Audio input: microphone and processing

Input device

Settings → Voice input → Input device, on both macOS and Windows, where it sits under the MICROPHONE heading. It lists every microphone the OS knows about: Bluetooth headsets, USB mics, built-in laptop mics, audio interfaces. Leave it on System default unless you have a reason not to.

If a dictation comes back empty, the OS most likely handed Dimmy the wrong device. Some laptops switch to HDMI when a monitor is plugged in. Pick the microphone by name instead of System default and try again.

Microphone volume

An Advanced setting, a gain slider at 50 % by default. Raise it if you speak softly and the transcript keeps losing the start of sentences. Leave it low if your mic is already loud.

Preprocessing

One toggle covering three stages: an 80 Hz highpass filter, voice activity detection, and automatic gain control. It ships off, and that is deliberate.

• The voice activity detection trimmed 40 to 60 % of the samples and misjudged exactly the quiet-microphone case it was meant to rescue. One 45-second dictation came back as "Ah!".
• The automatic gain control clipped every one of 38 measured captures, until its target was lowered and a true-peak ceiling added.
• Whisper is level-robust and noise-robust on its own, so the three stages were solving a problem it does not have.

Noise suppression

Dimmy has a neural denoiser that runs while you record rather than as a pass afterwards. It ships off and has no Settings toggle at all, because enhancement made transcripts measurably worse: on the dictations that changed at all, punctuation dropped by about a third and capitalisation by about half, and the denoiser removed none of the hallucinations it was there for. There is nothing to turn on here.

Always on, not a setting

A NaN-safe clamp keeps every sample inside [-1.0, 1.0] before anything else touches it. One bad sample used to be enough to crash Whisper. It is not user-toggleable and never has been.

The other Advanced toggles on this page

• Accelerate transcription: transcribes in short chunks while you talk, so the text is ready almost instantly when you release the key. Parakeet is realtime; Whisper needs a GPU to keep up.
• Live captions: shows a floating caption of what you are saying while you speak. Needs Accelerate transcription on.
• Realtime typing: writes each finished phrase straight into the app as you speak. Uses Deepgram if you saved a Deepgram key, otherwise your local model.

Bluetooth caveat

Bluetooth headsets often drop into the 8 kHz HFP profile when an app opens the mic, even though they support 48 kHz A2DP for output. The audio sounds tinny. Dimmy upsamples to 16 kHz before STT, which is enough for Whisper, but accuracy drops noticeably compared to a USB or built-in mic.