無音カット設定ガイド：しきい値dB・最小無音長・保持クッションの選び方

The 3 Fundamental Settings Behind Every Silence Remover

Automated silence removal is not magic; it is a deterministic calculation applied to root-mean-square (RMS) energy windows across an audio buffer. Understanding how the three core parameters interact allows you to achieve clean, transparent edits on any microphone recording.

ステップ 1：Decibel Sensitivity Threshold (dBFS / RMS Energy)

The decibel threshold defines the energy boundary below which audio is classified as silence. In digital systems, level is measured in decibels relative to full scale (dBFS), where 0 dBFS is digital maximum and negative numbers indicate quieter signals.

• Balanced (-38 dBFS): The industry baseline for treated home studios and controlled dialogue. Distinguishes conversational speech from baseline background air.
• Sensitive (-32 dBFS): A higher threshold that classifies soft room reverberation, subtle rustles, and shallow breathing intakes as silence.
• Conservative (-46 dBFS): A lower threshold reserved for recordings in quiet treated booths where you only wish to isolate deep dead-air stalls exceeding 1.5 seconds.

ステップ 2：Minimum Gap Duration (Trigger Length)

The minimum gap setting establishes how long an acoustic passage must remain below the decibel threshold before the engine flags it as a candidate for trimming.

Acoustic phonetics proves that human speech naturally contains micro-silences. Stop consonants (such as /p/, /t/, /k/, /b/, /d/, /g/) require an articulatory closure lasting 50ms to 140ms where zero air escapes the vocal tract. If your minimum gap is set too short (e.g., 200ms), the engine mistakenly flags these articulatory closures as dead air, cutting words in half. Setting the minimum gap to 0.60s to 0.80s protects internal word structure while capturing sluggish conversational pauses.

ステップ 3：Retained Pause / Speech Cushion (Safety Margin)

The speech cushion defines the amount of time preserved at the boundaries of every trimmed gap. When an engine detects a 2.0-second silence gap, it does not delete 2.0 seconds. Instead, it trims out the center of the gap while retaining a calibrated cushion (e.g., 0.32 seconds total, split as 0.16s after the preceding syllable and 0.16s before the incoming syllable).

This safety margin protects trailing phonetic decays (fricative trails like /s/ and /f/) and provides room for subtle, natural inhalations between sentences.

When Parameter Tweaking Helps vs When Raw Audio Must Be Fixed First

Adjusting detection settings can solve pacing and dead-air issues, but cannot overcome severe recording defects at the source.

Conditions Suited for Algorithmic Detection

• Consistent Signal-to-Noise Ratio (SNR > 25 dB): Clean recordings where speech peaks sit between -18 dBFS and -6 dBFS, and ambient noise sits below -45 dBFS.
• Single-Speaker Tracks: Solo narrations, voiceovers, and isolated microphone stems where speech levels are relatively uniform.
• Predictable Room Acoustics: Dry rooms without long reverberant decays exceeding 300 milliseconds.

Acoustic Issues That Settings Alone Cannot Solve

• Severe HVAC Rumble or Electrical Mains Hum: High low-frequency energy (e.g., 60Hz hum at -35 dBFS) tricks the silence detector into thinking speech is occurring continuously. Apply an 80Hz high-pass filter or 50/60Hz notch filter before running silence detection.
• Extreme Room Reverb (Echo): Long acoustic reflections in untreated tile or glass rooms decay slowly over 800ms, blurring the boundary between speech and silence.
• Severe Bleed on Multi-Mic Panels: If Host A's voice bleeds into Host B's microphone at -28 dBFS, standard silence detectors cannot isolate the true inactive gaps on Host B's stem.

Complete Calibration Matrix Across 6 Spoken-Word Formats

Select the appropriate parameter profile based on your project format and target audience expectations:

5-Step Calibration Sequence for Any Microphone Recording

ステップ 1：Measure Ambient Noise Floor in dBFS

Locate an unvoiced segment at the start or end of your track. Observe the RMS meter level. If the baseline bounces between -54 dBFS and -50 dBFS, your baseline noise floor is approximately -52 dBFS.

ステップ 2：Calculate the Optimal RMS Decibel Threshold

Apply the +12 dB Headroom Rule: add 12 dB to your measured noise floor. If your noise floor is -52 dBFS, set your threshold to -40 dBFS (-52 + 12 = -40). This guarantees that ambient air never triggers the detector while ensuring soft speech tails (-30 dBFS) are protected.

ステップ 3：Dial In Minimum Gap Duration Based on Speech Tempo

For normal conversational speech, set the minimum gap to 0.75 seconds (750ms). If the speaker talks rapidly, reduce to 0.55 seconds. If the delivery is deliberate and formal, increase to 1.00 second.

ステップ 4：Assign the Retained Speech Cushion

Default to 0.32 seconds (320ms). If you want a tight, high-energy delivery for social video, select 0.15 seconds. Never drop below 0.10s unless intentionally creating an artificial comedic jump-cut style.

ステップ 5：Perform A/B and Diff Playback Verification

Use the timeline audition tools. Listen to the Edited version for natural flow, then listen to the Diff mode (which plays strictly the audio segments marked for deletion). If you hear clear consonants or word fragments during Diff playback, your threshold is set too high or your cushion is too narrow.

実践ケーススタディ：Untreated Room with USB Mic & PC Fan Noise

A podcaster records a 40-minute interview using a USB condenser microphone placed 6 inches from their mouth. The room has bare hardwood floors, creating minor flutter echo, and a desktop PC fan produces a steady noise floor of -46 dBFS.

Calibration Breakdown:

• Initial Problem: Using a default -42 dBFS threshold failed to detect any pauses because the -46 dBFS fan noise was too close to the detection floor.
• Correction Applied:
  1. Applied an 80Hz high-pass filter to strip PC chassis rumble, dropping the noise floor to -50 dBFS.
  2. Set the RMS threshold to -36 dBFS (14 dB above the filtered noise floor).
  3. Set Minimum Gap to 0.80s and Retained Cushion to 0.32s with 8ms micro-crossfades.

Outcome: The engine cleanly identified 112 true dead-air intervals totaling 6 minutes and 14 seconds. Trailing vocal consonants remained fully audible, fan noise did not stutter, and the podcast dialogue flowed naturally.

5 Common Configuration Errors and How to Prevent Them

1. Setting Threshold Above -30 dBFS: This treats soft word tails, whispered emphasis, and trailing vowels as dead air, cutting off the ends of sentences.
2. Setting Minimum Gap Below 300ms: This chops intra-word stop consonants (such as the gap in "doc-tor"), creating a broken, stuttering cadence.
3. Zero-Cushion Gating: Snapping words together with 0ms padding makes speakers sound out of breath and mechanically spliced.
4. Failing to Re-Check After Changing Microphones: Dynamic broadcast microphones (e.g., Shure SM7B) have different sensitivity and noise floor characteristics than condenser mics. Always re-measure the baseline noise floor when switching hardware.
5. Skipping the "Diff Audition" Check: Always audition the cut audio in isolation to confirm no intelligible speech fragments are being discarded.