Dictating in a noisy room: microphone and silence settings

Why does a noisy room make dictation worse?

Because the speech model gets whatever the microphone picked up, and a microphone does not know which sound is you. A fan, a street, a coffee machine and a colleague on a call all arrive in the same recording as your voice. The model has to find the words under all of it, and the less your voice stands out, the more often it guesses.

Steady noise, like a fan or air conditioning, mostly costs accuracy on quiet syllables: word endings, short words, names. Speech in the background is harder, because to a recognizer it is not noise at all. It is more speech, and a loud enough neighbour can put a few of their words into your paragraph.

The second cost is easy to miss. If you use stop on silence, Dictera decides the recording is over by watching the input level. In a noisy room the level never really drops, so the recording may never end on its own. Or a single loud sound can start the clock before you have said anything.

What helps more, the microphone or the settings?

The microphone, by a distance. Nothing in Dictera filters the audio before the speech model hears it, so what the microphone captures is what gets transcribed. The silence settings decide when a recording ends. They do not remove noise from it, and raising the voice threshold does not make the background quieter.

What matters about a microphone in a noisy room is less its price than how close it is to your mouth. Sound from a nearby source gets louder quickly as you approach it, so halving the distance to your mouth lifts your voice by about 6 dB while the room stays where it was. A microphone a few centimetres away hears mostly you. The one in a laptop lid, half a metre away, hears you and the room in roughly equal measure.

Which microphone should I use?

In rough order, for a noisy room:

• A headset with a boom microphone near the corner of your mouth. The distance stays the same when you turn your head, which is the whole trick.
• Earbuds with built-in microphones. Further away than a boom, but they move with you, and many are designed to favour the wearer's voice.
• A desk microphone close to you, pointed at your mouth. Good in a quiet room; in a noisy one, only as good as how close you keep it.
• The built-in microphone. Fine in a quiet study, and the first thing to replace when the room gets loud.

Dictera uses the system default input unless you pick one. To pin a specific microphone, choose it under Microphone in Settings ▸ Dictation, or from the notch, which is quicker when you move between a desk and a café: hold the Microphone button and pick one under Dictate With, where System Default comes first. If the microphone you pinned is unplugged, Dictera falls back to the system default rather than failing to record.

A plain click on that same button does something else: it mutes every microphone on the Mac, for every app, and Dictera's own dictation hears the silence too. While it is muted, the notch says "Your microphone is muted" as you dictate. If a dictation ever comes back empty in a quiet room, check that first.

How do I see what my microphone hears?

With the input meter, Input Level in Settings ▸ Dictation ▸ Recording. Press Test and the bar shows the live level of the selected microphone without recording anything. Watch it for ten seconds with the room as it normally is and you are not talking: that is your room's floor. Then say a sentence at your normal volume: that is your voice.

The gap between the two is what everything else depends on. A wide gap means the model hears you clearly and stop on silence can tell you from the room. A narrow one means the microphone is too far away, or the room is too loud for it, and no slider will close that gap.

The bar turns yellow and then red as it nears the top. If your ordinary speaking voice reaches red, the microphone is close enough to distort, and moving it back a little helps more than it costs.

The meter is scaled in decibels, not in raw loudness. Its left edge is very quiet, −50 dB below the loudest signal the microphone can pass, and each tenth of the bar is about 5 dB. A fan that moves the bar from 0.1 to 0.2 is no small change: it is roughly three times the sound energy.

Why does stop on silence not stop?

Because the room is louder than your voice threshold. With stop on silence on, anything above the threshold counts as speech, and the recording only ends after the level has stayed below it for as long as Stop After says. The threshold starts a tenth of the way up the meter, near the very bottom, set for a quiet room. A fan or a busy office can sit above it all the time, so to Dictera you never stop talking.

With stop on silence on, the meter in Settings shows the threshold as a small tick on the bar. Move the Sensitivity slider until the tick sits clearly above the room's floor and clearly below your voice, roughly halfway between them: lowering Sensitivity raises the tick. Then test it: dictate a sentence, stop talking, and see whether the recording ends when it should.

If there is no position that is both above the room and below your voice, the gap is too narrow. Move the microphone closer, or leave stop on silence off in that room and end recordings with the shortcut instead.

Why does it stop before I have started?

Because a single loud sound counted as the start of your speech. The clock for Min Speech Duration starts at the first moment the level crosses the threshold, whether that moment was you or a door. Once that much time has passed and the room has then been quiet for as long as Stop After, the recording ends, even if you had not said a word yet.

In a room with sudden sounds rather than steady ones, raise Min Speech Duration, under Fine-tune, from its default of one second to two or three. A cough, a dropped cup or a slammed door is shorter than that, so the time since it has to stretch further before a pause can end the recording. It costs you nothing when you are actually talking.

The third slider, Stop After, is about how you speak rather than about the room. It runs from half a second to three seconds, two by default. Leave it alone for noise; it is covered in the post on recording modes, along with when stop on silence is worth having at all.

What if a few wrong words still get through?

Some will, in a loud enough room. If you have cleanup turned on, the model that runs after the speech model tidies punctuation and fillers and can fix an obvious slip in context, but it cannot recover a word the recognizer never heard clearly. Two things narrow what is left:

• Custom replacements for the names and terms that a noisy room mangles the same way every time. They run last, so they win. The tip on replacements shows how.
• A different speech model. The models Dictera offers are built differently and cover different languages, so one may cope with your voice and your room better than another, and switching is a setting. Choosing a speech model goes through them.

And the oldest fix of all: say the sentence again when the room is quieter. Esc ends a recording, and by default throws it away, which is what you want when a train went past mid-sentence.

A setup for common rooms

All of these settings are per Mac, not per room, so a setup tuned for the office is still in place when you get home. The meter takes ten seconds to read, and after a week you will know without testing which rooms need a different microphone.

None of this sends anything anywhere. The meter, the threshold and the recording itself stay on your Mac, as described on the dictation page; the full list of recording settings is on the features page.