音声の背景ノイズを除去する実践的エンジニアリングワークフロー

Noise Diagnostics: Identifying What Is in Your Background

The most critical mistake in spoken-word editing is treating all unwanted sound as a single generic problem called "noise." In physical acoustics, background noise consists of fundamentally different energy profiles that require completely different mathematical DSP operations.

Before touching any software controls, listen to 3 to 5 seconds of the recording's quiet room tone on closed-back headphones and diagnose whether the sound is stationary or transient.

Stationary Noise vs Transient Acoustic Interference

• Stationary (Continuous) Noise: Acoustic energy that maintains a predictable spectral distribution and amplitude over time. Examples include computer cooling fans, air conditioning airflow, refrigerator compressors, and 60 Hz AC electrical hum. These can be attenuated cleanly with linear filters and downward expanders.
• Non-Stationary (Transient / Dynamic) Noise: Acoustic events that occur unpredictably with rapid attack envelopes and wide frequency distributions. Examples include keyboard clicks, chair creaks, car horns, dogs barking, and door slams. These cannot be removed by static filters without devastating the speaker's vocal harmonics during overlap.

The Audio Restoration Decision Matrix

Use this decision framework to apply the exact DSP tool designed for each specific acoustic frequency band:

The 4-Stage Subtractive Restoration Chain

Clean audio is achieved through small, cumulative subtractions rather than a single heavy-handed plugin. Execute these 4 stages in strict sequential order:

ステップ 1：Subsonic & Handling Attenuation (80 Hz High-Pass)

The lowest fundamental tone of a deep bass voice is approximately 85 Hz (E2). The vast majority of spoken fundamentals reside between 100 Hz (male) and 220 Hz (female).

However, vibrations from desk taps, passing trucks, and microphone shock-mount resonance produce high-amplitude waveforms in the 15 Hz to 70 Hz range. While barely audible on small mobile speakers, this subsonic sludge consumes amplifier headroom and forces compressors into pumping.

• Filter Selection: 2nd-order Butterworth High-Pass (12 dB/octave slope).
• Corner Frequency: 80 Hz for general speech; 70 Hz for deep broadcast baritones.
• Phase Response: Minimum-phase Butterworth curve ensures smooth group delay without pre-ringing artifacts.

ステップ 2：Mains Ground Hum Elimination (50 Hz / 60 Hz Notch)

Alternating current (AC) power lines radiate electromagnetic fields that couple into audio cables and unbalanced interface circuits. In the Americas and Western Japan, this creates a 60 Hz hum; in Europe, Australia, and Asia, it manifests at 50 Hz.

A broad EQ cut at 60 Hz removes beneficial chest resonance from male voices. A surgical digital notch filter solves this:

• Set Center Frequency ($f_0$): 60.0 Hz or 50.0 Hz exactly.
• Set Quality Factor ($Q$): $Q = 8.0$ to $12.0$. The bandwidth of attenuation is restricted to just 5 Hz to 7.5 Hz wide.
• Gain: -24 dB. This eliminates the hum tone entirely while leaving vocal energy at 70 Hz and 90 Hz completely intact.

ステップ 3：Smooth Dialogue Pauses with Downward Expansion

A binary noise gate cuts audio to -inf dBFS the millisecond the signal drops below a threshold. This creates a jarring "on/off" pumping effect where ambient room hiss suddenly disappears when the speaker pauses and reappears the moment they speak.

A downward expander is a dynamic processor that lowers background noise smoothly and proportionately:

• Threshold: Position 3 dB above your measured quiet baseline (e.g., -40 dBFS if room tone is -43 dBFS).
• Ratio: 1:1.5 to 1:2.0. If the background noise is 4 dB below the threshold, the expander attenuates it by an extra 4 dB to 6 dB.
• Attack: 5 ms. Opens immediately on vocal consonant onset.
• Release: 250 ms. Allows natural vocal room decay without chopping sentence endings.
• Range / Floor: Clamp total attenuation to -10 dB. Room noise becomes inaudible without artificial dropouts.

ステップ 4：High-Frequency Hiss Taming & Dynamic De-Essing

Once low-end rumble and pauses are managed, evaluate the upper frequencies (above 5 kHz). Thermal preamp hiss from budget USB interfaces lives between 6 kHz and 14 kHz.

Do not use a low-pass filter (which makes the voice sound like an old telephone). Instead, combine a gentle high-shelf attenuation (-2 dB above 12 kHz) with a dynamic de-esser centered at 6.5 kHz. The de-esser acts only during harsh sibilant syllables ("s", "t", "z"), preventing ear fatigue without dulling overall vocal presence.

実践ケーススタディ：Untreated Bedroom Podcast Recording

The Problem: A podcast host records with a USB condenser mic in an untreated bedroom with hardwood floors. The recording has a continuous PC fan whine at -44 dBFS, a desk-bump rumble at 45 Hz, and occasional harsh "s" sounds.

The Resolution Workflow:

1. Import into AutoSilenceCut: The raw track is decoded directly in browser memory without lossy server re-compression.
2. Activate 80 Hz High-Pass: The low-frequency floor drops from -44 dBFS to -52 dBFS immediately as sub-rumble is cleared.
3. Engage 60 Hz Hum Notch: Clears minor USB grounding hum.
4. Apply Downward Expander: Set threshold to -42 dBFS with a 1:2 ratio. During speech gaps, PC fan noise is attenuated by an additional 8 dB, pushing the quiet noise floor down to -60 dBFS (broadcast standard).
5. Natural Pacing Cut: Long 3-second breathing pauses are trimmed to 0.32-second natural speech cushions with 8 ms micro-crossfades.
6. Export Master WAV: 16-bit 48 kHz uncompressed audio exported with zero clipping.

Pitfalls & Artifacts: Preventing Phase Distortion & Hollow Vocals

• Beware of Phase Smearing: Excessive spectral suppression damages the phase relationships between vocal harmonics, making the speaker sound "robotic" or like they are speaking through a metal pipe.
• Never Gate Overlap Noise: If a dog barks or a door slams while the speaker is talking, do not attempt to gate it. Re-record the sentence or use take-splicing to replace the phrase with a clean take.
• Protect Consonant Decay: Always verify the tail ends of words ending in "s", "k", "t", and "th". If word endings sound clipped, increase your expander release time from 150 ms to 300 ms.