Przejdź do treści
Biblioteka skilli
Marketing i treści · Przegląd i audyt

audio-verification

/audio-verification

Sprawdza wyeksportowany dźwięk z FL Studio pomiarami, wykrywając szumy, kliki, głośność i niespójne przejścia.

Licencja
MIT
Ostatnia zmiana repozytorium
08.11.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 08.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Sprawdza wyeksportowany dźwięk z FL Studio pomiarami, wykrywając szumy, kliki, głośność i niespójne przejścia.

Oryginalny opis (EN)

Check exported audio with measurements instead of guessing, since the agent can't listen. Use after rendering a mix or stem from FL Studio, when the user reports noise, clicks, thumps, harshness or low volume, to confirm that chords sound the notes that were written, to prepare level-matched versions for an A/B comparison, or to audit the continuity of the transitions between sections.

Jak działa

  1. Commands
  2. From a complaint to a measurement
  3. What the numbers looked like in practice
  4. Check notes before effects
  5. Comparing versions
  6. Transition Continuity Audit
  7. Limits
Sekcje z nagłówków SKILL.md, w oryginale · 116 linii, 1 337 słów
Podgląd SKILL.md
---
name: audio-verification
description: Check exported audio with measurements instead of guessing, since the agent can't listen. Use after rendering a mix or stem from FL Studio, when the user reports noise, clicks, thumps, harshness or low volume, to confirm that chords sound the notes that were written, to prepare level-matched versions for an A/B comparison, or to audit the continuity of the transitions between sections.
---

# Verifying exported audio

The agent can't hear the result. `bin/wav-check` turns a rendered WAV into numbers, which is enough to find
many concrete problems and never enough to say that something sounds good. Report what was measured and
leave taste to the person listening.

Peak meters read during playback only prove that a track has signal. Render a file and measure that.

## Commands

```sh
wav-check measure a.wav b.wav                 # peak, RMS, loudest passage, energy per band
wav-check timeline a.wav --step 4             # peak per interval
wav-check noise a.wav --start 24              # floor level, bands and slow level swings
wav-check thumps a.wav --bpm 95               # low attacks with almost no midrange
wav-check chords a.wav --bpm 95 'Bm:0.3-3.3:B,F#,C#'
wav-check chords a.wav --bpm 95 --skip 1 'Gm:28-32:G,Bb,D'
wav-check width a.wav                         # left/right correlation and side level
wav-check match out/ --peak -8.8 a.wav b.wav  # same level for A/B listening
wav-check cut out/ --bpm 95 --end 32 a.wav b.wav  # the same passage, levels untouched
```

Installed as a plugin, `wav-check` is on `PATH`. In a clone it is `bin/wav-check`, two folders above this file. It
needs numpy, so run `uv run --no-project --with numpy python <path to wav-check>` if numpy isn't installed.

## From a complaint to a measurement

| The user says | Run | Look for |
|---|---|---|
| Too quiet | `measure` | The peak. Then read the master and track faders before touching anything else |
| Harsh, metallic | `measure` | The share of 2-6 kHz, compared with an earlier version |
| Dull | `measure` | The same band falling too far, 120-500 Hz taking over |
| Hiss or a wave in the background | `timeline`, then `noise` | A floor that doesn't decay and a swing below a few Hz |
| Thumps or clicks in the middle | `thumps` | The beat, then the note at that beat |
| The render is much longer than the song | `timeline` | A constant floor keeping FL's tail detection open |
| Wrong notes, odd harmony | `chords` on the dry track | Missing pitch classes |

## What the numbers looked like in practice

These come from one clean electric guitar track at 95 BPM, and they are examples, not targets.

- **Background wave.** `timeline` showed a constant floor near -62 dBFS after the music ended, and the render
  ran 50 s for 20 s of song. `noise` put 94% of the energy above 2 kHz with a swing at 0.8 Hz. The cause was the
  `Noise` knob of Vintage Chorus, on by default. It emulates analog hiss, and the chorus modulates it, which is
  why it sounded like a wave. Turning it off dropped the hiss by 19 dB in quiet passages and cut the render to 28 s.
- **Thump.** `thumps` flagged beat 19.55, 16.8 dB above the midrange, and the note there was a low palm mute
  at velocity 0.41. Palm mute depth grows as velocity falls in Ample Guitar, so low velocities on a low string
  can read as a hit with no pitch.
- **Too bright.** Energy at 2-6 kHz went from 4% to 14% with the amp's bright switch and treble up, and the
  listener called it metallic. A setting near 8% was the middle ground.

## Check notes before effects

Run `chords` on the track with no distortion, delay or reverb. Each effect moves the reading.

- Distortion adds partials that were never played. A G chord with no third read B as a strong note after overdrive.
- A long reverb tail carries the previous chord, so a pitch class from the last chord can appear as `extra`.

A `DIFFERS` on a wet track is not a wrong note until the dry track says so. A passage should start after the attack
and end before the next chord.

A synth pad has the same problem with no effect on its mixer track, since a patch can carry its own release and
effects, so it can read the last chord for a while. `--skip` moves the start of every passage by that many beats, and the printed range shows what was read.
On a 4 beat chord at 95 BPM, two pads read `DIFFERS` over the whole chord and `ok` from the second beat, so the notes
were right and the change was slow.

## Comparing versions

Louder usually reads as better, so match levels before asking anyone to choose. Use `--peak` for whole takes
and `--rms` for short excerpts, where one hit would set the whole gain. Keep the original files, and say which
copies had their gain changed.

Use `cut` instead when the level difference is the thing being compared, such as a passage with and without a new
layer. It keeps each file's level. End the cut before the next section starts, because a fade over the entry of a
band leaves that entry in the file. Measure the same range as well, since a window that runs into the next section
counts its level.

Change one thing per render.

## Transition Continuity Audit

When a passage between sections sounds broken, read the notes of every channel on both sides of the boundary
(`fl-notes read`, or the score a generator wrote) before touching anything. Run a light pass over the main
boundaries before the final export of a piece with several sections too, when the notes and the section boundaries
are available. It doesn't need repeating after every small edit. Check four things in a window of a bar or two around
each boundary:

1. Gaps. A part ends before the part that takes over starts, or one line stops and resumes with a hole in between.
2. Stray notes. Short notes left at the start of a section by a range that ran one chord too far.
3. Rearticulations. The same pitch ends and starts again on one channel, which a legato instrument plays as a new
   attack. Overlapping notes of different pitches are a legato transition, not a new attack.
4. Abrupt controls. A large step between two consecutive CC1 or CC11 values, a CC64 release that cuts notes the pedal
   was holding, or note-offs on several channels at the same barline. A controller holds its value until the next
   event, so judge the size of each step, not the time between events. A ramp drawn in small steps isn't a jump, and
   a step while the channel is silent only prepares its next entry.

Compare sounding times, not written ones. Add the attack and legato delays measured for each instrument, and treat a
note under a held sustain pedal as lasting until the pedal lifts. In Cinematic Studio Strings the previous note kept
sounding through a legato transition until the new one was heard, so there the transition delay wasn't a gap. Check
that in a render before assuming it for another library. Say which delays and releases weren't measured, because a
long release can fill a gap that the notes show. To confirm that a gap is audible, render the stems and run
`timeline` with a short `--step` on each.

Report each finding with the channel, bar and beat, the events involved and why it looks suspicious. Don't fix
anything from the audit alone. A silence, a new bow stroke or a sudden contrast can be what the composer wanted, so
each finding is a question for the person who wrote it. Controls that a project draws as automation clips aren't in
the notes, so say when they couldn't be checked.

## Limits

- The tool knows nothing about taste, feel, timing or whether a part suits the song.
- `thumps` finds low attacks with little midrange. Its default threshold of 12 dB is calibrated on one guitar part,
  and it can miss hits in very short files, because the attack threshold comes from the file itself.
- `chords` compares pitch classes between 70 Hz and 1 kHz. It can't tell octaves or inversions apart.

Wersje

W bazie nie ma innych skilli o nazwie audio-verification.

Migawka bazy z 13.11.2026 · wersja paczki 82bcd3cf3486 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 100audio-review
    Ocenia dźwięk jak doświadczony projektant: głośność, dynamikę, balans, szumy i specyfikację, z punktacją i poprawkami.
  • 100sound-designer
    Odpowiada za dźwięk wideo: dialogi, podkład muzyczny, efekty zsynchronizowane z obrazem, miks i głośność na platformy.
  • 100audio-polish
    Normalizuje głośność wideo i audio dla YouTube, TikTok i podcastów oraz usuwa ciszę i kliki.
  • 100audio-and-sound
    Dodaje audio do gry w Phaser 4: wczytywanie dźwięku, odtwarzanie muzyki, głośność, dźwięk przestrzenny, Web Audio API i SoundManager.
  • 100audio-diagnostician
    Analizuje stemy audio przed miksem: katalogizuje problemy według wagi, wykrywa maskowanie częstotliwości między stemami i tworzy strukturalny brief miksu przez narzędzia Phantom MCP.
  • 100audio-stem-delivery
    Planuje nazewnictwo stemów, wyrównanie, zapas głośności i spójność eksportu.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.