Mix Preparation and Diagnostics

A practical engineering workflow for AI-assisted productions, separated stems and conventional recordings. Every value below is a listening starting point—not a signature that makes AI audio unrecognizable or a universal mix recipe.

1. Establish a reproducible session

• Keep the native sample rate of the supplied source unless the destination requires conversion.
• Confirm a shared timeline, channel order and duration for every stem. Import at unity and check synchronization.
• Retain unprocessed source copies and an A/B reference with matched perceived loudness.
• Set the listening level before evaluating tonal balance. A louder signal often appears more detailed even when it is not.
• Use clip gain to avoid overloading nonlinear plugins. There is no universal -6 dBFS or -18 dBFS requirement for a floating-point DAW; manufacturer calibration and downstream headroom determine useful levels.
• Keep a timecoded issue log: time / symptom / hypothesis / test / result.

2. Diagnose the cause, not only the frequency

Audible observation	Test before turning a knob	Candidate corrective action
Low-mid muddiness	Solo kick/bass, examine arrangement, room and reverb returns	High-pass only nonessential sub-energy; use a broad, small cut where masking is audible
Harsh consonants or cymbals	Compare at low monitoring volume; check codec or source distortion	Dynamic EQ or de-esser targeted to the actual band
Metallic or watery texture	A/B original against separated stems in context	Replace a stem or repair a short passage rather than adding global filtering
Chorus collapses in mono	Compare mid/side, correlation and fold-down	Reduce excessive widening or correct polarity/alignment
Unstable bass level	Listen to kick/bass interaction over sections	Clip automation or gentle targeted compression; avoid one-size-fits-all low-end rules
Transient smearing	Compare source and post-processing attacks	Reduce limiting/denoising, revise separation or replace damaged events

Frequency clues such as 200–500 Hz for low-mid congestion or 4–9 kHz for some vocal sibilance may help locate a problem. They are not automatic cut/boost bands; use playback and metering to identify the offending source.

3. Controlled processing starting points

Process	Example initial values	What to listen for / stop condition
Mix-bus compression	Ratio 1.5:1–2:1, attack 10–30 ms, release 50–200 ms or auto, about 1–2 dB gain reduction on strong sections	Punch and groove preserved; matched-bypass version not objectively preferable
Vocal compression	Ratio 2:1–4:1, attack 5–25 ms, release 40–150 ms; threshold set by actual performance	Intelligibility without pumping or distorted breaths
De-esser / dynamic EQ	Narrow the detected resonance or sibilant band; apply the minimum attenuation that solves the issue	No lisping or dull consonants
Reverb	Audition short rooms/plates; initial vocal pre-delay 15–40 ms if it improves separation	Source still readable and the decay does not cloud the next phrase
Saturation	Start with very low drive; gain-match before/after	Harmonics support the source without raising intermodulation or harshness
Stereo processing	Compare width, correlation, mid channel and mono fold-down	Center remains stable and important elements do not disappear

Threshold, attack and release are program-dependent. These ranges are examples, not official Spotify/Suno specifications. For a stereo-only file, prioritize controlled EQ, level automation and surgical repair. Recreating independent instruments with source separation is an estimation task and can introduce leakage.

4. A/B testing protocol

1. Define one hypothesis: “the synth masks the vocal in the second chorus.”
2. Loop only the relevant section, but then check the full arrangement.
3. Change one variable. Render or freeze if the plugin is nondeterministic.
4. Match before/after loudness closely; take short blind or alternated listens.
5. Check mono, headphones and a small loudspeaker at moderate listening levels.
6. Keep the change only if intelligibility, balance or emotional impact improves.

5. Example: distorted AI vocal with a good instrumental

Bad shortcut: apply high-shelf cuts, aggressive noise reduction and widening to the complete stereo song.

Better decision tree: determine whether the distortion exists in the original render, separated vocal or master processing. If the source is already damaged, request a revised vocal/render or use a legitimate replacement performance. If it appears after separation, adjust the separator or edit the affected segment. If it appears only after limiting, restore dynamics and check true peak before adding more processors.

6. Technical handoff checklist

• The approved source and unprocessed version are preserved.
• All stems start at the same sample and render over the same time range.
• No unintended clipping, clicks, truncated transitions or misrouted channels.
• Processing choices are documented as choices—not as “anti-detection” settings.
• Mono compatibility and a level-matched reference comparison pass.
• Mix and premaster exports can be identified by revision.

Next: Mastering for streaming, Provenance and release, or the producer's creative mixing guide.