← overview

LeVo 2 — prompting it for one voice, in the right genre

An evolutionary prompt search over LeVo 2 / SongGeneration-v2-large. Two independent failures — "opera" coming out as auto-tuned pop, and a backing choir arriving in clips that were supposed to be one person alone — located, explained and fixed by prompting. 2523 clips judged by Gemini 3.6 Flash on three axes. All audio 160 kbps stereo.

66% → 98%one-voice pass rate, 4 seed genres
71% → 95%genre pass rate, 4 seed genres
43% → 54%solo & clean pass, 50 genres
2523clips judged, 3 axes each

Why it happens — two different causes

1. The genre miss is an out-of-vocabulary problem. LeVo's descriptions field feeds the type_info conditioner, and the model ships its actual tag vocabulary in sample/description/: 2 genders, 27 genres, 8 emotions, 7 timbres. opera is not in it. Neither is choir, gospel, hymn or a cappella. Out-of-vocabulary words are not rejected — they are conditioned on weakly, and the model falls back to its prior, which is pop.

description given to LeVoopera Gchoir Gpop Gcountry Greading
(empty string)1.672.174.904.70With no conditioning at all, LeVo sings pop. That is why removing the prompt costs opera and choir everything and costs pop and country nothing.
female, opera — the out-of-vocabulary word3.17The word a human would write is the worst opera prompt in the search.
female, classical — the in-vocabulary anchor4.834.17Two words, both in the tag list. Best opera prompt found.
2. The backing choir is a chorus-section phenomenon. The benchmark already showed this could not be fixed downstream: LeVo scored B = 4.07 on the mix and 4.07 again after stem separation, while every other model's B rose once instruments were stripped. Nothing instrumental was there to remove — the non-compliance is a second voice. The search localised it: of the clips scoring S ≤ 3, the judge names the chorus unprompted in 50 % of cases.

"A female voice sings the verse alone, but multi-part vocal harmonies join in during the chorus"
"I hear a male singer start solo, but at the chorus (0:15) a full stack of vocal harmonies joins in behind him"
"The clip begins with a solo female voice, but at 0:08 a choir/backing vocal ensemble enters"

[chorus] is a learned cue for stacked, doubled, harmonised vocals, because that is what choruses are in the training corpus. Delete the chorus and the choir goes with it.

The recipe

descriptions = "<gender>, <IN-VOCABULARY genre tag>, <free style words>"
lyrics       = "[verse] Line one. Line two. Line three. Line four. Line five. Line six."
gen_type     = "vocal"
temperature  = 0.9    top_k = 50    top_p = 0.0    cfg_coef = 2.0
  1. Anchor the genre on one of LeVo's 27 genre tags, then add your real style words after it. classical carries opera; opera alone does not.
  2. One [verse] block. No [chorus]. If you need two sections, tag the second one [verse] as well.
  3. Never write [intro-*], [inst-*] or [outro-*]. They are instrumental by definition and they put instruments back into a vocals-only render.
  4. Do not write negations and do not write a cappella. Both cost score. The solo is bought by structure and by gen_type='vocal', not by asking.
  5. CFG 2.0. It is the genre-fidelity knob; 1.0 is as bad as an empty prompt.

The four seed genres — before and after

Baseline is the prompting the pipeline uses today. Every cell is the same lyrics and the same seeds; solo % is the fraction of takes the judge scored 4–5 for "exactly one vocalist", which is the number that matters because the failure is all-or-nothing rather than gradual.

genrearmnS one voiceG genreI no instr.solo %clean %genre %prompt / structure
operabaseline604.223.705.0073%73%60%female, classical, opera, emotional, a cappella, no instruments
vc · cfg 1.5 · top_k 50
evolved364.754.445.0092%92%86%female, classical
v_only · cfg 1.5 · top_k 250
choirbaseline603.433.334.9852%52%42%female, gospel, choir, uplifting, a cappella, no instruments
vc · cfg 1.5 · top_k 50
evolved244.964.885.00100%100%96%female, classical, hymn, sacred, emotional, a cappella, solo singer, single voice
v_only · cfg 2.0 · top_k 50
popbaseline574.355.004.8474%68%100%female, pop, emotional, a cappella, no instruments
vc · cfg 1.5 · top_k 50
evolved125.005.004.75100%92%100%female, pop, pop rock, emotional
vc · cfg 1.5 · top_k 50
countrybaseline603.984.334.8065%62%82%male, country, emotional, a cappella, no instruments
vc · cfg 1.5 · top_k 50
evolved305.005.005.00100%100%100%male, country, solo vocal
v_only · cfg 1.5 · top_k 50
One honest wrinkle. The choir winner the search converged on still carries a negation stack, which contradicts the advice above. It got there because the negations are inert for choir, not because they helped — the controlled probe below tested exactly that and found no effect in either direction. The load-bearing parts of that prompt are the classical anchor, the sacred style word, the single [verse] and CFG 2.0. Note also that choir itself must never appear: what is wanted is choral style sung by one person, and the baseline's gospel, choir is why it was the worst of the four.

Listen — worst baseline take vs best evolved take

Per genre: the baseline take the judge scored lowest for "one vocalist", a typical baseline take, and the best take from the evolved prompt. What the judge wrote is printed next to each.

genrearmaudioSGIwhat the judge heard
operabaseline — worst1.003.005.00A female soprano sings a classical operatic melody alone for the first half, before a full classical choir joins in for the chorus, with no instruments present.
baseline — best5.005.005.00I hear a single female classical opera singer performing the requested lyrics a cappella with natural room reverb and no instrumental accompaniment.
evolved — best5.005.005.00I hear a single operatic soprano voice singing classical lyrics with traditional bel canto technique and vibrato, completely unaccompanied.
evolved — worst2.005.005.00A trained classical female soprano sings the aria completely a cappella until the last three seconds, where a full classical choir suddenly enters singing 'Silent hour descend'. There are no instrumental sounds present at all.
choirbaseline — worst1.002.005.00The clip begins with a solo female voice, but at 0:08 a choir/backing vocal ensemble enters to provide full gospel harmonies behind the lead singer; there are no musical instruments present.
baseline — best5.005.005.00A single female voice singing a modal liturgical hymn unaccompanied, with cathedral reverb and no additional vocal layers or instruments present.
evolved — best5.005.005.00A single female voice sings a sacred liturgical hymn completely unaccompanied, with no additional voices or instruments present.
evolved — worst4.005.005.00A solo female soprano sings the sacred hymn completely unaccompanied with cathedral reverb, followed by a brief layered vocal phrase at the very end.
popbaseline — worst2.005.003.00I hear a female contemporary pop voice with rhythmic finger snaps in the verse, joined by stacked backing vocal harmonies during the chorus.
baseline — best5.005.005.00I hear a single female voice singing a contemporary pop melody completely solo with no backing vocals or instrumental accompaniment.
evolved — best5.005.005.00A single female voice sings the prompt's lyrics acapella with a contemporary pop delivery, featuring no extra vocal layers or instruments.
evolved — worst5.005.002.00A single male voice singing modern pop with rhythmic finger snaps in the background, but no other vocal parts or melodic instruments.
countrybaseline — worst1.002.005.00I hear a male lead vocalist singing in a pop/R&B style, joined in the chorus by a stacked vocal choir performing rhythmic 'oh-oh' vocal pads and harmonies, with no musical instruments present.
baseline — best5.005.005.00A single male country voice singing a cappella with natural vocal reverb and no instruments or second voices present.
evolved — best5.005.005.00A solo male country voice singing the lyrics acapella with no backing vocals or instruments present.

Lyric structure is the biggest single lever

Generation 2 swept the section scaffold with everything else held fixed (n = 6 per cell). Score shown is S — exactly one vocalist.

structureoperachoirpopcountryverdict
[verse] only, all lines in one block4.835.004.505.00best
[verse] ; [verse] (chorus retagged)5.005.004.674.67as good
[verse] ; [chorus] (the obvious thing)4.334.334.174.83baseline
[silence] ; [verse] ; [chorus]5.003.675.005.00erratic
[verse] ; [bridge]2.675.004.673.33erratic
[intro-short] ; … ; [outro-short]4.503.334.173.67hurts
[verse] ; [inst-short] ; [chorus]2.004.674.174.50worst
[inst-short] is the worst single intervention measured — it takes opera from S 4.33 to 2.00. An [inst-*] section is instrumental by definition; asking for a vocals-only render and then writing an instrumental section into the lyrics is a contradiction, and the model resolves it by adding instruments and voices back.

Negations backfire — the controlled test

Each row adds one phrase to the same base prompt, same seeds, everything else identical. The two rows marked manipulation check add the unwanted thing positively: if those make it worse, the model demonstrably responds to the phrase, and a negation that fails to help is evidence about how negation is handled — not evidence that the words are ignored.

probe: acappella

phrase addedopera Schoir Spop Scountry Smean Smean G
(base)5.005.004.175.004.794.71
+ 'a cappella'3.675.003.835.004.384.12
+ 'unaccompanied'5.005.003.675.004.674.92
+ 'solo voice'5.004.833.835.004.674.71
+ 'no instruments'5.005.004.505.004.884.88
+ 'piano' [manipulation check]5.005.004.334.674.754.79
+ 'a cappella, solo voice, unaccompanied'5.005.004.504.504.754.54
BASELINE (control)3.334.004.174.173.924.17

probe: negation

phrase addedopera Schoir Spop Scountry Smean Smean G
(base)4.835.005.005.004.964.75
+ 'backing vocals' [manipulation check]3.175.004.005.004.294.38
+ 'choir' [manipulation check]5.005.002.835.004.464.50
+ 'no backing vocals'5.004.502.835.004.334.71
+ 'no choir'5.005.003.005.004.504.54
+ 'no backing vocals, no choir, no harmony, no doubling'4.675.003.675.004.584.75
+ 'solo singer, single voice' (positive)5.005.003.005.004.504.75
BASELINE (control)3.504.004.174.173.964.25

What did not work

Transfer — 50 cluster-representative genres

The 50 medoids of a Ward clustering of the 153-genre taxonomy. Both arms get identical lyrics and identical seeds; base is the natural-language a-cappella caption the pipeline uses today, evo is the recipe above instantiated per genre. solo % = takes scored 4–5 on one-vocalist and 4–5 on no-instrumentation. PASS = solo % ≥ 83 and mean genre ≥ 4.

The recipe transfers, but it does not solve this set. It roughly triples strict solo compliance (3 → 10 of 50 genres) and doubles genre compliance (11 → 21 of 50) — and still leaves 46 of 50 failing one axis or the other (1 → 4 pass both). Mean S 3.44 → 3.94, mean G 2.84 → 3.23, mean I 4.84 → 4.67. Do not read the seed-genre result as the general case — the four seeds are broad, Western, slow and melodic, which is the easy end of this taxonomy.
familynbase clean%evo clean%base Gevo Gverdict
blues258%75%2.753.50solo transfers, genre does not
theatre242%67%2.502.50solo transfers, genre does not
folk339%67%3.783.72solo transfers, genre does not
metal429%62%2.793.67solo transfers, genre does not
classical750%60%2.193.10solo transfers, genre does not
functional523%59%2.302.55solo transfers, genre does not
world1150%52%1.791.53does not transfer — genre got worse
rock553%50%3.754.73flat
soul247%50%4.624.75flat
pop520%41%4.174.75transfers
urban462%38%3.383.88regresses
genrefamilybase S/G/Ibase solo%evo S/G/Ievo solo%Δverdictbaseline takeevolved takecluster it represents
Sámi joik
sami joik, cyclical wordless vocable chant, guttural and hypnotic
world3.3/1.2/5.033%4.5/1.0/4.567%+33FAIL8 genres: elderly_folk, georgian_solo, hawaiian_falsetto, opera_basso, sardinian_cantu, sean_nos, sertanejo
Andean huayno
andean huayno, very high piercing nasal quechua-inflected tone
world3.7/3.5/5.050%3.5/1.2/4.033%-17FAIL7 genres: ambient_wordless, min_yo, neo_soul, opera_sop_coloratura, pansori, whisper_sing
Bedroom pop
bedroom pop, whispered soft double-tracked intimacy, unforced small voice
pop2.3/4.0/4.517%3.2/4.4/4.220%+3solo fail7 genres: bossa_nova, gamelan_sindhen, lullaby, lullaby_grandmother, raga_bhajan, vocal_fry_sing
Keening / lament
ritual keening lament, sobbing wailing cries, grief-stricken and unmetred
functional2.7/1.0/4.70%4.2/1.3/4.767%+67FAIL6 genres: baroque_ornament, carnatic, enka, rnb_contemporary, trance_vocal
Bulgarian diaphonic
bulgarian village singing, bright nasal forward tone, tight ornament shake
world4.2/1.7/5.067%3.5/1.0/4.350%-17FAIL6 genres: country_pop, dnb_vocal, gospel_lead, inuit_katajjaq, schlager
Oratorio / sacred solo
oratorio soloist, clean sustained sacred line, minimal vibrato, devotional
classical3.0/2.5/5.033%4.2/3.2/4.550%+17FAIL6 genres: byzantine, canzone, klezmer_krekhts, opera_tenor_lirico, peking_opera
Teen pop
teen pop, bright youthful thin-bodied voice, eager and unweathered
pop3.2/3.8/5.033%2.8/4.7/5.033%0solo fail6 genres: citypop, cumbia, hyperpop, kpop, musical_belt
Field holler
unaccompanied field holler, long falling cries, free rhythm, raw and open-
blues4.0/3.7/5.067%4.5/2.5/5.083%+17genre fail6 genres: flamenco_cante, jazz_crooner, outlaw_country, soul_classic, uk_drill_sung
Samba
samba, buoyant percussive syncopated phrasing, open joyful tone
world3.5/2.0/4.533%2.5/2.3/5.017%-17FAIL6 genres: gospel_bluegrass, gregorian, shape_note, synthpop_80s, throat_overtone
Cantorial
cantorial chazzanut, virtuosic ornamented liturgical cries, huge emotive r
world4.0/3.2/5.067%4.5/2.8/5.083%+17genre fail5 genres: arabic_mawwal, persian_avaz, spiritual, synthwave
Mezzo-soprano
mezzo-soprano, warm dark middle register, rich operatic legato
classical3.0/2.0/5.033%4.5/3.2/5.083%+50genre fail5 genres: balkan_open, fado, kulning, musical_legit
Sprechgesang
sprechgesang, half-spoken half-sung expressionist delivery, pitched speech
classical4.5/2.7/5.083%4.2/3.0/5.067%-17FAIL5 genres: countertenor, humming, rebetiko, vaporwave_croon
Beatboxing
beatboxing, percussive mouth drum sounds, rhythmic and unpitched
functional2.5/1.2/4.50%3.2/1.0/4.017%+17FAIL4 genres: melodic_rap, vocalese, whistling
Maximum belt
maximum-effort belting, full-throated sustained loud chest voice, strained
functional3.5/3.2/4.550%4.6/3.6/4.460%+10FAIL4 genres: europop_dance, house_diva, pop_ballad
Punk
punk, shouted snarling barked delivery, aggressive and untrained
rock3.2/3.7/5.033%3.7/4.8/5.033%0solo fail4 genres: grime_hook, hardcore_scream, trap_sung
Clean metal baritone
clean metal baritone, dark brooding sustained tone, gothic gravity
metal3.5/2.8/4.733%4.7/4.2/4.067%+33solo fail4 genres: industrial_shout, noh_utai, vaudeville
Operatic baritone
operatic baritone, dark resonant chest voice, noble sustained line
classical4.0/1.5/5.067%3.3/4.3/5.050%-17solo fail4 genres: jazz_scat, rock_classic, tango
Appalachian ballad
unaccompanied appalachian ballad, plain hard modal tone, ornamented free r
folk2.8/3.0/5.017%3.5/1.8/5.050%+33FAIL3 genres: blues_classic_female, cantonese_opera
Roots reggae
roots reggae, nasal plaintive tone, off-beat phrasing, prophetic delivery
urban4.7/4.5/5.0100%4.0/5.0/4.550%-50solo fail3 genres: blues_chicago, salsa_sonero
Saeta
spanish saeta, unaccompanied processional cry from a balcony, raw and pier
world3.5/1.0/4.750%3.8/1.2/3.840%-10FAIL3 genres: chanson, torch_song
Playground chant
children's playground chant, sing-song taunting cadence, shouted and rhyth
functional4.2/2.5/4.750%4.2/2.3/5.067%+17FAIL3 genres: child_folk, nursery_rhyme_child
Lovers rock
lovers rock, sweet gentle reggae-soul, tender high register
urban4.3/4.5/5.067%4.0/3.5/4.233%-33FAIL3 genres: jazz_standard, scots_gaelic_waulking
German Lied
art-song lieder delivery, intimate recital voice, careful diction, restrai
classical2.7/2.2/4.817%4.7/2.5/5.083%+67genre fail3 genres: muezzin_style, qawwali
Black metal shriek
black metal shriek, high rasping strangled screech
metal2.5/0.8/4.517%4.2/3.3/4.767%+50FAIL3 genres: metal_death, turkish_uzun_hava
Singer-songwriter
singer-songwriter, plainspoken confessional delivery, small honest untrain
folk3.5/4.5/4.750%4.3/5.0/5.083%+33PASS2 genres: post_rock_wail
Patter song
comic patter song, extremely fast precise consonants, breathless wit
theatre3.0/1.0/4.333%5.0/1.3/4.783%+50genre fail2 genres: auctioneer
Tuvan throat singing
tuvan khoomei throat singing, low fundamental drone with whistling overton
world3.0/1.2/4.517%3.5/0.7/3.833%+17FAIL2 genres: bluegrass_high_lonesome
Delta blues
delta blues, rough weathered voice, bent notes, moaning field-holler infle
blues3.5/1.8/4.750%4.2/4.5/5.067%+17solo fail2 genres: cowboy_yodel
Dancehall
dancehall, rapid patois toasting sliding into melody, rhythmic and percuss
urban3.7/1.3/5.050%3.7/2.2/4.717%-33FAIL2 genres: boy_treble
Cabaret
cabaret, arch knowing delivery, spoken asides sliding into sung line
theatre3.3/4.0/5.050%3.7/3.7/5.050%0FAIL2 genres: maori_waiata
Ranchera
mexican ranchera, huge open-throated grito, proud sustained belt
world4.5/1.7/5.083%4.0/1.0/4.250%-33FAIL2 genres: country_classic
Emo
emo, cracking earnest voice, whiny nasal edge tipping into desperate yells
rock3.2/3.8/5.033%3.8/5.0/5.050%+17solo fail2 genres: teen_choir_solo
Hindustani khayal
hindustani khayal, slow raga exposition, microtonal meend glides, sustaine
world3.0/0.8/5.033%4.5/0.8/4.567%+33FAIL2 genres: ghazal
J-pop
j-pop, high clear cute-inflected tone, wide melodic leaps, earnest
pop3.2/4.7/5.033%3.2/4.8/4.517%-17solo fail2 genres: teen_bedroom
Gothic metal soprano
gothic symphonic metal soprano, operatic vibrato over heaviness, ethereal
metal3.2/4.5/5.033%4.0/2.2/4.250%+17FAIL2 genres: opera_sop_dram
Sea shanty
sea shanty, hearty rolling call, blunt communal projection
folk3.3/3.8/5.050%3.8/4.3/5.067%+17solo fail2 genres: mongolian_urtiin
Glam rock
glam rock, theatrical androgynous sneer, camp vibrato
rock4.0/4.5/4.550%2.5/5.0/5.017%-33solo fail2 genres: work_song
Afrobeats
afrobeats, lilting melodic pidgin-inflected phrasing, relaxed buoyant swin
urban3.5/3.2/4.733%3.8/4.8/5.050%+17solo fail1 genres:
Bel canto
bel canto, effortless even legato across registers, elegant messa di voce
classical3.5/2.7/4.850%2.5/2.7/5.017%-33FAIL1 genres:
Bolero
latin bolero, velvet romantic croon, dramatic sustained romantic line
world4.0/2.3/5.067%4.2/3.2/4.550%-17FAIL1 genres:
Britpop
britpop, adenoidal regional-accented drawl, cocky flat delivery
rock4.0/3.2/5.067%4.5/4.0/4.583%+17PASS1 genres:
Chorister solo
solo chorister boy, pure vibrato-free treble, echoing cathedral clarity
classical4.0/1.8/5.067%4.0/2.8/5.067%0FAIL1 genres:
Doo-wop lead
doo-wop lead, sweet earnest tenor, swooping falsetto leaps, vintage warmth
soul3.8/4.4/5.060%3.7/4.7/5.050%-10solo fail1 genres:
Pure falsetto
pure falsetto, light hooty unsupported head voice, fragile
functional2.8/3.7/5.017%4.3/4.5/5.083%+67PASS1 genres:
Forró
forro, warm northeastern brazilian accent, rustic bouncing delivery
world3.5/1.2/4.850%5.0/1.7/4.783%+33genre fail1 genres:
Grunge
grunge, strained throaty rasp, apathetic verses erupting into raw shouting
rock4.4/3.6/5.080%5.0/4.8/4.067%-13solo fail1 genres:
Indie pop
indie pop, slightly nasal untrained charm, close-mic breathiness, casual
pop2.8/4.7/5.017%3.3/4.8/5.050%+33solo fail1 genres:
Power metal
power metal, soaring operatic high tenor, sustained ringing screams
metal3.3/3.0/4.233%4.2/5.0/5.067%+33solo fail1 genres:
Boy-band tenor
boy-band lead, light high tenor, earnest and smooth, adolescent brightness
pop1.8/3.7/5.00%4.5/5.0/4.883%+83PASS1 genres:
Quiet storm
quiet storm R&B, smooth low seductive croon, silky sustained falsetto
soul3.3/4.8/4.833%3.8/4.8/4.750%+17solo fail1 genres:

Method