Play any song and ask what you would miss first if one sound vanished. For most people it is the voice. Listeners follow the words and the feeling, and they notice a flat note or a harsh "s" in half a second. Drums can sit a little wrong and nobody complains. A voice that slips behind the guitars gets reviewed as a bad song.
This is a piece for the desk. The earlier ones explained the tools; this one says in what order to reach for them, what the typical settings are, and how to know when to stop. Everything you can tweak below is built from numbers (a voice made of harmonics shaped by three resonances, plus filtered noise for the "s"), so you can hear what a setting does and also read it.
- Who takes it
- Part 1 of the Mixing & Mastering specialization, inside the Music Production specialization of Music; also for Sound Engineering students. Bands that mix their own records need it more than anyone.
- What it is for
- Taking a recorded vocal from "a good take" to a voice that is clear, even, in tune enough for the style, and sits where you want it against the band.
- Comes after
- Session Prep, Gain Staging & Mix Workflow (36/84) and Mixing & Signal Processing (36/63), which explain EQ, compression, delay and reverb. This piece does not re-teach them.
- Leads to
- Finishing the Mix: Space, Automation & Translation (36/88) and Mastering in Practice (36/89). Its neighbours in Part 1 are Mixing Drums (36/85) and Mixing Bass & Guitars (36/86).
- What this piece gives you
- A step order, ranges to start from, a checklist to tick at the desk and a free session to practise on. It will not give you taste: that comes from mixing a lot of voices. Easier first reads: Music Technology & Production (35/73) and How a Song Gets Made (09/07).
Fix the take first: comping, clean-up and clip gain
Every effect on a vocal makes the problems in it louder. A compressor lifts the breaths and the clicks along with the words; an EQ boost for "air" lifts the esses. So the first job is not a plug-in but an edit. Comp the best lines from the takes into one performance, joining at breaths or noisy consonants where a join is hidden. Mike Senior (Sound On Sound, 2022 and 2025) recommends an equal-gain crossfade where two pieces of the same signal meet, and short fades for mouth clicks, which last roughly one or two milliseconds.
Then use clip gain (a volume change on a piece of the audio, before any plug-in) for three things: turn breaths and harsh consonants down a few decibels, bring up weak words, and even out whole phrases. Senior's point is that clip gain is better done before compression, because a compressor reacts to the worst syllable. A plosive ("p", "b") can be handled with a high-pass filter of around 100 Hz applied only to that moment; across the whole vocal the same filter can thin the voice.
Afterwards, ride. A volume ride is an automation pass on the fader, word by word, that does what no compressor can: it knows which word matters. Do clip gain before the compressor and rides after it.
Pitch correction: how much is too much
Two kinds of tool exist. Graphical editors (Celemony's Melodyne, shown publicly at the Winter NAMM show in 2001) turn the recording into note blobs you move, one at a time, offline. Real-time correctors (Antares Auto-Tune, released 19 September 1997 and designed by Andy Hildebrand) watch the pitch as it plays and pull it toward the nearest note of a scale you set. The key control is retune speed: how quickly the pull happens. Slow keeps the singer's natural slides and vibrato; instant snaps to the note and removes the vibrato. Cher's 1998 single "Believe" is the usual first example of that snap used on purpose.
Whether it is a repair or an effect is a matter of genre and intent. In much pop, folk and rock, the goal is that nobody hears it. In trap, reggaetón and some Afrobeats the hard-tuned, slightly robotic voice is part of the sound, and singers sing toward it. Senior (2025) describes using simple pitch offsets on single syllables rather than automatic correction.
Sound isn't available in this reader; the pitch trace and numbers still work.
Four invented notes, each sung 22 to 40 cents off, scooping up into the pitch and with a 5.5 Hz vibrato of about 25 cents either way. The corrector snaps the sung pitch to the nearest note of the chosen scale, then moves toward that target with a one-pole lag whose time constant is the retune speed. A cent is a hundredth of a semitone. The sound is a synthetic vowel; nothing here is a recording.
Without JavaScript: in the A minor scale with a 60 ms retune speed, the average distance from the intended note falls from 31 to 9 cents and the vibrato on the second note keeps most of its width (50 cents peak to peak before, 41 after). At a retune speed of 0 the distance is 0 and the vibrato is gone.
Two habits keep it honest. Tune after comping and before heavy compression, and judge the result against the finished backing, not solo: a note that is 20 cents off sounds fine alone and sour against a guitar. And leave the scoops and the vibrato alone unless the style asks for the snap.
Serial compression: a catcher, then a leveller
One compressor asked to do everything is a compressor under strain. To close a 16 dB gap between the loudest and the softest word you need a low threshold and a high ratio, so it works hard, the gain slams up and down, and the voice pumps. The alternative, long used for vocals, is serial compression: a fast compressor or limiter first to catch the peaks, then a slower, gentler one to even the phrases. Craig Anderton (Sound On Sound, 2018) describes a transparent limiter taking no more than about 3 to 6 dB, followed by a compressor taking about 6 dB, with an attack of 20 to 30 ms to keep the front of each word, a starting ratio of 2:1, and a meter that moves between 0 and 6 dB.
Try the presets. They are gain-matched to the same overall loudness, so a louder-sounding option is not winning by being louder (Mixing & Signal Processing, 36/63).
Sound isn't available in this reader; the traces and numbers still work.
Five invented sung vowels at 16 dB between the loudest and the softest. Each compressor has a 6 dB soft knee and the same level-detector and smoothing design as in 36/63, run on every sample. "Deepest gain reduction" is the largest value on any waveform peak, so it reads higher than a slow meter would. "Gain movement" adds up how much the combined gain changes every 1 ms. Make-up gain is set so the whole phrase has the same RMS as before.
Without JavaScript: the catcher alone (threshold −34 dB, ratio 4:1) brings the 16.0 dB gap down to 5.2 dB but needs up to 16.2 dB of gain reduction and 124 dB per second of gain movement. The serial preset (catcher −18 dB at 10:1, leveller −30 dB at 3:1) reaches 5.9 dB with 5.5 and 6.2 dB of reduction and 36 dB per second.
The two ways agree on how even the phrase is; they differ in how hard each device works and how much the gain jumps about. That is the case for the pair. It is also not a law: plenty of excellent vocals are one compressor, and the right amount depends on the performance. The ranges below are the places to start.
EQ, de-essing and saturation
Treat EQ as two jobs. The corrective job comes first: a high-pass filter to remove rumble and proximity-effect boom (Anderton suggests starting at 12 dB per octave), and a slight, broad cut around 300 to 400 Hz if the voice sounds boxy. The colour job comes after the compressors: a gentle boost in the 2.5 to 4.5 kHz region for presence, and a high shelf around 10 kHz for air (Senior, 2022). If the vocal is already hissy, Anderton boosts widely in the 4 to 7 kHz region instead of reaching above 8 kHz. For a female voice, shift the regions up by about half an octave (Anderton). These are starting regions: Mixing & Signal Processing (36/63) shows why a cut often beats a boost.
A de-esser is a compressor whose detector listens to a band of high frequencies, typically somewhere in 4 to 10 kHz where the "s" and "sh" of speech live, and turns the voice down only when that band spikes. Two designs exist. A split-band de-esser lowers only the band; a wide-band one lowers everything. Order is argued, and both are common: after the compressors, since compression lifts sibilance, or at the end of the chain. Senior's advice is to reduce the worst consonants by clip gain first so the de-esser has less to do.
Sound isn't available in this reader; the traces and numbers still work.
Three invented vowels, each followed by a burst of filtered noise standing in for an "s". The detector is a band-pass filter (Q 0.8) at the centre you choose, followed by a 2 ms attack and 40 ms release. Split-band subtracts the lowered band from the signal; wide-band turns the whole signal down. Figures compare the processed signal with the original in each region, measured as RMS.
Without JavaScript: at the defaults (7 kHz, −34 dB, 6:1) split-band lowers the "s" sounds by 4.6 dB and leaves the vowels alone; wide-band lowers them by 10.5 dB but also pushes down the end of each vowel by 4.1 dB. Moving the detector to 1 kHz makes it hear the vowels: they fall 4.1 dB and the "s" is untouched.
Saturation is the last small step: a little harmonic distortion (tape, tube or transformer models, or a clipper) adds upper harmonics that help a voice stay audible on a small speaker, the same idea as in Mixing Bass & Guitars (36/86). Add it until you hear it, then take some back.
Space: delay, reverb and pre-delay
A dry vocal sounds like it is inside your head. Delay and reverb put it in a place. The time of a delay should belong to the song: a quarter note lasts 60,000 ÷ BPM milliseconds, and an eighth note is half of that. Anderton suggests an eighth-note delay to thicken a voice. Pre-delay is the gap before the reverb starts; a copy later than about 40 ms starts to be heard as a separate event (36/63), so a short pre-delay keeps the reverb attached to the voice, while a longer one opens a gap that keeps the words clear. A reverb tail that has not faded by the next beat smears the words.
Sound isn't available in this reader; the table and numbers still work.
The sound is two short synthetic vowels with a soft tick on every beat. The "late" version stretches the delay by 20%, so the repeats drift off the ticks. Reverb decay is shown as a straight fall in decibels (what RT60 means); no reverb is rendered.
Without JavaScript: at 100 BPM a quarter note is 600 ms, a dotted eighth 450 ms, an eighth 300 ms, an eighth triplet 200 ms and a sixteenth 150 ms. A reverb with a 1.8 s decay has fallen by only 10 dB by the next eighth note (300 ms).
Two habits matter more than any setting. Put the effects on sends to their own tracks, so you can filter and compress them separately. And automate them: Senior (2025) rides the effect sends as well as the fader, so a delay can throw the last word of a line into the gap and stay out of the way of the next.
Doubles, harmonies and ad-libs
A double is the same line sung twice. Mixed low and panned apart, it widens the chorus; mixed under the lead, it thickens it. Harmonies usually go lower than the lead in level, with the highs rolled off a little so they support rather than compete. Ad-libs belong in the gaps between phrases, where they can be louder than they could be on top of the words. Check all three in mono (Mixing & Signal Processing, 36/63; and 36/88 for the full check) because panned doubles can thin out when the two sides add.
Starting points, not rules
These are ranges to start from, with where each comes from. Your ears overrule every one of them: a voice, a mic and a song are never the average of anything. Treat a number as the place to put the knob before you listen, not a target.
| Move | Starting range | Where it comes from |
|---|---|---|
| High-pass on the vocal | about 80 to 120 Hz, 12 dB per octave to start | Anderton (SOS, 2018); Senior (SOS, 2025) uses about 100 Hz on plosives only |
| Mud cut | slight, broad, around 300 to 400 Hz | Anderton (SOS, 2018) |
| Presence | gentle boost, 2.5 to 4.5 kHz | Anderton (SOS, 2018) |
| Air | high shelf around 10 kHz; 4 to 7 kHz wide if hissy | Senior (SOS, 2022); Anderton (SOS, 2018) |
| Peak catcher | up to about 3 to 6 dB of reduction | Anderton (SOS, 2018) |
| Leveller | 2:1 to start, attack 20 to 30 ms, about 6 dB or less | Anderton (SOS, 2018) |
| De-esser band | roughly 4 to 10 kHz; start where the meter moves only on esses | Typical practice; Anderton's 4 to 7 kHz for hissy voices |
| Delay time | 60,000 ÷ BPM × ½ (eighth) to start | Arithmetic; Anderton (SOS, 2018) |
| Pre-delay | under about 40 ms attaches, over detaches | The fusion limit in 36/63 |
Common mistakes
Compressing before cleaning up. The compressor lifts breaths, clicks and plosives. Fix: edit and clip-gain first.
One compressor doing 15 dB. The voice pumps and loses its life. Fix: two stages of 3 to 6 dB, or clip gain and rides for the big differences.
Boosting air on a sibilant voice. The esses go from annoying to painful. Fix: de-ess first or boost a lower band.
Tuning in solo. The note sounds right alone and wrong in the song. Fix: tune against the backing.
Judging by loudness. A louder vocal always sounds better for a moment. Fix: gain-match every A/B.
Reverb that outlasts the beat. The words blur. Fix: shorten the decay, high-pass the return, or duck it under the voice.
Burying the voice to make the band fit. Fix: carve the band's mid-range under the voice (36/63) and set the level for the genre, below.
Genre and world differences
How loud the voice sits, and how tuned, is a style choice more than a rule. These are common practices in each style, not measured averages.
- Pop and hip-hop: the vocal is the loudest thing, close, dry and heavily compressed. In trap the lead often has doubles and ad-libs left and right, and audible tuning is part of the sound.
- Reggaetón and dembow: a forward, bright, closely doubled voice over a sparse beat; hard-tuned effects are common (see Beatmaking, 36/69). Latin pop mixes tend to keep the lead wide open and dry in the middle.
- Rock and metal: the voice sits level with or slightly inside the guitars, with more room sound and less top-end polish. Shouted and sung parts need different de-essing and different compression.
- Cumbia, Afrobeats and highlife: often a lead plus a group of backing singers, so the group is a part of the arrangement and needs its own level and width; response vocals interact with a busy percussion section.
- Jazz, folk and acoustic: the voice is part of a small ensemble. Light compression, natural room and breath left in; a heavily processed voice sounds out of place.
The vocal checklist
Vocal mix checklist (36/87)
Ticks are not saved when you leave the page; copy the list if you want to keep your progress. To print, use your browser's print command: only the checklist is printed.
Practise on a real session
Cambridge-MT's Mixing Secrets free multitrack library, run by Mike Senior, is a large collection of recorded songs, split into their individual tracks, offered for mixing practice and teaching. It is free for educational use; check the terms on its FAQ page before you post a mix publicly, and contact the artist or rights holder directly for any commercial use. Three sessions listed in its Pop and Singer-Songwriter area are good for vocal practice: Georgia Wonder's "Siren" (a single comped lead with a pre-mixed submix of doubles), Lyndsey Ollard's "Catching Up" (a pop song with a double in the chorus), and Amber Skye's "Howlin'". Start with the lead alone against a rough balance, and work down the checklist.
What a graduate just sees
- Every effect makes problems louder, so edit and clip-gain before processing.
- Compress in two gentle stages instead of one hard one; gain-match every comparison.
- Retune speed decides between a repair and an effect; slow keeps the vibrato, instant removes it.
- A de-esser is a compressor with an ear: aim the detector at the "s", not the vowel.
- A delay belongs to the tempo: 60,000 ÷ BPM, times a note fraction.
- Pre-delay under about 40 ms stays attached; over it, it detaches.
- The level of the voice is a choice made for the genre, and every number here is a start.
Where it shows up next
- Finishing the Mix (36/88): depth, width and automation for the vocal, and the translation check on phones and in the car.
- Mastering in Practice (36/89): a vocal that is already even and de-essed survives a limiter better.
- Producing Records (36/70): how to get a vocal recorded so that it needs less of this.
- Sound for Film, TV & Games (36/67): dialogue gets the same chain, with a different idea of "natural".
- In the world: podcasts, voice notes, radio presenters and a phone's "voice isolation" all use these tools.
If you want to go deeper
- 📘 Textbook. Mike Senior, Mixing Secrets for the Small Studio (Routledge, 2nd edition, 2018), written for people mixing at home, and Roey Izhaki, Mixing Audio (Routledge).
- 📘 Short and practical. Bobby Owsinski, The Mixing Engineer's Handbook, and the Sound On Sound articles listed below.
- 🌐 Free reading. Sound On Sound: Mike Senior, Mixing Pop Vocals (2025) and Vocal Mix Prep (2022); Craig Anderton, How To Make Your Vocals Twice As Good, Part 2 (2018).
- 🌐 Free course. MIT OpenCourseWare 21M.380 Music and Technology: Recording Techniques and Audio Production (Fall 2016).
A vocal mix is a short sequence done in order: clean the take, tune what needs it, level the words, compress in two soft steps, de-ess, add colour, place it in time and space, and then compare it at matched loudness. The ranges are a place to start; the voice and the song decide where to stop.
Try this: take any song you like, loop one line, and turn the vocal up and down by 1 dB with your eyes shut until you can hear the exact point where it feels "in" the band. Then check that point at low volume. Does it still feel the same?
Sources. Mike Senior, "Mixing Pop Vocals" (Sound On Sound, November 2025) and "Vocal Mix Prep" (Sound On Sound, February 2022); Craig Anderton, "How To Make Your Vocals Twice As Good! Part 2: Fitting Vocals In The Mix" (Sound On Sound, August 2018); Wikipedia, "Auto-Tune" (released 19 September 1997 by Antares Audio Technologies, developed by Andy Hildebrand; Cher's "Believe", 1998) and "Melodyne" (first public viewing at Winter NAMM 2001; Celemony, Peter Neubäcker); Cambridge-MT, Mixing Secrets free multitrack library and FAQ; the fusion limit of about 40 ms is from 36/63; the compressor design follows Giannoulis, Massberg and Reiss, JAES 60(6), 2012. Every number in the interactives was computed in node.
- 📘 Senior, Mixing Secrets for the Small Studio
- 🌐 Sound On Sound: Mixing Pop Vocals
- 🌐 MIT OCW 21M.380