{"course":{"slug":"motion-and-audio","title":"Motion, Sound and Music","url":"https://addaly.com/learn/motion-and-audio","level":"intermediate","track":"create"},"lesson":{"slug":"noise-floor","title":"What the noise floor costs you later","url":"https://addaly.com/learn/motion-and-audio/noise-floor","position":51,"of":72,"minutes":11,"answer":"Room noise is diffuse and roughly constant while the voice obeys the inverse square law, so every halving of distance buys 6dB of signal-to-noise for nothing.","keyIdea":"Room noise is diffuse and roughly constant while the voice obeys the inverse square law, so every halving of distance buys 6dB of signal-to-noise for nothing; and spectral subtraction leaves a residue of isolated tonal bursts because the noise it subtracts is an average of something random."},"body":{"format":"markdown","text":"## The cost of noise is paid later, not now\n\nA recording with a quiet background hiss sounds acceptable on its own. The problem appears three steps downstream: you compress it and the hiss comes up with everything else, you cut between two takes and the background jumps, you add music and dialogue and the mix turns grey, and every attempt to fix it makes something else worse.\n\nNoise is the constraint that limits what you can do to a recording afterwards. It is worth measuring rather than sensing.\n\n## Where it comes from, in the usual order of size\n\n1. **The room.** Ventilation, a computer fan, a fridge, traffic, a light fitting's transformer. Almost always the dominant source in a domestic recording, and almost always underestimated because you stop hearing it within a minute of entering.\n2. **The preamp.** Its own thermal and circuit noise, amplified with the signal. The figure to look for is equivalent input noise, typically −125 to −129 dBu for a good preamp; a cheap one at high gain will be audibly worse.\n3. **The microphone.** A condenser's self-noise, published as an A-weighted figure. Under 15 dB-A is quiet; over 20 dB-A is audible on a soft voice.\n4. **The converter.** Usually the smallest contributor in any modern interface.\n\nThe order matters because it tells you where to spend effort. Turning off the ventilation is free and typically wins 10dB. A better microphone might win 3.\n\n## Measure it in thirty seconds\n\nRecord ten seconds of the room with nobody speaking, then:\n\n```bash\nffmpeg -i roomtone.wav -af \"astats=measure_overall=RMS_level\" -f null -\n```\n\nThen measure the speech. The difference is your signal-to-noise ratio.\n\n- **60dB or more**: very good, you can do anything to it.\n- **45 to 60dB**: normal for a decent home setup; usable with care.\n- **30 to 45dB**: the noise will be audible after compression and will rise in every quiet passage.\n- **Under 30dB**: you are going to hear it, and denoising will cost you quality.\n\n## The inverse square law is a noise tool\n\nRoom noise is largely diffuse: it arrives from everywhere and its level does not change much as you move. Your voice obeys the inverse square law.\n\n> Figure: Where the noise goes, and what each route costs. Bought before the record button: Ventilation off: typically about 10 dB; Sixty centimetres to fifteen: 12 dB, free; A window shut and a fridge unplugged; A quieter microphone: perhaps 3 dB. Bought afterwards, and paid for: Spectral subtraction leaves musical noise; Consonants are noise-like, so they are dulled; Breaths and reverb tails get chopped; Six to ten decibels is usually invisible; twenty is not. Room noise is diffuse and roughly constant wherever you stand; your voice obeys the inverse square law. So every halving of distance is 6 dB of signal-to-noise for nothing, and it is the only improvement on this page that costs neither money nor quality.\n\nSo halving the distance from your mouth to the microphone raises the voice by 6dB and leaves the noise where it was. **Every halving of distance buys 6dB of signal-to-noise ratio, free.** Going from 60cm to 15cm is 12dB: more than the difference between a budget microphone and a good one.\n\nThis is the same mechanism as the direct-to-reverberant argument, and it points the same way: distance is the most powerful control you have and it costs nothing.\n\n## Room tone, and why ten seconds saves an hour\n\nRecord the room in silence for at least ten seconds at the same gain, in the same position, before or after every setup.\n\nIt has two uses. First, **edits**. Cut a sentence out of an interview and the background stops for a frame; drop room tone underneath and the join disappears. Every dialogue editor does this and it is invisible when done and glaring when not.\n\nSecond, **a noise profile**. Every denoiser works better when given a clean sample of exactly the noise it is being asked to remove.\n\nTen seconds. Do it before anyone speaks, while everyone is still settling.\n\n## Why denoising leaves a strange texture\n\nClassical noise reduction works by **spectral subtraction**. It measures the noise's average level in each of several hundred frequency bands, then reduces each band of the signal by that amount.\n\nThe noise is random, so in any given moment a band's actual noise level is above or below its average. Subtracting the average leaves a residual that is sometimes zero and sometimes a short isolated burst of energy in one band. Played back, those residuals are heard as brief tonal chirps appearing and vanishing: **musical noise**, and once you can name it you will hear it in every over-processed podcast.\n\nPush the reduction harder and two further things happen. Consonants (s, t, k, f) are broadband and noise-like, so the algorithm treats them as noise and dulls them. And breaths and room reverb tails get chopped, which makes the voice sound as if it is being switched on and off.\n\nMachine-learning denoisers, including the free **RNNoise** and the models built into several editors, are considerably better because they have learned what speech looks like rather than subtracting an average. They fail differently: they can remove non-speech content you wanted, smooth the voice into a slightly synthetic texture, and behave unpredictably on accents or languages under-represented in training.\n\nThe practical discipline is the same for both: **use the least reduction that solves the problem.** 6 to 10dB is usually enough and often inaudible. 20dB is nearly always audible and usually worse than the noise was.\n\n## Hum is a different problem\n\nA steady tone at 50 or 60Hz with harmonics is mains hum, caused by a ground loop or an unbalanced cable near a power supply. It is not broadband noise and a denoiser is the wrong tool.\n\nFix it at the source: a balanced cable, a different socket, moving a power brick away from the microphone cable, or a ground-lift on a DI box. Failing that, a narrow notch filter at the fundamental and its harmonics removes it with far less collateral damage than broadband reduction.\n\n## The free path\n\n**Audacity** has a noise-profile-based reduction: the classic spectral subtraction, with the residual characteristics described above. **RNNoise** is open source and available as an LADSPA plugin and inside several editors. **ffmpeg** has `afftdn` for spectral denoising and `anlmdn` for a non-local-means approach that is gentler on speech. **DaVinci Resolve**'s free edition includes its own voice isolation. **iZotope RX** is the paid industry standard and is named in job adverts; nothing here requires it.\n\n## The limitation\n\nDenoising cannot add back what the noise masked. If a word sits at the same level as the hiss, removing the hiss removes the word with it. The information about what was said is not separable from the noise at that point: they are the same samples.\n\nAnd the most advanced tools are the hardest to evaluate, because a machine-learning denoiser produces confident, clean output whether or not the result resembles what was said. On a quiet or accented passage it can smooth a word into something else entirely, and the result sounds plausible. Anything that will be quoted, transcribed or relied on should be checked against the original rather than trusted because it came out clean."},"check":{"question":"After heavy noise reduction, an interview sounds clean but is dotted with brief tonal chirps and the consonants have gone dull. What produced each symptom?","options":[{"text":"The noise profile was taken from a section that contained speech, so the algorithm is subtracting speech formants as well as noise.","correct":false,"tempting":"A contaminated profile is a real error, but it removes vowel content broadly rather than producing isolated chirps."},{"text":"Subtracting an averaged noise spectrum from a random signal leaves isolated residual peaks, and broadband consonants resemble noise closely enough to be attenuated with it.","correct":true},{"text":"The reduction was applied twice, and cascading two passes of spectral subtraction produces intermodulation between the two residuals.","correct":false,"tempting":"Repeated passes worsen both symptoms, but neither symptom requires a second pass to appear."},{"text":"The denoiser is operating on too short an analysis window, so its frequency resolution is insufficient to separate narrow noise components from speech harmonics, and lengthening the window would resolve both the chirps and the dulled consonants without any loss of reduction depth.","correct":false,"tempting":"Window length is a genuine tuning parameter, but no window setting removes the fundamental problem of subtracting an average from a random signal."}]},"prev":{"slug":"sample-rate-and-bit-depth","title":"Two numbers doing two different jobs","url":"https://addaly.com/learn/motion-and-audio/sample-rate-and-bit-depth"},"next":{"slug":"phase-and-two-microphones","title":"Two microphones and a comb filter","url":"https://addaly.com/learn/motion-and-audio/phase-and-two-microphones"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'What the noise floor costs you later', in Motion, Sound and Music. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/motion-and-audio/noise-floor (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What the noise floor costs you later","courseTitle":"Motion, Sound and Music","module":{"index":6,"slug":"capturing-sound","title":"Capturing sound"},"answer":null,"keyIdea":"Room noise is diffuse and roughly constant while the voice obeys the inverse square law, so every halving of distance buys 6dB of signal-to-noise for nothing; and spectral subtraction leaves a residue of isolated tonal bursts because the noise it subtracts is an average of something random.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"left":{"title":"Bought before the record button","items":["Ventilation off: typically about 10 dB","Sixty centimetres to fifteen: 12 dB, free","A window shut and a fridge unplugged","A quieter microphone: perhaps 3 dB"]},"type":"compare","right":{"title":"Bought afterwards, and paid for","items":["Spectral subtraction leaves musical noise","Consonants are noise-like, so they are dulled","Breaths and reverb tails get chopped","Six to ten decibels is usually invisible; twenty is not"]},"title":"Where the noise goes, and what each route costs","caption":"Room noise is diffuse and roughly constant wherever you stand; your voice obeys the inverse square law. So every halving of distance is 6 dB of signal-to-noise for nothing, and it is the only improvement on this page that costs neither money nor quality."},"alt":"Figure: Where the noise goes, and what each route costs. Bought before the record button: Ventilation off: typically about 10 dB; Sixty centimetres to fifteen: 12 dB, free; A window shut and a fridge unplugged; A quieter microphone: perhaps 3 dB. Bought afterwards, and paid for: Spectral subtraction leaves musical noise; Consonants are noise-like, so they are dulled; Breaths and reverb tails get chopped; Six to ten decibels is usually invisible; twenty is not. Room noise is diffuse and roughly constant wherever you stand; your voice obeys the inverse square law. So every halving of distance is 6 dB of signal-to-noise for nothing, and it is the only improvement on this page that costs neither money nor quality."}},"blocks":[{"t":"h2","id":"the-cost-of-noise-is-paid-later-not-now","runs":[{"text":"The cost of noise is paid later, not now"}]},{"t":"p","runs":[{"text":"A recording with a quiet background hiss sounds acceptable on its own. The problem appears three steps downstream: you compress it and the hiss comes up with everything else, you cut between two takes and the background jumps, you add music and dialogue and the mix turns grey, and every attempt to fix it makes something else worse."}]},{"t":"p","runs":[{"text":"Noise is the constraint that limits what you can do to a recording afterwards. It is worth measuring rather than sensing."}]},{"t":"h2","id":"where-it-comes-from-in-the-usual-order-of-size","runs":[{"text":"Where it comes from, in the usual order of size"}]},{"t":"ol","start":1,"items":[[{"text":"The room.","b":true},{"text":" Ventilation, a computer fan, a fridge, traffic, a light fitting’s transformer. Almost always the dominant source in a domestic recording, and almost always underestimated because you stop hearing it within a minute of entering."}],[{"text":"The preamp.","b":true},{"text":" Its own thermal and circuit noise, amplified with the signal. The figure to look for is equivalent input noise, typically −125 to −129 dBu for a good preamp; a cheap one at high gain will be audibly worse."}],[{"text":"The microphone.","b":true},{"text":" A condenser’s self-noise, published as an A-weighted figure. Under 15 dB-A is quiet; over 20 dB-A is audible on a soft voice."}],[{"text":"The converter.","b":true},{"text":" Usually the smallest contributor in any modern interface."}]]},{"t":"p","runs":[{"text":"The order matters because it tells you where to spend effort. Turning off the ventilation is free and typically wins 10dB. A better microphone might win 3."}]},{"t":"h2","id":"measure-it-in-thirty-seconds","runs":[{"text":"Measure it in thirty seconds"}]},{"t":"p","runs":[{"text":"Record ten seconds of the room with nobody speaking, then:"}]},{"t":"code","lang":"bash","label":"Bash","text":"ffmpeg -i roomtone.wav -af \"astats=measure_overall=RMS_level\" -f null -"},{"t":"p","runs":[{"text":"Then measure the speech. The difference is your signal-to-noise ratio."}]},{"t":"ul","items":[[{"text":"60dB or more","b":true},{"text":": very good, you can do anything to it."}],[{"text":"45 to 60dB","b":true},{"text":": normal for a decent home setup; usable with care."}],[{"text":"30 to 45dB","b":true},{"text":": the noise will be audible after compression and will rise in every quiet passage."}],[{"text":"Under 30dB","b":true},{"text":": you are going to hear it, and denoising will cost you quality."}]]},{"t":"h2","id":"the-inverse-square-law-is-a-noise-tool","runs":[{"text":"The inverse square law is a noise tool"}]},{"t":"p","runs":[{"text":"Room noise is largely diffuse: it arrives from everywhere and its level does not change much as you move. Your voice obeys the inverse square law."}]},{"t":"p","runs":[{"text":"So halving the distance from your mouth to the microphone raises the voice by 6dB and leaves the noise where it was. "},{"text":"Every halving of distance buys 6dB of signal-to-noise ratio, free.","b":true},{"text":" Going from 60cm to 15cm is 12dB: more than the difference between a budget microphone and a good one."}]},{"t":"p","runs":[{"text":"This is the same mechanism as the direct-to-reverberant argument, and it points the same way: distance is the most powerful control you have and it costs nothing."}]},{"t":"h2","id":"room-tone-and-why-ten-seconds-saves-an-hour","runs":[{"text":"Room tone, and why ten seconds saves an hour"}]},{"t":"p","runs":[{"text":"Record the room in silence for at least ten seconds at the same gain, in the same position, before or after every setup."}]},{"t":"p","runs":[{"text":"It has two uses. First, "},{"text":"edits","b":true},{"text":". Cut a sentence out of an interview and the background stops for a frame; drop room tone underneath and the join disappears. Every dialogue editor does this and it is invisible when done and glaring when not."}]},{"t":"p","runs":[{"text":"Second, "},{"text":"a noise profile","b":true},{"text":". Every denoiser works better when given a clean sample of exactly the noise it is being asked to remove."}]},{"t":"p","runs":[{"text":"Ten seconds. Do it before anyone speaks, while everyone is still settling."}]},{"t":"h2","id":"why-denoising-leaves-a-strange-texture","runs":[{"text":"Why denoising leaves a strange texture"}]},{"t":"p","runs":[{"text":"Classical noise reduction works by "},{"text":"spectral subtraction","b":true},{"text":". It measures the noise’s average level in each of several hundred frequency bands, then reduces each band of the signal by that amount."}]},{"t":"p","runs":[{"text":"The noise is random, so in any given moment a band’s actual noise level is above or below its average. Subtracting the average leaves a residual that is sometimes zero and sometimes a short isolated burst of energy in one band. Played back, those residuals are heard as brief tonal chirps appearing and vanishing: "},{"text":"musical noise","b":true},{"text":", and once you can name it you will hear it in every over-processed podcast."}]},{"t":"p","runs":[{"text":"Push the reduction harder and two further things happen. Consonants (s, t, k, f) are broadband and noise-like, so the algorithm treats them as noise and dulls them. And breaths and room reverb tails get chopped, which makes the voice sound as if it is being switched on and off."}]},{"t":"p","runs":[{"text":"Machine-learning denoisers, including the free "},{"text":"RNNoise","b":true},{"text":" and the models built into several editors, are considerably better because they have learned what speech looks like rather than subtracting an average. They fail differently: they can remove non-speech content you wanted, smooth the voice into a slightly synthetic texture, and behave unpredictably on accents or languages under-represented in training."}]},{"t":"p","runs":[{"text":"The practical discipline is the same for both: "},{"text":"use the least reduction that solves the problem.","b":true},{"text":" 6 to 10dB is usually enough and often inaudible. 20dB is nearly always audible and usually worse than the noise was."}]},{"t":"h2","id":"hum-is-a-different-problem","runs":[{"text":"Hum is a different problem"}]},{"t":"p","runs":[{"text":"A steady tone at 50 or 60Hz with harmonics is mains hum, caused by a ground loop or an unbalanced cable near a power supply. It is not broadband noise and a denoiser is the wrong tool."}]},{"t":"p","runs":[{"text":"Fix it at the source: a balanced cable, a different socket, moving a power brick away from the microphone cable, or a ground-lift on a DI box. Failing that, a narrow notch filter at the fundamental and its harmonics removes it with far less collateral damage than broadband reduction."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"Audacity","b":true},{"text":" has a noise-profile-based reduction: the classic spectral subtraction, with the residual characteristics described above. "},{"text":"RNNoise","b":true},{"text":" is open source and available as an LADSPA plugin and inside several editors. "},{"text":"ffmpeg","b":true},{"text":" has "},{"text":"afftdn","code":true},{"text":" for spectral denoising and "},{"text":"anlmdn","code":true},{"text":" for a non-local-means approach that is gentler on speech. "},{"text":"DaVinci Resolve","b":true},{"text":"’s free edition includes its own voice isolation. "},{"text":"iZotope RX","b":true},{"text":" is the paid industry standard and is named in job adverts; nothing here requires it."}]},{"t":"h2","id":"the-limitation","runs":[{"text":"The limitation"}],"role":"limit"},{"t":"p","runs":[{"text":"Denoising cannot add back what the noise masked. If a word sits at the same level as the hiss, removing the hiss removes the word with it. The information about what was said is not separable from the noise at that point: they are the same samples."}]},{"t":"p","runs":[{"text":"And the most advanced tools are the hardest to evaluate, because a machine-learning denoiser produces confident, clean output whether or not the result resembles what was said. On a quiet or accented passage it can smooth a word into something else entirely, and the result sounds plausible. Anything that will be quoted, transcribed or relied on should be checked against the original rather than trusted because it came out clean."}]}],"check":{"question":[{"text":"After heavy noise reduction, an interview sounds clean but is dotted with brief tonal chirps and the consonants have gone dull. What produced each symptom?"}],"options":[{"text":[{"text":"The noise profile was taken from a section that contained speech, so the algorithm is subtracting speech formants as well as noise."}],"correct":false,"tempting":[{"text":"A contaminated profile is a real error, but it removes vowel content broadly rather than producing isolated chirps."}]},{"text":[{"text":"Subtracting an averaged noise spectrum from a random signal leaves isolated residual peaks, and broadband consonants resemble noise closely enough to be attenuated with it."}],"correct":true,"tempting":null},{"text":[{"text":"The reduction was applied twice, and cascading two passes of spectral subtraction produces intermodulation between the two residuals."}],"correct":false,"tempting":[{"text":"Repeated passes worsen both symptoms, but neither symptom requires a second pass to appear."}]},{"text":[{"text":"The denoiser is operating on too short an analysis window, so its frequency resolution is insufficient to separate narrow noise components from speech harmonics, and lengthening the window would resolve both the chirps and the dulled consonants without any loss of reduction depth."}],"correct":false,"tempting":[{"text":"Window length is a genuine tuning parameter, but no window setting removes the fundamental problem of subtracting an average from a random signal."}]}],"order":[3,2,1,0],"why":null},"prev":{"slug":"sample-rate-and-bit-depth","title":"Two numbers doing two different jobs","minutes":10},"next":{"slug":"phase-and-two-microphones","title":"Two microphones and a comb filter","minutes":10}}}