Voice activity detection: measure energy, inspect mistakes

Exact evidence inspector

Read the selected state numerically. This includes the details behind the chart rather than a preset verdict.

Measure a frame before deciding

This energy detector processes six seconds of deterministic synthetic audio at eight kilohertz. Every twenty-millisecond frame contains 160 samples. The model squares those samples, averages them and takes the square root to obtain RMS amplitude. It compares that measurement with a threshold rather than inventing an energy score. The blue ground-truth row marks two known speech-tone intervals; the green detector row shows decisions, including mistakes and delayed release. A tone is only a teaching surrogate for speech, not recorded conversation or a trained voice classifier.

Tune activation and release

Activation RMS threshold determines when an idle detector turns on; the comparison is strictly greater than the threshold. Release threshold ratio multiplies that threshold while the detector is active. A smaller ratio provides hysteresis so small dips do not immediately switch it off. Hangover keeps activity for the chosen time after energy falls below the applicable threshold, quantized to twenty-millisecond frames. Uniform noise amplitude changes the same seeded noise samples. Include non-speech tone burst inserts a loud interference interval from 0.8 to 1.1 seconds. Changing any input recomputes the signal and pauses the inspection at the beginning.

Separate sensitivity from correctness

Disable the burst, set noise to zero, hangover to zero and release ratio to one. At an activation threshold of 0.10, the loud speech interval is detected but the quiet interval is missed: the final counts are sixty correctly detected speech frames and sixty missed speech frames. Lower the threshold to 0.05 and both intervals should be detected. Re-enable the burst: it can trigger false speech because an energy threshold does not understand the sound source. Increase hangover and inspect the frames just after a speech interval; the detector stays active on frames whose ground-truth label is silence. These are deliberately counted as false speech, not hidden from the result.

Inspect the evidence and limits

Play reveals three hundred frames over sixty inspection seconds, ten times slower than the six-second signal. Pause holds the view; Step adds one inspection second; Restart rewinds; Speed changes viewing pace; Inspect time selects a frame boundary. Restore defaults resets the policy. The inspector gives the latest frame’s samples, RMS, effective gate, activity and cumulative confusion counts. Counts describe this synthetic fixture only: they are not accuracy claims for a deployed detector. The model omits frequency features, learned classifiers, microphone processing, real background datasets, endpointing policies and language-dependent pauses. No microphone permission is requested, no audio is played, and no claim of near-zero false positives is made.