無音部分で音声を分割する実践的テイク仕分けワークフロー

Why Splitting at Silence Is the Foundation of Production Pipelines

In high-volume spoken-word production, capturing audio is only the first step. The real operational bottleneck lies in take preparation: slicing hours of continuous multi-take recordings into hundreds of discrete, named, normalized, and error-free audio assets.

Whether you are preparing dialogue for a video game dialogue engine (such as Wwise or FMOD), assembling an AI voice synthesis training dataset, cataloging sound design foley hits, or delivering audiobook pickup corrections, manual slicing in a DAW is repetitive, error-prone, and slow. Slicing at silence boundaries automates this process by treating deliberate pauses as natural segment delimiters, producing sequentially indexed assets with zero quality loss.

When to Split by Silence vs. When to Use Transcripts or Timecodes

Selecting the correct segmentation strategy depends entirely on the structure and noise characteristics of your source material:

Technical Slicing Specifications: Format, Framing, and Buffers

To ensure exported takes integrate seamlessly into downstream DAWs, game engines, and audio middleware, every slice must adhere to strict digital audio engineering standards:

• Bit Depth & Encoding: Slices should be exported as uncompressed 16-bit or 24-bit linear PCM WAV. Avoid MP3 or AAC for intermediate assets, as lossy encoders add 576–1152 silent delay samples that disrupt sample-accurate timing.
• Canonical 44-Byte RIFF Headers: Ensure each WAV file includes valid RIFF, WAVE, fmt (format tag 1 for PCM), and data chunk descriptors so asset loaders parse durations instantly without header errors.
• Manifest Ledger (RFC 4180 CSV): A structured metadata file bundled inside the export archive containing:Index, Filename, Start_Timecode, End_Timecode, Duration_Sec, Peak_dBFS, RMS_dBFS.

Acoustic Mechanics: Attack Cushions, Decay Tails, and Micro-Crossfades

Acoustic take slicing requires more than simply cutting where RMS energy drops below a threshold. Human speech phonetics require specialized buffer handling at both ends of every take:

ステップ 1：Leading Attack Cushion (100ms – 200ms)

Before vocal cords phonate, the speaker takes a silent or quiet preparatory inhale, and the lips or tongue position for the initial consonant. Unvoiced plosives (like /p/ in "power" or /t/ in "target") have a pre-burst silence closure lasting 40ms to 80ms. Without a 100ms–200ms leading cushion, the algorithm will slice immediately at the acoustic spike, creating a sharp, unnatural attack.

ステップ 2：Trailing Decay Cushion (200ms – 350ms)

When a sentence ends in an unvoiced fricative (such as /s/, /f/, or /th/) or a soft nasal decay (/m/, /n/), the energy drops steeply below -38 dBFS well before the sound has fully decayed into the ambient room floor. A 250ms trailing cushion ensures the complete phonetic decay and acoustic room response are preserved intact.

ステップ 3：Zero-Crossing Alignment & 8ms Micro-Fades

If a file is sliced while the digital waveform is at non-zero amplitude, playing the file causes a voltage step discontinuity that manifests as an audible click. AutoSilenceCut applies an automatic 8ms equal-power micro-fade:

Fade In: g(t) = sin((pi * t) / (2 * T)) for 0 ≤ t ≤ 8ms
Fade Out: g(t) = cos((pi * t) / (2 * T)) for 0 ≤ t ≤ 8ms

Step-by-Step Production Workflow: From Master Track to Indexed Archive

Follow this standard workflow to turn long master tracks into production-ready takes:

1. Pre-Flight Recording Check: Ensure your source audio has at least 3 seconds of silent room tone at the beginning and that the speaker leaves deliberate 1.5s–2.0s pauses between takes.
2. Load Master Track in Browser: Open Split Audio by Silence. Drag and drop your raw WAV, MP3, or FLAC master recording. The file is decoded into memory locally with zero server transfers.
3. Configure Detection Parameters:
   • RMS Threshold: Set to -38 dBFS for studio vocals (or 8 dB above measured noise floor).
   • Minimum Gap Duration: Set to 1.50 seconds to avoid false triggers on intra-sentence pauses.
   • Leading Cushion: 150 ms | Trailing Cushion: 250 ms.
4. Review Slice Boundaries on the Waveform: Scan the interactive waveform visualizer. Yellow marker lines show slice cut points. Use the slice audition controls to spot-check the beginning of Take 1, Take 2, and Take 10.
5. Export Lossless ZIP Bundle: Click Export Take Archive. The engine renders discrete 16-bit PCM WAV takes (take_001.wav, take_002.wav...) and packages them with manifest.csv into a single downloadable ZIP.

Production Case Study: Preparing 120 Voice Lines for a Video Game

Context: A game studio recorded 120 character lines (battle cries, dialogue responses, error messages) in a continuous 52-minute session at 48 kHz / 24-bit PCM. The audio director needed all 120 lines exported as discrete, numbered WAV files for batch import into Audiokinetic Wwise.

Workflow Comparison

Common Slicing Pitfalls and the 5-Point Quality-Control Checklist

ステップ 1：Slicing During Mid-Sentence Thinking Pauses

Symptom: A single continuous line with a thoughtful 0.9-second pause is accidentally divided into two separate files.
Fix: Increase the Minimum Gap Duration to 1.6s–2.0s. Coach talent to maintain a continuous cadence within takes and pause only between takes.

ステップ 2：Missing Attack Transients on Explosive Consonants

Symptom: The initial /k/ in "Combat ready!" sounds truncated or muffled.
Fix: Increase the Leading Cushion from 50ms to 150ms–200ms to capture the full pre-phonation closure.

ステップ 3：Inconsistent File Naming in Downstream Pipelines

Symptom: Asset management tools fail because files are named with inconsistent numbering (take_1.wav vs take_001.wav).
Fix: Always use 3-digit zero-padded naming (take_001.wav through take_999.wav) to maintain alphabetical and chronological sorting across all operating systems.

ステップ 4：Noise Floor Pumping Between Takes

Symptom: Total digital silence between takes creates an abrupt drop in room ambience.
Fix: Use Room Tone Matcher to fill inter-take gaps with synthetic room ambience synthesized via Overlap-Add (OLA).

ステップ 5：Slicing Audio with High DC Offset

Symptom: Every sliced file exhibits a subtle pop when loaded into sample players.
Fix: Apply an 80Hz high-pass filter before slicing to remove DC bias, and ensure 8ms micro-crossfades are active at all slice boundaries.