1. What this lesson is for

In this lesson you turn raw spoken audio into clean, even, speech-first sound that stays clear on a phone speaker, on earbuds and on studio headphones. Keep the module principle in front of you: intelligible speech is non-negotiable. Music, effects and style come later (lesson 4.2) - none of them can rescue a voice the viewer cannot understand.

By the end of the module you will deliver one finalized mix where dialogue or voiceover stays clear across phone and headphone playback. This lesson builds the first half of that mix: cleanup, leveling and EQ.

2. The mixing flow

Always work in the same order. Each step depends on the one before it: denoise changes the level, leveling changes how EQ sounds, and ducking only works if the voice is already consistent.

3. Noise reduction basics

Noise is anything that is not the voice: room hum, fans, traffic, hiss from a cheap mic. CapCut offers a noise reduction toggle (sometimes with a strength setting, sometimes just on/off) in the audio panel of a selected clip. The exact name and options differ between desktop, mobile and web versions, so look for Reduce noise or similar under the clip's audio settings.

• Fix it at the source first. Closer mic, quieter room, soft furnishings. No plugin beats 20 cm less distance to the mouth.
• Denoise before anything else. Leveling and compression raise the noise floor together with quiet speech, so remove noise while it is still low.
• Less is more. Use the lightest setting that makes the noise unobtrusive, not absent. A little room tone sounds natural; zero noise with a metallic voice sounds broken.
• Listen to the pauses and the consonants. Artifacts show up first in gaps (a "watery" swirl) and on S, T, F sounds (a lisp or phasing).

4. Cleanup chain template

Use this as your default chain for a single voice clip. Parameter ranges are starting points; CapCut may expose only some of these controls (or none, as an on/off switch) in your version. If a control is missing, do the step in a dedicated audio tool and re-import.

Step	Tool	Parameter range	Warning signs
1. Trim silence and clicks	Split / trim, short fades	Fades of 3-10 ms at edit points	Clicks or pops at cuts
2. Noise reduction	Reduce noise	Lowest strength that works	Metallic, watery or lisping voice
3. Low cut (high-pass)	EQ	70-100 Hz, gentle slope	Thin voice if set above ~120 Hz
4. Clip gain / normalize	Volume, loudness normalization	Speech peaks around -6 to -3 dBFS	Some clips obviously louder than others
5. Presence / clarity	EQ	+1 to +3 dB around 2-5 kHz	Harsh, tiring or sibilant voice
6. De-ess (if available)	De-esser or narrow EQ cut	-2 to -6 dB around 5-8 kHz	Lisp, dull S sounds
7. Final ceiling	Limiter / export normalization	True peak -1 dBTP	Distortion on loud words, clipping

5. Loudness consistency

Two numbers matter. Peak is the loudest single moment; it tells you about clipping. Loudness (measured in LUFS, Loudness Units relative to Full Scale) is the perceived average; it tells you whether your video sounds as loud as the one before it in the feed. Platforms normalize playback toward a target, so being far louder than the target buys you nothing and costs you dynamics.

Reference targets per platform (approximate; platforms change these without notice):

Platform	Integrated loudness	True peak	Notes
YouTube	about -14 LUFS	-1 dBTP	Louder uploads are turned down; quieter ones are generally not turned up
TikTok / Instagram Reels / Shorts	-14 to -12 LUFS	-1 dBTP	No official public spec; mostly phone playback, so speech clarity beats loudness
Spotify / podcasts	-14 LUFS (Spotify), -16 LUFS (common podcast)	-1 to -2 dBTP	Relevant if you reuse the audio as a podcast
Broadcast (EBU R128)	-23 LUFS	-1 dBTP	Only if you deliver to TV

And the per-track reference for the whole module (the mix you finish in lesson 4.2):

Track type	Target level	Peak ceiling	Notes
Dialogue / voiceover	Loudest element, peaks -6 to -3 dBFS	-1 dBTP	Always on top; everything else is measured relative to it
Music under speech	15-25 dB below voice	-6 dBFS	Should feel present, never compete
Music without speech	6-10 dB below voice level	-3 dBFS	Intros, transitions, breaks
SFX accents	Short peaks near voice level	-3 dBFS	Brief; avoid masking words
Full mix	about -14 LUFS integrated	-1 dBTP	Measure after export

CapCut does not always show a LUFS meter. It may offer a loudness normalization option on clips or at export; if it does not, measure the exported file with a free tool such as ffmpeg:

# measure integrated loudness (I), true peak (TP) and range (LRA)
ffmpeg -i final_mix.mp4 -af loudnorm=print_format=summary -f null -

# typical output to read
Input Integrated:    -17.8 LUFS    -> about 4 dB too quiet for -14
Input True Peak:      -0.2 dBTP    -> too hot, aim for -1.0
Input LRA:             6.1 LU      -> fine for speech (roughly 4-10)

Log every clip while leveling so you can see what changed and why:

LOUDNESS LOG - episode_03
CLIP            PEAK     AVERAGE    GAIN APPLIED   RESULT
intro_vo        -1.5     -20 LUFS   -2 dB          peaks now -3.5, ok
interview_a     -9.0     -26 LUFS   +5 dB          matches intro
interview_b     -4.0     -22 LUFS   +1 dB          ok, slight hiss audible
street_take     -0.1     -16 LUFS   -4 dB + denoise  was clipping, now clean
outro_vo        -6.0     -21 LUFS   +1 dB          ok
FULL MIX        -1.0 dBTP  -14.2 LUFS  -           pass

6. EQ and clarity for speech

EQ shapes the tone of the voice. CapCut exposes EQ as presets or a simple band editor depending on the version; the frequency guide below applies to any EQ.

Range	What lives there	Typical move	Too much sounds like
Below 80 Hz	Rumble, handling noise, traffic	Cut (high-pass)	-
100-250 Hz	Body and warmth	Leave, or small cut if boomy	Cut: thin; boost: muddy
250-500 Hz	Boxiness, "room"	-2 to -4 dB narrow cut	Cut too much: hollow
2-5 kHz	Presence, intelligibility of consonants	+1 to +3 dB wide boost	Harsh, tiring
5-8 kHz	Sibilance (S, SH)	Cut or de-ess if sharp	Cut too much: lisp
Above 10 kHz	Air	Small shelf boost if dull	Hiss becomes obvious

7. Monitoring with headphones and phone

Laptop speakers hide both bass problems and hiss. Checking the mix only on laptop speakers is one of this module's listed pitfalls. Use this routine every time:

1. Headphones (closed-back if possible) at a moderate, fixed volume - hunt for noise, clicks, denoise artifacts and sibilance.
2. Phone speaker - export or preview on the phone. Is every word understandable without subtitles?
3. Earbuds - check that nothing is harsh or tiring at high volume.
4. Log the differences, then fix and re-check.

MONITORING LOG - episode_03
TIME      DEVICE       ISSUE                          FIX
00:12     headphones   hiss in pause                  denoise one step up
00:34     phone        "fifteen" sounds like "fifty"  +2 dB at 3 kHz on interview_a
00:51     earbuds      sharp S on "system"            de-ess / narrow cut at 6.5 kHz
01:05     phone        outro quieter than intro       +1.5 dB clip gain

8. QA checklist

• Noise floor reduced without metallic or watery artifacts.
• Levels normalized - no clip is noticeably louder or quieter than its neighbours.
• No clipping: true peak at or below -1 dBTP on the export.
• Consonants remain crisp; no harsh sibilance.
• Voice keeps its body - EQ cuts did not make it thin.
• Checked on headphones and a phone speaker, with differences logged.

9. Exercise

1. Record or pick a 60-90 second voice clip with some room noise and uneven volume. Duplicate the project so you keep the original.
2. Apply the cleanup chain from section 4 in order. Write down each setting you used.
3. Fill in a loudness log for every clip, then export and measure the full mix (target about -14 LUFS, -1 dBTP).
4. Run the monitoring routine on headphones and a phone speaker and log at least three differences.
5. Retrospective: where did denoise introduce a metallic sound, and where did an EQ cut reduce the body of the voice? What would you change at recording time?

Further reading

• CapCut Help Center
• YouTube - recommended upload encoding settings
• EBU R128 - loudness normalisation
• FFmpeg loudnorm filter documentation
• Wikipedia - LUFS