録音された会話音声からリップノイズや口内クリック音を低減する方法

The Biomechanics & Acoustic Anatomy of Mouth Sounds

In professional audiobook narration, voiceover, and podcast production, mouth noise (often called "saliva clicks" or "lip smacks") is the most tedious defect to clean manually. In an 8-hour audiobook, a single narrator may produce over 3,000 audible clicks.

Understanding the acoustic physics of mouth sounds is the key to preventing and repairing them:

• Mechanical Origin: As the narrator opens their mouth or repositions their tongue to articulate vowel and consonant transitions, thin liquid bridges of saliva stretch and snap.
• Acoustic Profile: The snap creates an impulsive acoustic shockwave with an instantaneous rise time (< 0.2 ms) and a total duration of 1 ms to 4 ms.
• Spectral Distribution: Saliva clicks contain concentrated broadband energy spanning 2 kHz to 10 kHz, with peak energy typically centered between 3.5 kHz and 6.5 kHz.

Micro-Transients (1–4 ms) vs Speech Consonants (t, k, p)

The challenge in de-clicking is distinguishing unintentional saliva pops from intentional speech consonants (like "t", "k", "p", and "ch").

• Natural Consonants ("t", "k"): Have an acoustic duration of 20 ms to 80 ms and exhibit complex formant transitions and harmonic resonance following the initial burst.
• Saliva Clicks: Have an acoustic duration of only 1 ms to 4 ms and exhibit an isolated, razor-thin impulse spike with zero harmonic body.

Prevention vs Post-Production Repair Matrix

The most efficient workflow combines physical prevention at the microphone with automated transient interpolation in software:

The 4-Step De-Clicking Workflow: Detection to Spline Repair

Execute this complete prevention and restoration protocol for pristine spoken-word clarity:

ステップ 1：Pre-Session Vocal Conditioning (Hydration & Malic Acid)

Saliva clicks are exacerbated by dehydration, caffeine, and dairy products, which cause saliva to become thick and sticky.

• 30 Minutes Before Recording: Drink 16 oz (500 ml) of room-temperature water. Avoid ice water (which constricts vocal cords).
• 15 Minutes Before Recording: Eat 2 to 3 slices of a tart green apple (Granny Smith). The natural malic acid breaks down salivary proteins and cleanses the tongue surface.
• During Recording: Keep room-temperature water mixed with a splash of fresh lemon juice at your desk and take small sips every 10 minutes.

ステップ 2：Capsule Deflection (The 45-Degree High-Frequency Null)

High-frequency acoustic transients radiate in a directional beam directly forward from the lips.

• Position the microphone 5 to 7 inches from your mouth, but angle the capsule 30° to 45° off-axis (aimed at the corner of your mouth rather than straight on).
• Directional high-frequency saliva clicks shoot straight past the capsule, while vocal resonance enters the cardioid pattern naturally.

ステップ 3：Statistical Transient Detection (>4.5σ Running Variance)

In post-production, AutoSilenceCut's de-clicker analyzes raw audio samples in client memory using a sliding variance window.

• The algorithm calculates a running standard deviation ($\sigma$) of sample amplitudes across 10 ms windows.
• When an instantaneous amplitude spike exceeds $4.5\times$ the running standard deviation within a window shorter than 4 ms, the system flags the index as a saliva transient anomaly.
• Because speech consonants (like "t" or "s") sustain energy across 20+ milliseconds, they never trigger false-positive transient flags.

ステップ 4：Hermite / Catmull-Rom Cubic Spline Reconstruction

Once a 1 ms to 4 ms click window is isolated, the corrupted samples must be repaired without altering the timeline length or creating a gap:

• The algorithm takes the 2 clean audio samples immediately preceding the click ($P_0, P_1$) and the 2 clean samples immediately following it ($P_2, P_3$).
• Using Catmull-Rom cubic spline interpolation, the DSP engine calculates a smooth, continuous polynomial curve through the corrupted window: $$p(t) = 0.5 \cdot \left[ (2P_1) + (-P_0 + P_2)t + (2P_0 - 5P_1 + 4P_2 - P_3)t^2 + (-P_0 + 3P_1 - 3P_2 + P_3)t^3 \right]$$
• The transient saliva pop is replaced by a mathematically natural transition curve that is completely imperceptible to the human ear.

Production Scenario: Cleaning an 8-Hour Audiobook Narration

The Project: An ACX audiobook narrator delivers an 8-hour spoken narrative recorded on a Neumann U87 condenser microphone. Despite good hydration, close microphone proximity captured hundreds of subtle saliva clicks on pre-sentence inhales and syllable transitions.

Restoration Pipeline in AutoSilenceCut:

1. Batch Audio Ingest: Narrator imports the uncompressed 24-bit 48 kHz WAV files directly into the browser workspace.
2. De-Clicker Analysis: The cubic spline engine scans the tracks, identifying 1,420 micro-transient spikes exceeding the 4.5σ threshold.
3. Micro-Window Spline Repair: All 1,420 click instances are repaired in client memory via Catmull-Rom interpolation within 8 seconds.
4. Natural Pacing Cut: Long breath pauses (>1.2s) between paragraphs are shortened to 0.32s speech cushions with 8 ms micro-crossfades.
5. ACX Compliance Check: Exported audio meets ACX standard specifications: -18 to -23 LUFS RMS with -3.0 dBFS true-peak headroom and zero audible mouth noise.

4 Mistakes That Muffle Audio When Trying to Remove Clicks

• Mistake 1: Scooping Out 3 kHz to 5 kHz with EQ. Trying to notch out mouth clicks with a static equalizer scoops out critical speech presence, making the narrator sound muffled and distant.
• Mistake 2: Applying a Heavy Noise Gate. Noise gates do not open and close during words. When a mouth click occurs inside a word (e.g. "par- [click] -ticular"), a gate does nothing.
• Mistake 3: Manually Slicing and Deleting Clicks. Slicing out 10 milliseconds of audio creates a jump in time and shifts video sync. Always use spline interpolation to keep timeline timecodes intact.
• Mistake 4: Drinking Dairy or Sugary Drinks Before Sessions. Milk, yogurt, sugary energy drinks, and chocolate coat the mouth in thick sugars that dramatically increase saliva snap sounds.