Dictation glossary

Voice activity detection (VAD)

Voice activity detection (VAD) is a filter that decides whether a slice of audio contains speech or is silence, noise, or other sound. Dictation and transcription use it to start and stop listening, to skip empty buffers, and to keep room tone out of the model. Neural VADs such as Silero score frames with a small network. Energy VAD uses amplitude as a fallback when no neural model is loaded.

In more detail

What is VAD?

What is voice activity detection doing in the pipeline? Audio arrives in chunks. Each chunk gets a speech-or-not label. Consecutive speech frames become an utterance; trailing silence becomes an endpoint that can trigger finalization. If VAD is too timid, you clip the first word. If it is too eager, fans and keyboards keep the recognizer busy. That is why a neural VAD, trained to ignore non-speech, usually beats a raw energy gate in an open office, while energy VAD remains useful when the neural file is missing.

Silero VAD is a widely used open model for this job: small enough to run on-device, accurate enough to replace many energy heuristics. It is not a speech-to-text model; it only answers 'is someone talking?' Endpointing for dictation still combines VAD with your push-to-talk or toggle key. A hold-to-talk session can ignore VAD for start, then use it to trim. Always-on capture without a key is a different product, and a much sharper privacy question, even with a good VAD in the loop.

For writers who speak

Why it matters for dictation

Dictation feels broken when the first syllable is missing or when the engine never stops. Those are often VAD problems, not model problems. A solid VAD also cuts work for on-device engines, which keeps latency and fans down. If you evaluate tools, ask what happens when the fancy VAD cannot load: a silent failure is worse than an honest energy fallback.

In this product

How WhisperJot handles it

WhisperJot uses a Silero VAD model when it is available, and falls back to an energy VAD if that model is not present. Capture is gated by your hotkey, not by always-on listening. Local engines transcribe on-device by default; Jot Cloud is opt-in and still depends on the same capture and VAD front end.

Questions

Straight answers.

What is voice activity detection?

Voice activity detection, or VAD, is software that labels short audio frames as speech or non-speech. Dictation apps use it to ignore silence, trim leading and trailing noise, and decide when an utterance ended. It is not transcription: it does not output words. A neural VAD such as Silero learns the difference between voice and a fan; an energy VAD only watches how loud the signal is.

What is Silero VAD?

Silero VAD is a popular open neural network for voice activity detection. It runs on-device, scores whether a frame is speech, and is small compared with a full speech-to-text model. Apps embed it so they do not have to hand-tune volume thresholds for every microphone. If the Silero file is absent, many pipelines fall back to energy-based detection rather than refusing to record.

How is VAD different from a wake word?

VAD asks whether speech is present. A wake word asks whether you said a specific phrase such as a product name. Wake-word systems often run VAD first, then a tiny keyword model. Push-to-talk dictation can skip wake words entirely because the key is the gate. Confusing the two leads people to think a dictation app is always listening when it is only watching the buffer while a hotkey is held.

One hotkey, any focused app.

Private local transcription by default, with an optional opt-in cloud engine.