Skip to content
Skills Library
Marketing & content · Automation

studio-voice-processing

/studio-voice-processing

Processing of a recorded human voice (the author's teleprompter takes, interviews, voice memos, a talking head's sound) — which module fixes what, in which order, with what settings, and how to prove it by measurement — voice_polish chain, trim_silence, declick, denoise, voice_eq, compress, deess, reduce_breaths, voice_isolate, normalize. Use when a recorded voice sounds noisy, quiet, uneven, harsh, boomy or clicky, or before mixing it with music.

Repository
smashcut ↗
Licence
MIT
Last repository change
12.11.2026

Assessment and security

100/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content35 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
1
Scripts
0
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 12.11.2026, without references to retired models or APIs.

Licence

MIT, open: use and modify while retaining author attribution.

What it does and when to use it

Processing of a recorded human voice (the author's teleprompter takes, interviews, voice memos, a talking head's sound) — which module fixes what, in which order, with what settings, and how to prove it by measurement — voice_polish chain, trim_silence, declick, denoise, voice_eq, compress, deess, reduce_breaths, voice_isolate, normalize. Use when a recorded voice sounds noisy, quiet, uneven, harsh, boomy or clicky, or before mixing it with music.

Original author description

Processing of a recorded human voice (the author's teleprompter takes, interviews, voice memos, a talking head's sound) — which module fixes what, in which order, with what settings, and how to prove it by measurement — voice_polish chain, trim_silence, declick, denoise, voice_eq, compress, deess, reduce_breaths, voice_isolate, normalize. Use when a recorded voice sounds noisy, quiet, uneven, harsh, boomy or clicky, or before mixing it with music.

How it works

  1. Start here: voicepolish
  2. The order, and why
  3. Modules
  4. Recipes
  5. Don't
Original SKILL.md headings · 93 lines, 1,137 words
SKILL.md preview
---
name: studio-voice-processing
description: Processing of a recorded human voice (the author's teleprompter takes, interviews, voice memos, a talking head's sound) — which module fixes what, in which order, with what settings, and how to prove it by measurement — voice_polish chain, trim_silence, declick, denoise, voice_eq, compress, deess, reduce_breaths, voice_isolate, normalize. Use when a recorded voice sounds noisy, quiet, uneven, harsh, boomy or clicky, or before mixing it with music.
---

# Processing a recorded voice

You cannot hear. Work from measurements: every step below returns numbers or can be measured (`voice_polish` report, `normalize`/`compress` loudness, `watch_video` / `analyze_audio` for levels and silences). Never say "sounds cleaner" — say what changed in dB.

Synthetic voice (ElevenLabs) needs none of this: it is already clean and levelled. This skill is for sound that came through a microphone.

## Start here: `voice_polish`

`run_media_tool {op:'voice_polish', params:{media}}` — or, for a teleprompter take, `process_voice_take {takeId, options}` (always starts from the raw recording and swaps the result on the timeline).

It runs the whole chain in the right order and returns a report:

```
before / after: lufs · truePeak · lra · noiseFloor · snr · duration
trimmed {start, end} · steps[] · mode · noiseMeasured
```

Read it and decide:

| Report says | Do |
|---|---|
| `after.lufs` within ±0.5 of the target, `truePeak` ≤ −1 | Level is right. |
| `after.snr` ≥ 35 dB | Clean. Stop. |
| `after.snr` 25–35 dB | Fine under music. For voice-only video try `noise:'strong'`. |
| `after.snr` < 25 dB with `noise:'strong'` | The room is the problem: tell the user (closer to the mic, quieter room) or, with their consent, `noise:'isolate'` (cloud, paid). |
| `before.snr` ≥ 40 dB | Already clean: `noise:'light'` or `'off'` — a denoiser on a clean voice only costs detail. |
| `before.lra` ≤ 4 LU | Already even (or already compressed): `compress:'light'` or `'off'`. |
| `before.lra` ≥ 10 LU | Whispers and shouts: `compress:'strong'`. |
| `mode: 'dynamic'` | The peak ceiling could not be met with plain gain, loudnorm rode the level. Fine for a voice; if it pumps, raise `compress` one step. |
| `noiseMeasured: false` | The take had no pause ≥ 0.7 s, the noise figures are a rough estimate — do not quote them as measured. |
| `trimmed.start` > 2 s | Long lead-in was cut; if captions were timed on the raw file, they shift by `trimmed.start`. |

Targets: **−16 LUFS** for a voice that sits over music in a −14 LUFS video (default), **−14** for voice-only videos, true peak ≤ **−1.5 dBTP**.

## The order, and why

1. **Trim the edges** (`trim_silence`) — the click of the record button and the breath before the first word.
2. **High-pass 80 Hz** (`voice_eq {highpass}`) — rumble first, or it drives the compressor.
3. **Declick** (`declick`) — before anything that raises level.
4. **Noise** (`denoise`, or `voice_isolate`) — before compression: a compressor lifts whatever is quiet, noise included (measured: 3:1 compression costs ~5 dB of SNR).
5. **Breaths** (`reduce_breaths`) — optional, see below.
6. **EQ** (`voice_eq`) — tone.
7. **Compressor** (`compress`) — evenness.
8. **De-esser** (`deess`) — last of the tone tools: compression and the presence boost are what make sibilants stick out.
9. **Loudness** (`normalize`) — always the final step; nothing after it.

Running single modules by hand: keep this order, feed each result into the next, and `normalize` at the end. Don't run a module twice "for more" — change its setting.

## Modules

**`trim_silence {media, padding}`** — edges only; inner pauses stay (they are the reader's rhythm). For jump-cutting pauses out of a talking head use `remove_silence` instead — never on narration that has to match a scene's timing.

**`declick {media, strength}`** — mouth clicks, lip smacks. `normal` by default. `strong` only if clicks are still reported/heard: it starts to dull t/k/p.

**`denoise {media, strength, noiseFloor}`** — steady noise: hiss, fan, hum, room tone. It cannot remove a dog, a keyboard, traffic, reverb — that is `voice_isolate`. Stand-alone it is gentle; inside `voice_polish` the noise stage is the denoiser plus a soft expander in the pauses: `light` ≈ 8 dB, `normal` ≈ 12 dB, `strong` ≈ 20 dB (measured on a voice over room noise: SNR +10 dB at `normal`, +20 dB at `strong`). `strong` on a quiet voice can sound watery — it is for laptop mics and fans.

**`voice_isolate {media}`** — ElevenLabs Voice Isolator: voice out of a bad room, street, music. **Cloud and paid** (≈ 1000 characters of quota per minute; the audio leaves the machine). Only when the user asked or agreed, and say what it will cost first. Needs the ElevenLabs key. In the chain: `voice_polish {noise:'isolate'}`.

**`voice_eq {media, highpass, mud, presence, air}`** — dB, 0 = off, stay within ±4.
- boomy / boxy room → `mud: -3`
- dull laptop or phone mic → `presence: 3.5, air: 2`
- harsh, thin condenser → `presence: 0, air: 0`, then `deess`
- voice disappears under music → `presence: 3` (and `duck` the music) rather than more volume
- deep male voice with thump → `highpass: 100`

**`compress {media, amount, ceiling}`** — `light` 2:1 (already steady reader), `normal` 3:1 (default), `strong` 4.5:1 (ads, energetic delivery, wide LRA). Threshold follows the file's loudness. Returns loudness before/after: LRA should drop; if LRA was already < 4, skip.

**`deess {media, strength, split}`** — only the band above `split` is turned down, and only on sibilants. `normal`; `strong` risks a lisp. `split` 4500 for a low voice, 6500 for a high one.

**`reduce_breaths {media, reduction, sensitivity}`** — an expander, **not** a breath detector: it lowers everything quiet, so soft word endings can go too. Off by default. Use when the user asks or breaths are obviously loud after compression; `reduction` ≤ 12 (removed breaths sound unnatural — lowered ones don't). If words lose their tails, raise `sensitivity`.

**`normalize {media, target}`** — the last step of any hand-made chain.

## Recipes

- **Teleprompter take, decent mic** — defaults: `process_voice_take {takeId}`.
- **Laptop mic, fan in the room** — `{noise:'strong'}`; if `after.snr` is still < 25 → talk to the user about `isolate`.
- **Good studio mic, quiet room** (`before.snr` ≥ 40) — `{noise:'light', declick:true, compress:'light'}`.
- **Energetic ad read** — `{compress:'strong', target:-14}`.
- **User says "too processed / robotic"** — `{noise:'light', compress:'light', deess:false}`; offer `options:false` (raw + nothing) as the reference.
- **User says "I hear breathing"** — `{breaths:true}`; mention that endings may soften.
- **Voice over music** — polish to −16 LUFS, then `duck {music, voice}` (studio-media-toolkit), then check the mix (`music_report {vo:true}`).
- **Interview / talking-head video** — `extract_audio` → `voice_polish` → put the result under the picture (mute the clip's own sound), or `denoise` + `normalize` directly on the video when only light cleaning is needed.

## Don't

- Don't process a take twice (polish on a polished file): start again from the raw (`process_voice_take` does).
- Don't normalise before compressing, or compress after normalising.
- Don't use `remove_silence` on scene narration — it changes the timing the scene was cut to.
- Don't send audio to the cloud isolator without the user's yes.
- Don't report a result without the numbers: LUFS before → after, noise floor before → after, what was trimmed.

Versions

The database has no other skills named studio-voice-processing.

Database snapshot: 13.11.2026 · package version 5d504b2287a5 (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 100studio-author-voice
    The author voices the video themself — preparing the script for the teleprompter, what happens to a recorded take (processing, word timings, placement on the voice-over track), checking takes, re-timing scenes to the recording, captions from the take, mixing with music. Use when…
  • 90shorts-audio
    Clean the audio on a talking-head recording without making it sound processed — constant-gain loudness, rumble removal, and when (not) to denoise. Use when recording audio sounds weird, artificial, pumping, quiet, or noisy, or when asked to denoise or normalize a video's audio.
  • 100audio-mixing
    Mix voiceover, background music and sound effects for video - ducking music under speech, fades, loudness normalization (LUFS), noise reduction and music selection. Use when audio sounds uneven, music overpowers speech, or the user asks to add or balance music.
  • 100x-voice
    Measure the user's actual X voice from their own past posts and write voice.md from evidence, then check any draft for drift before it goes out. Use when the user says "write my voice.md", "learn how I write", "does this sound like me", "this doesn't sound like me", "make it…
  • 100li-voice
    Measure the user's actual LinkedIn voice from their own past posts and write voice.md from evidence, then check any draft for drift before it goes out. Use when the user says "write my voice.md", "learn how I write", "does this sound like me", "this doesn't sound like me", sets…
  • 90talking-head-and-piece-to-camera
    The on-camera delivery craft — helping a real human film themselves talking to a lens and look like themselves doing it. Use when someone wants a "talking head video" or "piece to camera," says "film myself" or "I look stiff on camera," asks about a teleprompter, framing,…

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.