vad

Voice Activity Detection

Detect speech regions in an audio file. Two output modes:

1. Sidecar mode (default) — write a JSON + CSV of [start, end] speech segments next to the input.
2. Split mode — cut the input into per-utterance files in a vad-segments/ folder.

Arguments ($ARGUMENTS)

• <input> (required) — path to an audio file (any format ffmpeg handles).
• --mode=<sidecar|split> (optional, default sidecar).
• --out=<dir> (optional) — output directory. Defaults to the input's directory for sidecar mode, <input-dir>/vad-segments/ for split mode.
• --min-speech=<seconds> (optional, default 0.3) — minimum segment length to keep.
• --min-silence=<seconds> (optional, default 0.5) — minimum silence to treat as a break.
• --threshold=<dB> (optional, default -40) — peak threshold below which counts as silence (used when falling back to ffmpeg).
• --engine=<auto|silero|webrtc|ffmpeg> (optional, default auto).

Engine selection

1. If the user's goal is "chunk and transcribe"

Hand off to the Claude-Transcription-Plugin — it already wraps VAD-aware transcription presets (Gemini, Whisper, AssemblyAI). Only run this skill standalone when the user wants the raw VAD output (timing sidecar or split files) without transcription.

2. Prefer Silero VAD (Python)

If silero-vad is importable in the user's Python environment, use it:

import torch, torchaudio
from silero_vad import load_silero_vad, get_speech_timestamps, read_audio

model = load_silero_vad()
wav = read_audio("<input>", sampling_rate=16000)
segs = get_speech_timestamps(
    wav, model,
    sampling_rate=16000,
    min_speech_duration_ms=int(<min-speech>*1000),
    min_silence_duration_ms=int(<min-silence>*1000),
    return_seconds=True,
)

3. Fallback — WebRTC VAD

If Silero isn't available but webrtcvad is, use 30 ms frames at aggressiveness 2, coalesce adjacent voiced frames, and apply the same min-speech / min-silence filters.

4. Last-resort fallback — ffmpeg silencedetect

ffmpeg -hide_banner -i "<input>" -af silencedetect=noise=<threshold>dB:d=<min-silence> -f null - 2>&1

Parse silence_start / silence_end lines; invert to derive speech regions.

Tell the user which engine was used in the final report.

Sidecar mode

Emit two files next to the input:

• <basename>.vad.json — [{"start": <float>, "end": <float>, "duration": <float>}, ...]
• <basename>.vad.csv — index,start,end,duration with seconds to 3 decimal places.

Include a summary block in the chat response: engine used, segment count, total speech time, total silence time, speech ratio.

Split mode

For each detected segment, cut losslessly where possible:

ffmpeg -hide_banner -ss <start> -to <end> -i "<input>" -c copy "<out>/<basename>-seg<NNN>-<start>s.<ext>"

If stream-copy is not viable (e.g. the container/codec can't start at non-keyframe boundaries), re-encode to WAV PCM:

ffmpeg -hide_banner -ss <start> -to <end> -i "<input>" -c:a pcm_s16le -ar 16000 -ac 1 "<out>/<basename>-seg<NNN>.wav"

Report: number of segments written, output directory, total output size.

Safety

• Never overwrite the original input.
• If the output directory already contains vad-segments/, ask before overwriting.
• If the detected speech ratio is < 5% of the file, warn the user — the threshold may be too aggressive.

Typical use cases

• Transcribe long recordings — split before transcribing to stay inside model context windows.
• Remove dead air — reassemble only the speech segments for a condensed version (follow up with /audio-production:concat-audio).
• Diarisation prep — VAD segmentation gives the diariser clean boundaries to work from.
• Find silent gaps — invert the sidecar to get silence regions for trimming or cue-point markup.