamic credibility judgements, and a voice that sounds processed or strained is being evaluated differently from one that sounds present. For interview and testimonial content, where credibility is the entire product, an artefact ridden repair can undermine the material it was meant to save.

The audiovisual construction research supports treating audio as a performance variable rather than a technical afterthought. Zhang et al. (2025) found that audiovisual features of short video advertising contribute measurably to consumer engagement behaviours, and Xiao et al. (2026) reached compatible conclusions from a combined visual and audio perspective. Poor audio is not a small blemish on a good film; it is one of the properties that determines whether people continue watching.

Sound design can cover what repair cannot. A scene with unusable location audio can be rebuilt: room tone, footsteps, movement, ambience and effects placed deliberately, with the dialogue replaced or narrated. This is more work than cleanup and it produces a better result than a heavily processed original, and for generated footage it is the normal approach anyway since there is no location audio at all.

The prevention is cheap and consistently skipped. A lavalier microphone on the subject rather than relying on the camera. A backup recorder running. Room tone captured for thirty seconds at every location. The air conditioning switched off for the take, with a note to switch it back on. Levels checked rather than assumed. These cost minutes on the day and remove almost every problem described above.

For client supplied material the useful practice is to listen before quoting. A studio that agrees to fix the audio on footage it has not heard is accepting an unbounded obligation. Ten minutes with the files establishes whether the job is routine cleanup, substantial rebuilding or genuinely impossible, and setting that expectation before the project starts is the whole of the professional obligation here.

References

Peng, Z., Wang, C., & Jiang, X. (2025). On how vocal cues impact dynamic credibility judgments: Mouse-tracking paradigm examining speaker confidence and gender through voice morphing. Journal of Speech, Language, and Hearing Research, 68(11), 5261–5277. https://doi.org/10.1044/2025_JSLHR-24-00849

Zhang, Z., Qiu, K., & Ye, Y. (2025). Influence of audiovisual features of short video advertising on consumer engagement behaviors: Evidence from TikTok. Journal of Business Research, 201, Article 115662. https://doi.org/10.1016/j.jbusres.2025.115662

Xiao, L., Li, X., & Mou, J. (2026). Exploring user engagement behavior with short-form video advertising on short-form video platforms: A visual-audio perspective. Internet Research, 36(1), 154–188. https://doi.org/10.1108/INTR-07-2023-0521