Tuning

The defaults are tuned for dictation on a phone and are the right place to start. Change them when your users speak differently, for example in short commands or with long pauses mid-thought.

When a sentence ends

EndpointerConfig controls how the session splits speech into sentences. It is passed to start(), transcribeStream() or replay().

await session.start(
  endpointer: const EndpointerConfig(
    hangover: Duration(milliseconds: 700), // wait longer before ending
  ),
);

Setting	Default	What it does
hangover	400 ms	How much silence ends a sentence. Longer lets people pause mid-thought; shorter returns text sooner.
minSpeechDuration	250 ms	Sounds shorter than this, such as coughs and clicks, are ignored (and reported as SpeechDiscarded).
reattachWithin	400 ms	A short sound right after speech is kept anyway, so a quick “yes” or “no” survives.
maxUtterance	20 s	The longest sentence. Someone reading without pausing is split here, with a small overlap so no word is lost.
speechThreshold	0.5	How sure the voice detector must be that speech has started.
silenceThreshold	0.35	Below this, the detector counts silence. Lower than the start threshold, so a sentence is not cut off mid-word.
preRoll	300 ms	Audio kept from before speech starts, so the first syllable is not clipped.
postRoll	250 ms	Audio kept after speech ends.

Durations are rounded up to the voice detector’s 32 ms steps.

The example app offers three presets to try: eager (250 ms hangover), balanced (400 ms, the default) and patient (700 ms).

Automatic gain

Quiet speakers, or a phone held far away, can be too quiet for the voice detector to notice. The session boosts quiet audio and calms loud audio before anything listens to it. It is on by default; pass normalizer: null to turn it off.

The gain never boosts a noisy room’s background into “speech”: the boost is capped by how loud the room’s own noise is. The defaults rarely need changing; NormalizerConfig documents each setting.

Compare settings on the same audio

Two live runs never hear the same audio, so comparing settings live tells you little. Record once, then replay the recording under each setting:

await session.start(recordTo: '${dir.path}/take.wav');
// ...speak...
await session.stop();

final audio = Wav.read('${dir.path}/take.wav');
for (final config in [eager, balanced, patient]) {
  await session.replay(audio, endpointer: config);
  print(session.text.value.settled);
}

replay runs the recording through exactly what the microphone uses, as fast as the phone can. Its audio must be 16 kHz mono float, which is what recordTo writes.