ナレーション・オンライン講義・インタビュー別：無音検出設定の選び方

The Triad of Silence Detection: Threshold, Duration, and Cushion

Automated silence removal is not a simple on/off switch. Every speech editing engine relies on three mathematical parameters that interact dynamically with the speaker’s voice volume, room acoustics, and delivery tempo:

Mathematically, when a quiet interval between timestamp t1 and t2 exceeds T-min, the removed dead-air block is defined as:

Cut Block = [ t1 + Tcushion, t2 - Tcushion ]

If (t2 - t1) ≤ 2 × T-cushion, no cut occurs, guaranteeing that short pauses are never partially truncated or unnaturally squeezed.

When to Automate Silence Cuts vs. When Manual Editing Is Mandatory

Understanding the acoustic boundaries of automated silence processing prevents ruined takes and saves hours of revision time:

Ideal for Automated Processing

• Solo Presenter Screen Recordings & Tutorials: High signal-to-noise ratio, consistent microphone placement, and predictable cognitive pauses between steps.
• Scripted Corporate Voiceovers: Professional studio booths with steady -55 dBFS noise floors where pauses represent intentional breaks or retake markers.
• Solo Podcast Monologues: Structured narrative episodes where trimming dead air tightens pacing without altering guest interactions.

Requires Manual or Semi-Automated Audition

• Emotional Drama / Acting Audio: Long dramatic silences (1.5s–3.0s) convey character subtext and must not be mechanically compressed.
• Overlapping Panel Discussions: When multiple speakers laugh, interject, or speak simultaneously, automatic gating can clip quiet side-comments. Use multi-track timeline editing instead.
• Extreme Dynamic Whispers: If a speaker drops to a soft whisper at -45 dBFS while normal speech is at -18 dBFS, the dynamic range spans 27 dB, risking whisper deletion.

Calibration Matrix Across 6 Voice Production Scenarios

Use this field-tested parameter guide based on standard post-production guidelines:

Acoustic Engineering: Balancing Noise Floor, SNR, and Vocal Decay

Setting an accurate threshold requires measuring the track's Signal-to-Noise Ratio (SNR) and identifying the energy level of the softest spoken consonants.

The 10 dB Acoustic Rule

In human phonetics, vowels (such as /a/ and /o/) concentrate immense acoustic energy between 200 Hz and 2 kHz, typically peaking around -14 dBFS to -20 dBFS. In contrast, unvoiced fricatives and stop bursts (such as /s/, /f/, /t/, and /k/) have broad high-frequency spectrums (3 kHz–8 kHz) with RMS energy often hovering between -35 dBFS and -42 dBFS.

To ensure seamless dialogue editing:

• Your silence threshold must sit at least 6 dB above the steady-state noise floor to avoid false triggers from air conditioners or pre-amp hiss.
• Your threshold must sit at least 6 dB to 10 dB below the lowest consonant energy to prevent clipping word endings.

Optimal Threshold = Noise Floor (dBFS) + (0.35 * (Speech Peak dBFS - Noise Floor dBFS))

Step-by-Step Parameter Tuning Protocol for Any Spoken Track

Follow this 5-step engineering checklist to dial in perfect settings for any audio recording:

1. Measure the Baseline Noise Floor: Open Split Audio by Silence or Remove Silence from Audio and check the delivery metrics panel. Note the measured ambient noise floor (e.g., -54 dBFS).
2. Identify the Spoken Dynamic Range: Find a quiet sentence ending in an unvoiced consonant (e.g., "We need to fix this aspect"). Check the peak meter for the word "aspect" (e.g., -32 dBFS).
3. Set Initial Threshold: Place the threshold midway between the room noise floor and the quietest consonant (for example, with a -54 dBFS floor and -32 dBFS consonant, set threshold to -38 dBFS).
4. Set Minimum Gap and Cushion: For standard content, choose the Natural preset (0.75s minimum gap, 0.32s retained cushion). If working on high-energy promotional video, choose Tight (0.50s gap, 0.15s cushion).
5. Perform the Sibilant Audition Test: Use the interactive player to preview proposed cuts on 3 key phonetic scenarios:
   • A sentence starting with /p/ or /b/ (verify pre-speech cushion captures the inhale).
   • A sentence ending in /s/ or /t/ (verify post-speech cushion captures the decay).
   • A mid-sentence thoughtful comma (verify the pause is preserved or trimmed comfortably).

Production Case Study: Re-Pacing a 30-Minute Technical Lecture

Project Overview: A computer science professor recorded a 32-minute screen walkthrough explaining distributed databases. The recording had excellent technical content but suffered from heavy cognitive pauses, slide transitions, and throat clearing.

Iterative Calibration Results

• Test 1 (Over-Aggressive Settings): Threshold = -30 dBFS | Min Gap = 0.30s | Cushion = 0.05s.
  Result: Cut 8.2 minutes, but sounded unnatural and breathless; sliced off the /t/ in "commit" and "socket". Unusable.
• Test 2 (Over-Conservative Settings): Threshold = -48 dBFS | Min Gap = 1.50s | Cushion = 0.50s.
  Result: Cut only 45 seconds; left numerous awkward 1.2-second dead-air pauses during code compilation screens.
• Test 3 (Optimized Calibration): Threshold = -39 dBFS | Min Gap = 0.65s | Cushion = 0.28s with 8ms micro-crossfades.
  Result: Removed 5 minutes 42 seconds of pure dead air. Retained 100% of student comprehension pauses and natural breathing beats. Video length reduced to a crisp 26:18.

Top 5 Calibration Traps and How to Avoid Them

ステップ 1：Chasing "Zero Silence"

The Trap: Believing that deleting 100% of quiet gaps creates the fastest, most engaging content.
Acoustic Reality: Complete absence of pauses triggers listener fatigue within 90 seconds. Human cognition requires brief 250ms–350ms pauses to assimilate complex data. Always maintain at least a 0.25s–0.32s cushion.

ステップ 2：Neglecting the Pre-Speech Inhale Cushion

The Trap: Setting leading cushion to 0ms.
Acoustic Reality: The vocal cords tense and an acoustic intake breath occurs 80ms–150ms before explosive phonemes. Cutting directly at the first high-amplitude sample produces an unnatural, clipped start.

ステップ 3：Using One Threshold for Untreated Multi-Mic Audio

The Trap: Applying identical -36 dBFS thresholds to two hosts with different microphones and vocal projections.
Acoustic Reality: The louder host will trigger properly, while the softer host's words will be cut in half. Process individual speaker stems independently before stem summation.

ステップ 4：Compressing Audio Before Silence Trimming

The Trap: Applying a dynamic compressor or loudness normalizer before editing silence.
Acoustic Reality: Compressors boost room tone during quiet intervals, reducing the gap between speech and silence from 25 dB down to 10 dB, causing false detection triggers.

ステップ 5：Ignoring Non-Zero Crossings at Edit Boundaries

The Trap: Splicing chunks without micro-fades.
Acoustic Reality: If the waveform voltage is +0.2V when the cut starts, the jump to 0.0V creates an instantaneous DC step click. Always ensure 8ms linear or equal-power crossfades are applied to all cuts.