mc-vocal-direction

Vocal Direction and Harmony Design (Vocal Direction)

What this skill governs is: how the vocal shouldbe sung**, not just what notes to sing.**

This is especially important for this library, because—

The most consistent flaw of AI vocals is "perfection": pitch always spot-on, dynamics flat, every chorus sung identically. Real people can't do this, and don't want to.

So half of this skill is teachinghow to demand "imperfection", andhow to write such demands in language the backend can consume。

Which sections to read by task

Symptom / Task	Read
Vocal sounds flat, expressionless	§2 Vocabulary of vocal direction
Want the vocal to have more impact	§2.3 Consonants
Vocal is "too in tune", sounds like a machine	§4 Philosophy of pitch
Vocal won't push through, buried in the accompaniment	§4.1 (it's mostly not a volume issue)
Want to add harmony, don't know which kind	§5 The four types
Harmony overpowers the lead / too grabby	§5 Delivery for the corresponding type + EQ approach
Whether to do a double	§6
Writing vocal prompts for Suno/YuE2	§7
Filling in the vocal block of ARR-SPEC	§8 ＋ §9

Boundaries: what this skill does not govern

Not here	Where it belongs
How to write lyrics, rhyme, inverted characters/harmonized sounds	Lyrics library（`lyric-writing`，LYR-SPEC）
Melodic pitch contour, range design	`mc-melody`（L1）
How words are set to notes (note-splitting/extra syllables)	Lyrics library—it's the lyrics-music interface, belongs to the lyrics library
Specific vocal EQ frequencies, compression parameters	`mc-mix-intent`(intent) / outbound mixing library (practice)
Which instrument the vocal competes with for frequency range	`mc-texture-layering`
Synthesized vocal sources (engine parameters of Vocaloid, Synthesizer V)	Not done. This library only writes vocal intent (§7 maps to each backend), not engine parameters

1. Required before writing

Required	Question
① Who is singing this	Vocal persona: gender, register, texture, era feel.Not just to the level of "female voice"
② Delivery for each section	At least split into verse/chorus/bridge. Write using §2's vocabulary, not "with emotion"
③ Dynamic contour	Relative strength section by section—lint #15 checks this; a single dynamic level across the whole song is an automatic fail
④ Chorus variation per pass	What changed in the second and third passes respectively (lint #16)
⑤ Harmony plan	Which type (§5), in which sections, above or below

2. Vocabulary of vocal direction

2.1 Only two musical criteria

Whether it suits this song

Is it suitable for this singer (似合っているか)

Just these two.Technical skill is not the criterion— technique is a means. Applied to this library, item 2 becomes "Is it suitable for this voice persona"。

2.2 ★ Don't say "sing louder"

This is the most practical item in the whole section:

When a grand arrangement is paired with a muddy vocal,don't say "sing louder"— a volume fader can solve volume. What you actually want is probably "a brighter voice", so saying "please sing brighter" is more effective.

Visual instructionsare equally effective:

"sing with a smile"
"corners of the mouth raised"
"sing toward somewhere far away"

Why this item is especially important for this library: these are exactlythe words the text backend can consume。 sing brighter / smiling delivery / sung toward a far distance are effective prompt words, while louder has almost no meaning to the generative model—it does not control volume.

2.3 ★ Think of consonants and vowels separately

This is common foundational knowledge for both vocal direction and post-production editing:

One syllable = consonant + vowel("か" = K + A, clearest when written in romaji)
When you want impact, saying "put force into the consonants" is more effective than saying "louder"
When editing, distinguish clearly: is thisa consonant timing problemora vowel length problem
In the mix,just controlling the consonant levelcan change the vocal's expression

2.4 The magic consonant "S"

S ≈ white noise made with the throat—it is essentially noise.

Corollaries (all very practical):

You can cut in the middle of an S without hearing a seam
Its length can be freely stretched or shortened without feeling wrong
Transplanting an S from another take is basically seamless

★ The beat lives on the vowel, not on the consonant. When "さ" (= S + A) is sung on beat 2, the S actually occurs on the tail of beat 1— what aligns with the beat is the vowel A.

This item has two uses：

When editing AI-generated vocals,"there's an S here" = "this is a good edit point"
When judging whether a vocal is "rushing/dragging,"look at the position of the vowel, not the start of the syllable— otherwise you will misjudge every word beginning with S, SH, or TS as rushing

3. Analogy of the production chain

A useful comparison (film → pop music):

Film	Music
Script	Lyrics and music
Sets and costumes	Arrangement
Actor	Singer
Hair and makeup	Live vocal direction
Cinematography	Recording
Editing	treatment (vocal editing)
MA	Mixing

"Live direction + recording + treatment"—only the three together areVocal direction in the broad sense。

Significance for this library: When using a generative model, the "recording" stage disappears, but **both the "hair and makeup" (direction) and "editing" (refinement) stages remain, and are even more important**— because the only place you can control the model is direction, and the only place you can fix things is refinement.

4. ★ Pitch philosophy: the most counterintuitive section of this skill

4.1 Two empirically tested lessons

A vocal that sings flat won't "come out" no matter how high you push the fader

A vocal that is slightly sharp has better "nuke" (penetration)—good singers unconsciously sing slightly high

The first point explains a common misjudgment:When the vocal is buried in the accompaniment, check pitch first, then volume。

4.2 Don't fix everything

"It's in tune, therefore it's well sung" is incomplete; the reverse is what's true:People who sing well are, by and large, in tune.

So:

In sparse accompaniment, moments that are flat or sharpare themselves "flavor"
Places you want to emphasizebeing sharp is acceptable—not pulled up after the fact, butdeliberately preserved those places where the singer "couldn't help singing high in the emotion"
It's not aligning everything, but deliberately leaving part of it untouched—that is natural pitch correction.

This is the core vocal rule of this library for fighting AI flavor. The vocals output by generative models are naturally perfectly in tune; what you need to do is not "tune them more accurately," butspecify where inaccuracy is allowed。

How to write it: In the ARR-SPEC's vocal.pitch_policy , write "which positions retain offset," rather than writing "tune accurately."

Failure condition: In sections with extremely dense arrangement and thick stacked harmonies, pitch deviation becomes muddiness rather than flavor— those places should be fixed.The criterion is accompaniment density, not section name.

4.3 The three-step order of treatment

Clean up the junctions between multiple takes
Align the rhythm
Fix the pitch

Why this order: If you move the rhythm (shift time) first and then join, the junctions have to be redone.

5. The four harmony types

Each type is defined by three things: purpose / delivery / how to keep it from stealing from the lead vocal. Note that delivery is part of this skill, while EQ is part of mc-mix-intent —here we only give the approach.

Type 1: Singing-up upper harmony (full-voice upper harmony, usually a third)

Purpose: Keep the main melody unchanged and make the wholebrighter and more brilliant
Internal contradiction: It is higher than the lead vocal and naturally stands out, yet it must not overpower the lead vocal
★ Solution：Don't solve it by "singing softer"—singing softer loses the brilliance. Keep the full-voice delivery and solve it with EQ(Approach: make the harmony "escape" from the lead vocal's frequency range— cut the lows heavily, scoop out the midrange "core," and if necessary emphasize the highs instead)

Type 2: Upper harmony that hugs the lead vocal

Purpose: Not brilliance, butCreate an airy feel and bring out the sense of chords
★ The key is changing the vocal delivery: not an open sound, buta light, breathy, clean vocal delivery
Note: the volume will drop, butdo not compensate by moving closer to the mic(the proximity effect will boost the low end); use gain control

Type 3: Restrained parallel lower harmony

Image："the shadow cast by the lead vocal". The illustration in the original book is exactly the singer and their own shadow: the vocal delivery and diction should match the lead vocal as closely as possible, and the shadow itself has no independent expression (p0186)
Reduce presence: cut the low end. The reference EQ given in the original book (6-band EQ): low cut at about 110 Hz, then add a −10 dB cut at 123 Hz, a −4.9 dB cut at 247 Hz, and a −4.2 dB cut at 621 Hz. This is an overall reduction from low to low-mid frequencies, not a single-point notch (p0186)

Type 4: Aino-kote (short response phrases between lines, sung in 2–3 voices in unison)

Positioning: it is a **"different object"** from the lead vocal
Approach：first reach for the EQ to change the texture (a radio tone is fine too; just cutting the low end can change the atmosphere)
Improve the unity of the multiple vocal parts: align levels with faders + compression → put the same spatial effect on this group → or even have the whole group share one modulation effect (chain: EQ → compression → chorus → delay)

Reverb is also part of the model

Types 1 and 2 can usereverb/delay rich in high-frequency contentto add brilliance or airiness; this extends the idea of "escaping the lead vocal's frequency range" from EQ to spatial effects (p0186)
Type 3 reduce direct sound and increase the reverb componentcan reduce presence; as the dry/wet ratio moves toward wet, the sense of physicality that "this person is really singing here" fades accordingly (p0186)
The spatial processing goal for Type 4 is tightness: give the whole group the same spatial effect to unify the texture, see Type 4 in this section. For how to place the reverb's panning together with the lead vocal's panning, see §6.2

6. Double and Panning

6.1 Doubling the lead vocal: a clear trade-off

Doubling a lead vocal with excellent diction will dilute the feel of the words

but the stereophonic solidity unique to doubling bringsa warmth and thickness that machine replication cannot produce

★ Note the phrase "machine replication cannot produce"—it also explains why "two tracks" output directly by a generative model sound fake: that is replication, not a second performance. For a true double, the backend mustgenerate a second time, not copy one track and tweak it.

Key operational points：

The lead vocal take must be chosen first; the double is sung while listening to the lead vocal and sticking close to it
But choosing slowly makes the singer wait idle—it is better to record the double immediately while the singer still remembers the diction(a time vs. quality trade-off)
For songs with complex diction:record carefully in segments of a few lines at a time
Checking method: hard-pan the two tracks left and right, ormute the lead vocal, and strictly check whether the double sticks to it
Guiding principle：Don't let the lead vocalist's enunciation get erased—have the singer usea delivery that emphasizes enunciationto record the double, and it fits perfectly

6.2 Panning strategy

Form	Approach
Single	If there's no particular reason, place itdead center; use volume, reverb, and delay to create distance and depth from the lead vocal; for harmonies in extremely high/low registers considerswitching to a different microphone than the lead vocal
Double	Can be spread left and right;the more extreme the panning, the more it stands out, and it's emphasized even at low volume—pan the ones you want to emphasize far out. When both the upper and lower harmonies have doubles,pan only one of the pairs, if you pan both pairs there are too many elements. If only the harmonies are wide in the arrangement it will seem odd (you can use that deliberately),the reverb's panning must be considered together
Triple	Understand it as "single + double": the single handles depth, the double handles width;pull down the single's volumecleaner

6.3 Guidance for harmony recording

The lead vocalist sings the main melody well, but singing harmonies is unexpectedly hard—because the harmony is subtly similar to the main melody, it's hard to remember。

A guide track (ガイド) is the most commonly used solution：

Create a new track, usea soft timbre with a weak attackto play out the harmony line
Tips:deliberately play it an octave higherso it's easier to hear; make a 1-bar line into a 2-bar loop; you canpan it to only one side of the headphones
After the singer is familiar with itgradually lower the guide track volume, then record the real take
The most effective thing is actually to sing it for them—instrument timbres and the human voice are completely different things, it's extremely difficult to ask a lead vocalist who doesn't play an instrument to "mentally swap the instrument notes for the human voice," especially when the syllable arrangement of the lyrics is complex

Mapping to this repository: when using a generation backend, the "guide track" corresponds togiving the model a reference melody or reference audio. Same reasoning:giving the model human vocals to listen to is more effective than showing it instrument MIDI。

7. Mapping to each backend

Backend	What it can take in	How to write it
Suno / MiniMax	Natural-language delivery description	Use the vocabulary from §2.2:`brighter`、`breathy`、`harder consonants`、`smiling`。Don't write `louder`
YuE2	Same as above, and can differ per section	Writing the delivery per section lines up perfectly with §1's required item ②
MIDI / DAW (recorded by a real person)	The complete §2–§6	Use it directly as a direction sheet

Two things common to all backends：

Dynamics must differ per section（lint #15）
Each pass of the chorus must have variation(lint #16)—add harmony, change the delivery, add a double, choose at least one of the three

Handling backend capability differences: if some backend doesn't accept delivery descriptions at all (honors: none), when calculating the compliance rate you shouldexclude itthese items, rather than giving a score of 0—see the backend profile rules in L4 for details.

8. Correspondence with ARR-SPEC fields

Field	Which section of this skill
`vocal.persona`	§1①
`vocal.delivery`(can be paragraph by paragraph)	§2
`vocal.dynamics_by_section`	§1③，lint #15
`vocal.chorus_variation`	§1④，lint #16
`vocal.pitch_policy`	§4.2 (writewhere to preserve the offset, not "correct it")
`vocal.stacking`	§5 four types + §6 double (including type / paragraph / delivery / panning)

9. Must pass after writing

The vocal persona is specific enough toimagine who it is, not "female voice"
Each paragraph has its own delivery, using executable words like "bright/breathy/hard consonants"
Did not write "louder"
Dynamicsdiffer paragraph by paragraph（lint #15）
Choruschanged something each time, and can say what was changed (lint #16)
pitch_policy What is written iswhere inaccuracy is allowed, not "correct it"
The harmony chose a model, and that model's delivery and "how not to steal from the lead vocal" are both written
Type 1 harmonydid not avoid by "singing softly"
If doing double, what is required isa second vocal take, not duplicating one track
When both upper and lower harmonies have doubles,only threw one pairto extreme panning
When judging whether vocals rush/drag, look atvowels' position

Appendix: Sources

Japanese package jp-vocal-lyrics-production—§2, §4, §5, §6 all main trunk. The original book is from a Japanese pop music production practice series.
The EQ reference example for §5 type 3 and the three points of "reverb is also part of the model" come from 《プロの曲作りが分かる本》(p0186).
The lyrics-music alignment part (符割り／字余り／イントネーション) is not in this skill, belongs tolyric library lyric-writing, because it is a lyric process.
The specific parameters of the vocal mixing chain (compression ratio, EQ frequencies, the 309.6 Hz example of pop noise removal, etc.) are organized in reference.md, with practical output going to the mixing library.
For the judgment methods of self-checks #15/#16, see mc-workflow §3.0。