# Porcelain, the song — Mila Zima

> Listen to Porcelain, the first song of the AI pop star Mila Zima: its 13 parts, its loudness, its sound colour and the range of the voice as measured, and how the master was made.

Page: https://milazima.com/porcelain/song/ · Site index for machines: https://milazima.com/llms.txt

---

![Mila Zima in a vocal booth with headphones on, at the microphone](https://milazima.com/media/cover/03-booth-1800.webp)

Audio file: https://milazima.com/media/porcelain/song-2/porcelain.mp3

Song 01 · November 2026

**The song · 3:58** · 107.7 BPM · −14.0 LUFS · 48 kHz · stereo · 13 parts · she sings in 11 of them · measured from the master · not verified by ear

The whole song, part by part. Height is loudness; rose is her voice. Click a part to play from there.

01 · The song

## Porcelain, as sound.

One recording, 3:58 long. It is the recording the film was cut to.

Three minutes fifty-eight, in thirteen parts. She sings in eleven of them, and in two of those there are no words.

This page is the song as sound: where it is loud, where the voice is alone, where the colour changes, how high the voice goes, and how the master was made. Every number was read from the files. None of it was checked by ear by the agent that measured it; it cannot hear.

[The film](https://milazima.com/porcelain/) · How it was mastered

- **Length:** 3:58 · 238.4 s
- **Tempo:** 107.7 BPM · one bar 2.23 s
- **Key:** F minor by measurement · not verified by ear
- **Loudness:** −14.0 LUFS · range 4.1 LU
- **True peak:** −1.0 dBTP
- **Master:** 48 kHz · stereo · 24 and 16 bit
- **The tune:** mostly between C4 and E♭5
- **Her voice:** heard for 198 s · absent for 40 s
- **Voice:** Mila's own, generated · Suno
- **Style:** soul-pop with a rock edge

02 · The parts

## Thirteen parts.

Click a part to play from there. There are no lyrics on this page; each part is described by what was measured in it.

1. **0:00**
   **Intro**
   The band alone. No voice and no humming before the first line.
   19 s · band only
2. **0:19**
   **Verse 1**
   Sung low: most of it sits around C4, the lowest of any part. One word in its last line is not clear in her diction.
   36 s · 61 words
3. **0:57**
   **Pre-chorus 1**
   The voice moves up to around A♭4 and the band gets louder under it.
   16 s · 28 words
4. **1:16**
   **Chorus 1**
   The hook. The tune sits around C5, an octave above the first verse, with the band at its loudest so far.
   20 s · 23 words
5. **1:36**
   **The rise**
   No words. One long note, then a high melody over the full band. The strongest line in the voice is around C6 here, with a much weaker one an octave under it.
   16 s · no words
6. **1:53**
   **Verse 2**
   Shorter than the first verse, and higher: around E♭4.
   14 s · 34 words
7. **2:10**
   **Pre-chorus 2**
   The same climb as before. Twice, at 2:12 and at 2:16, the band stops for a moment and comes back.
   16 s · 29 words
8. **2:28**
   **Chorus 2**
   The hook again, with the full band.
   20 s · 26 words
9. **2:48**
   **Bridge**
   The band drops out four times and comes back each time. The voice goes on through the gaps.
   15 s · 33 words
10. **3:04**
    **Last chorus**
    From 3:06 to 3:14 the band is almost gone and the voice carries the hook. The band is back at 3:15.
    20 s · 27 words
11. **3:24**
    **The peak**
    No words. The loudest seconds of the song, and the voice at its loudest.
    7 s · no words
12. **3:31**
    **Closing**
    The last words, with one of them held for three seconds. The voice stops at 3:42.
    11 s · 12 words
13. **3:42**
    **Outro**
    The band alone, fading out.
    16 s · band only

03 · Sound colour

## Where the sound sits.

Five readings on one time line, from the first second to the last. Click anywhere in a chart to play from there. Measured from the files; not verified by ear.

*The whole song · 40 Hz to 16 kHz, logarithmic · brighter is louder*

![Spectrogram of the whole song Porcelain, from 40 Hz to 16 kHz](https://milazima.com/media/porcelain/song-2/full.webp)

100 Hz · 1 kHz · 10 kHz ·

Spectrogram of the master. The display is tilted 3 dB per octave so that the top is visible; the range from dark to bright is 62 dB.

*Loudness · short-term, 3 smomentary, 0.4 s*

−14 LUFS · −20 LUFS · −26 LUFS ·

EBU R 128. The whole song reads −14.0 LUFS with a loudness range of 4.1 LU. The loudest three seconds are at 3:25 (−12.2 LUFS).

*Voice and band · lead vocaleverything else*

−20 dB · −30 dB · −40 dB ·

Level of the take's lead-vocal channel and of the rest, half a second at a time, before mastering. The rest includes the backing voices. Where the upper line falls away she is alone.

*Colour · LowLow midMidPresenceAir*

50 % ·

From the bottom up: low (20 Hz to 120 Hz), low mid (120 Hz to 500 Hz), mid (500 Hz to 2 kHz), presence (2 kHz to 6 kHz), air (6 kHz to 20 kHz). Over the whole song: low 25 %, low mid 21 %, mid 39 %, presence 11 %, air 5 %. The energy-weighted centre is 1310 Hz.

*The voice alone · 100 Hz to 10 kHz, logarithmic*

![Spectrogram of the lead vocal of Porcelain, from 100 Hz to 10 kHz](https://milazima.com/media/porcelain/song-2/voice.webp)

1 kHz ·

Spectrogram of the lead-vocal channel. The lowest bright line is the note she sings; the lines above it are its overtones, and their weight is the colour of the voice.

*Average spectrum · verseschorusesbridge*

0 dB · −10 dB · −20 dB

Third-octave bands from 50 Hz to 16 kHz, relative to the loudest band of the whole song.

*Where the voice sits · seconds on each note*

5 s · 10 s · 15 s

pYIN on the lead-vocal channel, 158 s of pitched voice. Most of the singing lies between C4 and E♭5; the note held longest in total is A♭4. The channel also carries 22 s at G5 and above, up to C6, in the choruses and in the rise (the rose bars). In 11 of those seconds there is a second, much weaker line an octave lower, about 29 dB down. One voice with a very strong overtone, or two voices an octave apart: the measurement cannot tell.

Part by part

| Part | Starts | Length | Words | Voice | Band | Loudness | Centre | Notes: low · most · high |
|---|---|---|---|---|---|---|---|---|
| Intro | 0:00 | 19 s |  |  | −23.5 dB | −13.4 LUFS | 1178 Hz |  |
| Verse 1 | 0:19 | 36 s | 61 | −23.5 dB | −25.6 dB | −14.4 LUFS | 1241 Hz | A♭3 · C4 · B♭4 |
| Pre-chorus 1 | 0:57 | 16 s | 28 | −22.8 dB | −24.0 dB | −13.6 LUFS | 1435 Hz | E♭4 · A♭4 · B♭4 |
| Chorus 1 | 1:16 | 20 s | 23 | −22.5 dB | −22.9 dB | −13.6 LUFS | 1575 Hz | B3 · C5 · C6 |
| The rise | 1:36 | 16 s | none | −21.1 dB | −22.9 dB | −13.9 LUFS | 1211 Hz | F4 · C6 · D♭6 |
| Verse 2 | 1:53 | 14 s | 34 | −22.6 dB | −24.9 dB | −15.0 LUFS | 1488 Hz | B♭3 · E♭4 · A♭4 |
| Pre-chorus 2 | 2:10 | 16 s | 29 | −22.8 dB | −26.3 dB | −14.5 LUFS | 1510 Hz | E♭4 · A♭4 · A♭5 |
| Chorus 2 | 2:28 | 20 s | 26 | −22.3 dB | −23.9 dB | −13.4 LUFS | 1419 Hz | B♭3 · C5 · C6 |
| Bridge | 2:48 | 15 s | 33 | −22.1 dB | −25.4 dB | −14.4 LUFS | 1417 Hz | F4 · A♭4 · D♭5 |
| Last chorus | 3:04 | 20 s | 27 | −21.2 dB | −24.4 dB | −13.6 LUFS | 1477 Hz | G4 · C5 · C6 |
| The peak | 3:24 | 7 s | none | −16.8 dB | −22.7 dB | −12.4 LUFS | 863 Hz | E♭5 · F5 · A♭5 |
| Closing | 3:31 | 11 s | 12 | −19.3 dB | −22.4 dB | −12.4 LUFS | 1287 Hz | A♭4 · B♭4 · E♭5 |
| Outro | 3:42 | 16 s |  |  | −33.0 dB | −21.9 LUFS | 475 Hz |  |

Voice and band: level of each channel before mastering, dB below full scale. Loudness: median short-term LUFS of the master. Centre: energy-weighted centre of the spectrum. Notes: 5th percentile, the note held longest, 95th percentile.

04 · How the sound was made

## One take, mastered twice.

Her voice is generated. The choices were made by a person, by ear; the measuring was done afterwards.

1. **01**
   **The voice**
   One generated voice, made once in Suno and kept. She sings every song with it.
2. **02**
   **The take**
   The song was generated more than once. The director listened to the takes and chose this one by ear.
3. **03**
   **The check**
   The voice of that take was separated from the band and every word was timed. The words she sings were compared with the words as written. One word is not clear in her diction.
4. **04**
   **The master**
   Made from the full mix of the take at 48 kHz, not from summed stems. The voice is separated, raised 2.5 dB and put back. Then a high shelf of +1.2 dB at 10 kHz and a bell of +0.8 dB at 3.5 kHz, Q 1.2. Then loudness normalisation in two passes, to −14.0 LUFS with the true peak at −1.0 dBTP.
5. **05**
   **Twice**
   The first master, on 13 November, started from a 44.1 kHz file. The second, a day later, started from the take's own 48 kHz mix, with the same chain and the same targets. The second one plays on this page.
6. **06**
   **The files**
   A 24-bit WAV and a FLAC at 48 kHz, a 16-bit WAV, and an MP3 at 320 kbit/s. This page plays the MP3.
7. **07**
   **The voice alone**
   Two separations of the voice exist. One is made locally with Demucs and is part of the master chain. The voice charts on this page are read from the other: the take's own lead-vocal channel, as the generator gives it.

What was measured with what.

Loudness: ffmpeg ebur128. Spectrum, tempo, key estimate and pitch: librosa (STFT, beat tracker, chroma, pYIN). Separation: Demucs. Word times: faster-whisper. The charts are drawn by the page generator from those numbers.

- **Loudness:** −14.0 LUFS integrated · range 4.1 LU
- **True peak:** −1.0 dBTP · sample peak −1.0 dBFS
- **Peak to RMS:** 16.3 dB
- **Stereo:** left–right correlation 0.83 · side 10.3 dB below mid
- **Format:** 48 kHz · stereo · 24 and 16 bit

05 · What is not known

## Nine open points.

The song is finished. These are the things on this page that nobody has confirmed.

1. **01** · The key. Three estimates do not agree. From the master: F minor (0.79), F major (0.68), C minor (0.64). From the band alone: F minor (0.77), F major (0.71), C major (0.56). From the notes the voice spends its time on: C minor (0.82), A♭ major (0.69), F minor (0.63). The style brief asked for F sharp minor. Nobody has checked it at an instrument.
2. **02** · The pitch readings come from an algorithm, not from a musician's transcription. Of the 22 seconds it places at G5 and above, 11 also show a weaker line an octave lower. In those seconds the sung note may be the lower one.
3. **03** · One word at the end of the first verse, around 0:51 to 0:55, is not clear in her diction. A transcription model hears two other words there.
4. **04** · Voice and band levels are read from the take's channels before mastering. The master raises the voice by 2.5 dB, so in what you hear the voice sits a little higher against the band than these two lines show.
5. **05** · The word counts come from a transcription of the first master. In the parts without words it sometimes writes words that are not there; those parts are counted by hand as wordless.
6. **06** · The tempo is one number from a beat tracker for the whole song. It was not checked against a click.
7. **07** · The film was cut to the first master, at 44.1 kHz. This page measures the second. Same take, same chain; the two files are not identical.
8. **08** · This page has no lyrics, no score and no chords.
9. **09** · The agent that measured all of this cannot hear. It reads levels, spectra and pitch curves. Whether the song is good was decided by a person.

The same song in 56 shots

[The film](https://milazima.com/porcelain/)

Pause

Intro · 0:00 / 3:58

---

Mila Zima is an AI pop star. Her face, her voice and the rooms she sings in are generated. The decisions are made by people.
Contact: mila@milazima.com · https://www.instagram.com/milazimacom/ · https://www.tiktok.com/@milazimacom · https://x.com/milazimacom · https://www.youtube.com/@milazimacom
