Talking Head Audio Cleaner

Shorts audio

by timscheuerai · MIT license · ★ 37 on GitHub

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…

---
name: shorts-audio
description: 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.
---

# shorts-audio

Stage 2 of shorts-edit. **Two traps, both silent, both cost hours.**

Quiet recordings cause both. The reference takes sat at **−43 LUFS** and needed **+26 dB**.

---

## TRAP A: `loudnorm` silently pumps

`loudnorm` with `linear=true` promises a single constant gain.

**It abandons that promise without warning** whenever the required gain would push the true

Overview

Stage 2 of shorts-edit. Two traps, both silent, both cost hours.

Manual install (npx, ZIP)

Claude Codenpx skills add timscheuerai/content-vault --skill shorts-audio -g -a claude-code

Claude Code (degit)npx degit timscheuerai/content-vault/skills/shorts-audio#main ~/.claude/skills/shorts-audio

Claude.ai / desktopZip the folder and upload it under Settings → Capabilities → Skills.

Codex, Cursor, other agentsPoint them at SKILL.md. The other files are plain text.

Connects with Agent Alley MCP · agentalley.io/mcp

Files included

SKILL.md4.1 KB

Read SKILL.md here

timscheuerai/main1 file shown

SKILL.md4.1 KB

Browse every file on GitHub↗

SKILL.md

Show the full text93 lines

shorts-audio/SKILL.md93 lines · 4.1 KB

PreviewCode

Outline

RawView on GitHub

shorts-audio

Stage 2 of shorts-edit. Two traps, both silent, both cost hours.

Quiet recordings cause both. The reference takes sat at −43 LUFS and needed +26 dB.

TRAP A: `loudnorm` silently pumps

loudnorm with linear=true promises a single constant gain.

It abandons that promise without warning whenever the required gain would push the true peak past the ceiling — which +26 dB always does. It falls back to dynamic mode: per-frame gain riding. Pumping, breathing, a squashed unnatural tone.

This — not the denoiser — was the actual cause of "the audio sounds artificial". Removing the denoiser alone did not fix it, which is exactly why it cost so much time.

Fix: never use loudnorm for gain. Measure the true peak, apply ONE constant volume=XdB that lands it at −1.5 dBTP, and accept whatever LUFS falls out.

Verified pure: correlation 0.999975 with the source waveform, residual 43 dB down (just AAC noise), no limiter ever engaged. Mathematically, the output is the input times a number.

TRAP B: every denoiser reconstructs the voice, and you can hear it

afftdn with tn=1 (blind tracking) guesses the noise spectrum moment to moment. Guessing wrong is precisely what makes musical burbling.

Learning the profile from real silence in the take is far better — asendcmd … sn start/stop around the longest silence:

chain	noise reduction	speech level
highpass 60Hz only	11.2 dB	−0.1 dB
RNNoise (arnndn)	17.6 dB	−0.1 dB
learned print, nr=12	23.0 dB	−0.1 dB

And it still didn't ship. On headphones, every spectral denoiser — blind, learned, even a 50/50 wet/dry blend — was audible as an unnatural tone, because they all reconstruct the signal rather than merely attenuating it. The user listened to all of them and chose the honest one.

SHIPPED: no denoiser. highpass=f=60 + a constant gain. Nothing else.

If a genuinely noisy take needs help, print12@50 (a 50% blend) was the best compromise.

Two ordering rules that silently defeat the filters

The denoiser must run BEFORE the trims. Its asendcmd timestamps point at silence in the ORIGINAL take, and the edit throws that silence away. Applied after the cut, the sample window falls outside the clip, the command never fires, and afftdn quietly reverts to blind mode — the exact thing that sounded fake. The floor betrays it: −54 dB (blind) vs −64.8 dB (learned).
Gain must precede ANY denoiser. RNNoise is trained on speech at normal levels. Fed −68 dBFS audio it sees silence and does literally nothing (measured: 0 dB of extra reduction). Gain it first and the same filter removes 23 dB.

What did NOT work

A noise gate. Intuitive — mute the gaps — and useless: after the edit the cut is 96% speech. There are no gaps left to gate. The audible noise is under the voice.
Notch filters. The noise looked like 60Hz mains harmonics at coarse FFT resolution. At fine resolution it isn't: 69% of the voice-band noise is broadband room tone. Notching mangled the voice (correlation 0.77) to buy 3 dB.

The mono-SFX clipping trap

amix negotiates a common channel layout. A mono SFX bed makes it pick mono, which silently downmixes the stereo voice — and that downmix pushed the peak from −1.62 dB to +0.83 dB (clipping). Pin aformat=channel_layouts=stereo on both sides of the mix.

Known tradeoff: the clips are quiet (−18 to −23 LUFS)

Peak-normalizing with no limiting and no compression lands below the −14/−16 social norm. That is the honest ceiling for zero dynamics processing; getting louder means compressing, which is what was rejected. Platforms normalize on playback.

The real fix is at the source: raise the input gain when recording.

1	---
2	name: shorts-audio
3	description: 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.
4	---
5
6	# shorts-audio
7
8	Stage 2 of shorts-edit. Two traps, both silent, both cost hours.
9
10	Quiet recordings cause both. The reference takes sat at −43 LUFS and needed +26 dB.
11
12	---
13
14	## TRAP A: `loudnorm` silently pumps
15
16	`loudnorm` with `linear=true` promises a single constant gain.
17
18	It abandons that promise without warning whenever the required gain would push the true
19	peak past the ceiling — which +26 dB always does. It falls back to dynamic mode: per-frame
20	gain riding. Pumping, breathing, a squashed unnatural tone.
21
22	This — not the denoiser — was the actual cause of "the audio sounds artificial".
23	Removing the denoiser alone did not fix it, which is exactly why it cost so much time.
24
25	Fix: never use `loudnorm` for gain. Measure the true peak, apply ONE constant
26	`volume=XdB` that lands it at −1.5 dBTP, and accept whatever LUFS falls out.
27
28	Verified pure: correlation 0.999975 with the source waveform, residual 43 dB down (just
29	AAC noise), no limiter ever engaged. Mathematically, the output is the input times a number.
30
31	---
32
33	## TRAP B: every denoiser reconstructs the voice, and you can hear it
34
35	`afftdn` with `tn=1` (blind tracking) guesses the noise spectrum moment to moment.
36	Guessing wrong is precisely what makes musical burbling.
37
38	Learning the profile from real silence in the take is far better — `asendcmd … sn
39	start/stop` around the longest silence:
40
41	\| chain \| noise reduction \| speech level \|
42	\|---\|---\|---\|
43	\| highpass 60Hz only \| 11.2 dB \| −0.1 dB \|
44	\| RNNoise (arnndn) \| 17.6 dB \| −0.1 dB \|
45	\| learned print, nr=12 \| 23.0 dB \| −0.1 dB \|
46
47	And it still didn't ship. On headphones, every spectral denoiser — blind, learned, even
48	a 50/50 wet/dry blend — was audible as an unnatural tone, because they all reconstruct the
49	signal rather than merely attenuating it. The user listened to all of them and chose the
50	honest one.
51
52	> SHIPPED: no denoiser. `highpass=f=60` + a constant gain. Nothing else.
53
54	If a genuinely noisy take needs help, `print12@50` (a 50% blend) was the best compromise.
55
56	---
57
58	## Two ordering rules that silently defeat the filters
59
60	1. The denoiser must run BEFORE the trims. Its `asendcmd` timestamps point at silence
61	in the ORIGINAL take, and the edit throws that silence away. Applied after the cut, the
62	sample window falls outside the clip, the command never fires, and afftdn **quietly
63	reverts to blind mode** — the exact thing that sounded fake. The floor betrays it:
64	−54 dB (blind) vs −64.8 dB (learned).
65
66	2. Gain must precede ANY denoiser. RNNoise is trained on speech at normal levels. Fed
67	−68 dBFS audio it sees silence and does literally nothing (measured: 0 dB of extra
68	reduction). Gain it first and the same filter removes 23 dB.
69
70	---
71
72	## What did NOT work
73
74	- A noise gate. Intuitive — mute the gaps — and useless: after the edit the cut is
75	96% speech. There are no gaps left to gate. The audible noise is under the voice.
76	- Notch filters. The noise looked like 60Hz mains harmonics at coarse FFT resolution.
77	At fine resolution it isn't: 69% of the voice-band noise is broadband room tone.
78	Notching mangled the voice (correlation 0.77) to buy 3 dB.
79
80	## The mono-SFX clipping trap
81
82	`amix` negotiates a common channel layout. A mono SFX bed makes it pick mono, which
83	silently downmixes the stereo voice — and that downmix pushed the peak from −1.62 dB to
84	+0.83 dB (clipping). Pin `aformat=channel_layouts=stereo` on both sides of the mix.
85
86	## Known tradeoff: the clips are quiet (−18 to −23 LUFS)
87
88	Peak-normalizing with no limiting and no compression lands below the −14/−16 social
89	norm. That is the honest ceiling for zero dynamics processing; getting louder means
90	compressing, which is what was rejected. Platforms normalize on playback.
91
92	The real fix is at the source: raise the input gain when recording.
93