Canonical Technical Definition

An audio signal processing technique that detects the presence or absence of human speech in an audio stream, filtering out background silence and noise.

Voice Activity Detection (VAD)—also known as speech activity detection or silence suppression—is a fundamental audio processing technique that continuously analyzes an audio stream to classify whether each time frame contains human speech or non-speech (silence, room reverberation, air conditioning hum, or keyboard typing).

In telecommunications, VAD was originally developed to conserve bandwidth by suspending voice transmission during natural conversational pauses. In modern desktop artificial intelligence, VAD is a vital efficiency gateway that guards expensive neural networks from processing dead air.

How VAD algorithms operate

VAD systems have evolved across two distinct architectural paradigms:

1. Energy and feature-based heuristics (Classic VAD)

Legacy VAD algorithms compute mathematical metrics across 10–30 millisecond audio frames:

  • Short-Time Energy: Total signal amplitude; speech typically exhibits higher energy than room ambience.
  • Spectral Centroid & Flux: The frequency center of gravity; human vocal tracts produce energy concentrated between 300 Hz and 3,400 Hz.
  • Zero-Crossing Rate (ZCR): How frequently the waveform crosses zero amplitude; unvoiced consonants exhibit high ZCR, while vowels exhibit low ZCR.

While computationally cheap, classic heuristics fail in noisy environments (such as coffee shops or open offices) where typing or passing traffic triggers false speech positives.

2. Neural VAD (Modern Standard, e.g., Silero VAD)

Modern note software utilizes compact neural networks (often under 1 MB in size) trained specifically to recognize human voice acoustic signatures:

[Incoming Audio Frame (30ms)]
              ↓
  [Compact 1D CNN / LSTM]
              ↓
[Speech Probability: 0.0 - 1.0]
              ↓
  [Hysteresis Thresholding] ───> State: SPEECH / SILENCE

Neural VAD delivers 99%+ precision, cleanly ignoring mechanical keyboard clatter, paper shuffling, and ambient fan noise.

The operational benefits of VAD in note applications

MetricWithout VAD (Continuous Audio Feed)With Neural VAD (Oyma)
Inference Duty Cycle100% of meeting duration~45% (only when participants speak)
Battery Life ImpactHigh sustained GPU/NPU power drawMinimal; hardware sleeps during pauses
Chunking Boundary QualityRandom fixed cuts; words get choppedNatural cuts at natural conversational pauses
Hallucination RiskHigh (Whisper hallucinates on silence)Zero (silence is never fed to Whisper)

VAD in Oyma on macOS

In Oyma, Voice Activity Detection runs continuously on high-efficiency Apple Silicon cores:

  • When a speaker finishes a thought and pauses for 400 milliseconds, VAD marks the boundary and dispatches the accumulated buffer to Whisper.
  • Whisper transcribes a coherent, complete grammatical sentence.
  • During quiet stretches of a video call, neural inference pauses entirely, preserving your MacBook’s battery and keeping system fans silent.

Frequently asked questions about Voice Activity Detection (VAD)

Why is Voice Activity Detection critical for on-device transcription?

Running neural speech models on empty silence or keyboard clicks wastes CPU/GPU cycles and battery life. VAD ensures models only execute when actual human speech is present.

How do modern VAD algorithms distinguish speech from background noise?

Modern algorithms (like Silero VAD) use compact convolutional neural networks trained on acoustic features, spectral zero-crossing rates, and harmonic energy distributions.

How does Oyma use VAD during meetings?

Oyma integrates low-overhead VAD to segment audio streams at natural pauses, reducing energy consumption and flushing complete sentences to Whisper cleanly.