Skip to content

VAD Chunking

Voice Activity Detection (VAD) chunking automatically splits long recordings into smaller chunks at natural speech boundaries, enabling real-time background transcription.

What is VAD Chunking?

VAD chunking uses AI-powered voice detection (Silero VAD) to identify when you're speaking and when you're silent. When a chunk reaches the minimum duration and a natural silence is detected, Voicepad:

  1. ✂️ Splits the audio at the silence boundary
  2. 🎯 Transcribes the chunk in the background
  3. 📝 Appends to the markdown file in real-time
  4. ♻️ Continues recording without interruption

The result: see your transcription update live while you're still recording!

Why Use VAD Chunking?

Without VAD (Traditional)

[Record 30 minutes] → [Stop] → [Wait 5 minutes for transcription] → [Read result]

With VAD Chunking

[Record 30 minutes with live transcription updates] → [Stop] → [Read result immediately]

Benefits:

  • 🚀 Reduced wait time - No waiting at the end, transcription happens during recording
  • 👀 Live feedback - Watch transcription appear in real-time
  • 💾 Memory efficient - Processes chunks instead of entire recording
  • 🎯 Natural splits - Chunks break at silences, not mid-sentence

How It Works

graph LR
    A[Microphone] --> B[Audio Buffer]
    B --> C{Duration >= Min?}
    C -->|No| B
    C -->|Yes| D[VAD Analysis]
    D --> E{Silence Found?}
    E -->|No| B
    E -->|Yes| F[Extract Chunk]
    F --> G[Background Transcription]
    G --> H[Append to Markdown]
    F --> I[Continue Recording]
    I --> B

Step-by-Step Process

  1. Audio Buffering - Incoming audio accumulates in a rolling buffer
  2. Duration Check - Once min_chunk_duration is reached, VAD analysis begins
  3. Speech Detection - Silero VAD identifies speech vs. silence regions
  4. Boundary Detection - Finds the last speech segment with sufficient silence after it
  5. Chunk Extraction - Splits audio at the silence boundary
  6. Background Transcription - Chunk transcribes while recording continues
  7. Markdown Update - Transcription appends to the file (thread-safe)
  8. Buffer Reset - Remaining audio stays in buffer for next chunk

Configuration Parameters

Enable VAD Chunking

vad_enabled: true

voicepad record start --vad

Minimum Chunk Duration

vad_min_chunk_duration - Seconds before allowing splits

vad_min_chunk_duration: 60.0  # Default: 60 seconds

voicepad record start --vad --min-chunk-duration 120

What it does:

  • Audio must accumulate this long before VAD looks for boundaries
  • Ensures chunks have enough context for accurate transcription
  • Prevents tiny, fragmented chunks

Recommended values:

  • Short answers (Q&A): 30-45 seconds
  • Normal conversation: 60 seconds (default)
  • Lectures/presentations: 90-120 seconds

Minimum: 10 seconds

Values below 10 seconds are not allowed to maintain transcription quality.

Speech Detection Threshold

vad_threshold - Probability threshold for detecting speech (0.0 - 1.0)

vad_threshold: 0.5  # Default: 0.5

voicepad record start --vad --vad-threshold 0.45

What it does:

  • Higher values = more strict (less likely to detect as speech)
  • Lower values = more lenient (more likely to detect as speech)

Recommended values:

  • Noisy environment: 0.6-0.7 (stricter)
  • Normal environment: 0.5 (default)
  • Quiet environment: 0.4-0.45 (more sensitive)

Tuning the Threshold

If chunks split too early (mid-sentence), increase threshold to 0.6. If chunks never split, decrease threshold to 0.4.

Minimum Silence Duration

vad_min_silence_duration_ms - Milliseconds of silence needed to trigger split

vad_min_silence_duration_ms: 1000  # Default: 1000ms (1 second)

What it does:

  • Controls how long a pause must be before creating a chunk boundary
  • Prevents splitting at brief pauses (breaths, short hesitations)

Recommended values:

  • Fast speech/Q&A: 500-800ms
  • Normal conversation: 1000ms (default)
  • Slow/deliberate speech: 1500-2000ms

Speech Padding

vad_speech_pad_ms - Milliseconds added to each side of speech segments

vad_speech_pad_ms: 400  # Default: 400ms

What it does:

  • Adds buffer before and after detected speech
  • Prevents cutting off beginning/end of words

Recommended values:

  • Usually keep at 400ms (default)
  • Can reduce to 300ms for very clear audio
  • Increase to 500-600ms if words are being cut off

When to Use VAD Chunking

  • Long recordings (10+ minutes)
  • Meetings and interviews
  • Lectures and presentations
  • Dictation sessions
  • When you want live transcription feedback
  • Short recordings (< 5 minutes) - overhead not worth it
  • Continuous speech with few pauses - may never chunk
  • Music or non-speech audio - VAD won't detect proper boundaries
  • When you need exact timing - chunk splits may not align perfectly

Output Format

With VAD Enabled

Markdown file updates in real-time with timestamped chunks:

# Transcription: recording_20260218_033000.wav

**Status:** Recording in progress...

---

## Chunk 1 (0:00 - 1:12)

This is the transcribed text from the first chunk...

## Chunk 2 (1:12 - 2:45)

Continuing with the second chunk...

## Chunk 3 (2:45 - 4:20)

The third chunk appears here...

---

**Recording Complete**
- Total chunks: 3
- Total duration: 4:20

Without VAD

Single transcription block after recording:

# Transcription

**Language:** English (98.5%)
**Date:** 2026-02-18 03:30:00

---

All the transcribed text in one block...

Troubleshooting

Chunks Are Too Small

Problem: Chunks are 10-20 seconds, shorter than min_chunk_duration

Solution: The minimum is enforced, so this shouldn't happen. If it does, check your configuration and verify vad_min_chunk_duration is set correctly.

Chunks Never Split

Problem: Recording for 10 minutes, but only one chunk created

Possible causes:

  1. No silence detected - Continuous speech with no pauses
  2. Threshold too high - VAD not detecting speech boundaries
  3. Minimum silence too long - Pauses aren't long enough

Solutions:

  • Lower vad_threshold to 0.4-0.45 (more sensitive)
  • Reduce vad_min_silence_duration_ms to 500-800ms
  • Speak with intentional pauses

Chunks Split Mid-Sentence

Problem: Chunks break in the middle of sentences

Solutions:

  • Increase vad_threshold to 0.6 (stricter speech detection)
  • Increase vad_min_silence_duration_ms to 1500-2000ms
  • Speak more continuously with fewer pauses

Performance Considerations

Memory Usage:

  • Buffers audio until chunk boundary found
  • Typically 60-120 seconds of audio in RAM (~2-4 MB)

CPU Usage:

  • VAD analysis runs periodically (not every frame)
  • Minimal overhead (~1-2% CPU)

Transcription Speed:

  • Chunks transcribe in parallel with recording
  • GPU: ~10-20x real-time (60s chunk → 3-6s transcription)
  • CPU: ~1-5x real-time (60s chunk → 12-60s transcription)

See Also