AIMC Notes · Case Study 09 · Controlled Voice Swap

Sound Chronicles Theme — Same Song, Different Voice

Sound Chronicles Theme. One song. Three audio files. The instrumental never changes — only the voice does. What do machine listening tools hear when an AI voice is replaced by a human singer?

Alex Tolkmit · EUMAK4U · Berlin · Measurements October–November 2026

Same notes. Different silence. The AI left room to breathe — the human filled it.

1 · Question

If a human singer re-sings an AI-generated vocal over the identical instrumental, which differences do Cyanite (audio tagging) and Klangio (transcription) report — and which of those differences come from the voice, and which from the tool itself?

2 · Setup

FileContentRole
A · Min (AI)Sound Chronicles Theme, original release mix. Vocal by the AI persona Min (Suno), incl. Suno backing vocals.AI condition
B · Min – Human AvatarSame song, sung by a professional session singer (R&B/Soul, booked via Fiverr) over the instrumental of A.Human condition
C · InstrumentalThe instrumental of A without lead and backing vocals, mixed from the Suno stems. This is exactly what the singer received.Control

3 · Results — Cyanite (three-way)

3.1 Tempo & key

C · InstrumentalA · MinB · Min – Human Avatar
BPM7575150
KeyG majorG majorG major

Finding: identical instrumental, identical tempo — but the human version is reported at double tempo. The voice alone triggered a tempo octave error.

3.2 Instruments

InstrumentC · InstrumentalA · MinB · Min – Human Avatar
Acoustic GuitarThroughoutThroughoutThroughout
Electric Piano—Throughout—
Electric Guitar—Frequently—
Synthesizer—Frequently—
Bass GuitarFrequentlyFrequentlyFrequently
Percussion—FrequentlyFrequently
Drum KitPartiallyFrequentlyFrequently
PianoPartially—Frequently
StringsPartially—Partially

Finding: with the AI voice on top, Cyanite reports four instruments that are not in the instrumental (electric piano, electric guitar, synthesizer, percussion). With the human voice, it reports one (percussion) and upgrades piano and drums. The voice changes what the tool believes the band is playing.

Cyanite instrument detection, three conditions

3.3 Mood

MoodC · InstrumentalB · Min – Human AvatarA · Min
Romantic0.730.850.89
Chill0.570.730.79
Calm0.360.460.49
Sad0.310.390.45
Happy0.200.350.25
Ethereal0.110.190.19
Sexy—0.090.12
Uplifting0.070.110.07
Epic0.05——

Finding: both voices raise romance and calm above the instrumental; the AI voice raises them most. The human voice is the only condition that pushes “happy” up.

Cyanite mood scores

3.4 Emotional profile & energy

C · InstrumentalA · MinB · Min – Human Avatar
Emotional ProfileBalancedBalancedBalanced
Emotional DynamicsMediumMediumLow
Energy LevelLowLowLow
Energy DynamicLowLowLow
Valence curve≈ −0.4 → peak +0.4 (1:30) → −0.25starts ≈ −0.6, rises to ≈ 0starts ≈ −0.3, positive from 0:30, peak ≈ +0.35 (2:00)
Arousalnegative throughout (≈ −0.25 to −0.48)negative throughoutnegative throughout (≈ −0.33 to −0.50)

Finding: the AI voice keeps the emotional movement of the music (Medium) and starts darker. The human voice flattens it (Low) and lifts the whole song into positive valence.

3.5 Genre

A · MinB · Min – Human Avatar
Main genresPop 0.42 · Folk/Country 0.39Pop 0.32 · R&B 0.28 · Folk/Country 0.26
Free genresPop · Acoustic · ChillPop · Acoustic · Chill

Genre for the instrumental was not captured.

Finding: R&B appears only with the human singer — although he imitated the AI melody note for note. His own style is audible to the tool even when he tried to hide it.

3.6 Voice detection

C · InstrumentalA · MinB · Min – Human Avatar
Predominant voiceNoneMaleMale
Voice presenceNoneHigh (≈ 1.0)High (≈ 1.0, dip to 0.78 at 2:15)
“Female” at 2:15≈ 0.08≈ 0.6≈ 0.24
False voice in instrumentalmale up to 0.23 · female up to 0.14——

Finding: at 2:15 both singers go high, and the tool reads a high male note partly as a female voice. The control shows no such spike. In the instrumental, the tool still reports a faint “male” voice where there is none.

4 · Results — Klangio transcription (vocal stems)

Min (AI)Min – Human Avatar
Key · tempo marks1 sharp (G) · ♩ 63 / 79 / 631 sharp (G) · ♩ 63 / 79 / 62
Notes202222
Rests (in beats)88.367.9
Held notes (≥ 2 beats)18
Repeated same pitch5971
Leaps (≥ 5 semitones)2936
Notated rangeD3 – A4D3 – A4
Most frequent pitchesG · B · A · D · EG · B · A · D · E

Finding: the written skeleton is almost identical — same key, same range, same pitch hierarchy. The difference is in the space: the AI leaves more rests and moves on; the human holds notes and fills the gaps.

Klangio comparison

5 · Results — pitch analysis (between the notes)

Own measurement on source-separated vocal stems (pitch tracking, frame-level). Range = 5th–95th percentile.

Min (AI)Min – Human Avatar
Sung rangeD3 – F♯4 (16 semitones)E3 – E4 (12 semitones)
Median pitchB3 (247 Hz)A3 (223 Hz)
Pitch changes per second10.67.6
Phrases5133
Breaths detected2420
Dynamic range (P10–P90)10.4 dB9.0 dB
Notes with vibrato42 / 6046 / 75
Vibrato rate · depth5.7 Hz · 84 cents6.1 Hz · 92 cents
Jitter · shimmer · HNR1.11 % · 6.0 % · 17.6 dB0.78 % · 5.6 % · 16.7 dB
Spectral centroid1,939 Hz (darker)2,291 Hz (brighter)

Finding: Klangio rounds every bend to the nearest note, so both singers look the same on paper. Between the notes, the AI voice moves more: wider range, more pitch movement, more phrases, more breaths, more dynamics. Where Min bends or rises (0:13, 0:20, 2:00, 2:51), the Min – Human Avatar sings straight; where Min stays plain (2:45), the Min – Human Avatar adds vibrato.

5b · Direct comparison — how closely does the human reproduce the AI identity?

Both vocals sit on the same instrumental, so their time axes align. Values on a 10-ms grid.

MeasureValueReading
Time both sing72.9 sshared sung time
Only Min (AI) sings14.8 sthe human is silent here
Only Min – Human Avatar sings37.6 sthe human sings into Min's silences
Same note (±50 cents) in shared time56.6 %within one semitone: 68.0 %
Median pitch offset (Human − AI)-0.06 semitonesno transposition — same key, same register
Melody contour correlationr = 0.609lowest window 2:20–2:30 (r ≈ −0.04), highest at the end (r ≈ 0.92)
Phrase onsets matched27median offset 0 ms, typical deviation 40 ms — he starts exactly where Min starts
Median phrase length1.46 s · 3.23 sAI · Human — the human sings phrases twice as long
Longest phrase5.64 s · 11.98 sAI · Human
Spectral bandwidth2891 Hz · 3011 HzAI · Human
Timbre distance (MFCC 1–12)51.3clearly different voice colour despite same notes

Finding: the human copies the AI's timing almost perfectly (phrase starts within ~40 ms) and stays in the same register. He departs in two places: he fills Min's silences (37.6 s sung where Min is silent, against 14.8 s the other way) and he binds Min's short phrases into long lines. The identity he reproduces is the melody and the timing; the identity he does not reproduce is the silence.

Pitch contour overlay
Voicing and contour agreement
Pitch distribution
Performance metrics

6 · What this case shows

7 · What this case does not claim

8 · Evidence

ItemSource
Cyanite · A · MinContest-MP3 320k_Sound Chronicles Theme.mp3 · Cyanite library · 08.10.2026
Cyanite · B · Min – Human AvatarHuman vocal version, full mix · genre, mood, energy, instruments, voice · 06.11.2026
Cyanite · C · InstrumentalInstrumental (control), as sent to the singer · mood, energy, instruments, voice · 06.11.2026
KlangioVocal stem Min (AI) · vocal stem Min – Human Avatar · MusicXML · 06.11.2026
Pitch analysisSeparated vocal stems of A and B · 01.11. and 06.11.2026
All valuesCSV: AIMC Fallstudie SC Theme Messwerte.csv (66 rows, source per value)
Listening comparison6 passages, Min (AI) first, then Min – Human Avatar: 0:13 · 0:20 · 1:51 · 2:00 · 2:45 · 2:51

Screenshots and files available to researchers on request.