Dictamic

dicta·mic — an adaptive microphone pipeline for dictation.

Fix terrible voice-recognition / dictation accuracy from your laptop's internal microphone on Linux (PipeWire).

Many laptops — especially those with Realtek HDA codecs like the ALC257 — have a nasty trap: the desktop microphone volume slider maps directly onto analog gain stages (Capture up to +30 dB plus Internal Mic Boost up to +30 dB). Anywhere near the top of the slider, speech saturates the ADC. The result is clipped, crackling audio that no app, filter, or software AGC can repair — because the damage happens before the digital domain. Speech-to-text accuracy collapses while the mic still "works."

Two more common aggravators:

• HDA power-save wake-up pop: with snd_hda_intel power_save=1 the codec suspends after ~1 s idle, and on resume some ADCs rail at full scale for ~0.5 s — every recording starts with a deafening pop that confuses recognizers.
• Aggressive VAD/gating filters (e.g. Speex voice-activity detection) that chop word onsets.

What this package does

mic (fixed, conservative analog gain — slider stays ~35%)
        │
        ▼
PipeWire module-echo-cancel (WebRTC audio processing)
  • adaptive gain control — adapts to room, distance, mic
  • noise suppression  • high-pass filter  • echo cancellation
        │
        ▼
virtual source "ASR Mic (AGC + noise suppression)"  → set as default input

The analog stage keeps generous headroom so speech can never clip; the WebRTC AGC (the same engine browsers and Zoom use) levels speech digitally per environment. Works identically for any plugged-in mic — keep its raw device volume low and the AGC does the rest.

Measured on the reference laptop (30 s speech test, internal mic):

Metric	Before	After
Clipped samples	hundreds–thousands	0
Noise floor	−34 dB	−54 dB
Speech level	rail-to-rail	~−10 dB, AGC-held
Startup pop	0.5 s full-scale rail	none

Also included: Dictamic, a small GTK app showing pipeline status, a live level meter of what apps actually hear, a bypass toggle, and an autostart-at-login option.

Install

From the latest release:

sudo apt install ./dictamic_*.deb
systemctl --user restart pipewire pipewire-pulse wireplumber
pactl set-default-source asr_mic

Then set the raw mic's volume to ~35% (once):

pactl set-source-volume alsa_input.<your-device> 35%

Launch Dictamic from your app menu to verify: it shows whether the pipeline is running, warns if the raw volume is dangerously high, and lets you bypass/re-enable processing or enable autostart.

The wake-up-pop fix (power_save=0) fully applies after the next reboot; the postinst applies it to the running kernel immediately.

Uninstall

sudo apt remove dictamic
systemctl --user restart pipewire pipewire-pulse wireplumber

Files installed

Path	Purpose
/etc/pipewire/pipewire.conf.d/99-asr-mic.conf	WebRTC AGC virtual source
/etc/wireplumber/wireplumber.conf.d/99-mic-no-suspend.conf	keep ADC powered (no wake pop)
/etc/modprobe.d/50-audio-no-powersave.conf	disable HDA power save
/usr/bin/dictamic	GTK monitor/toggle GUI
/usr/share/applications/dictamic.desktop	app menu entry

Building the package

./build-deb.sh 1.1.0        # → build/dictamic_1.1.0_all.deb

Tuning

• Never push the raw device volume past ~50% — that's where the slider starts engaging analog boost and clipping returns. ~35% is the sweet spot on the reference hardware (speech peaks ≈ −10…−6 dBFS pre-AGC).
• If a recognizer prefers rawer audio, disable noise suppression by setting webrtc.noise_suppression = false in 99-asr-mic.conf (AGC alone is usually the win).
• ASR/dictation apps should record from the default source (asr_mic).