[audio]

Audio is captured through CPAL (PipeWire or PulseAudio on Linux). v1 requires 16 kHz mono output for the ASR pipeline regardless of the input device format.

[audio]
backend = "cpal"
device = "default"
target_sample_rate = 16000
channels = 1
max_record_seconds = 300

[audio.gates]
min_record_ms = 350
min_rms_energy = 0.003
min_peak_energy = 0.015
notify_on_no_speech = true

[audio] options

Option	Type	Default	Description
backend	string	"cpal"	Audio capture backend. Only cpal is supported in v1.
device	string	"default"	Input device name or "default" for the system default microphone.
target_sample_rate	integer	16000	Sample rate written to WAV files and sent to Whisper. Must be 16000 in v1.
channels	integer	1	Output channel count after downmix. Must be 1 in v1.
max_record_seconds	integer	300	Safety cap on recording length for a single job (5 minutes by default).

[audio.gates] options

Gates run when a recording stops. Failed gates produce a no_speech error instead of running ASR.

Option	Type	Default	Description
min_record_ms	integer	350	Minimum recording duration in milliseconds. Shorter clips are rejected.
min_rms_energy	float	0.003	Minimum RMS energy across the clip. Lower values accept quieter speech.
min_peak_energy	float	0.015	Minimum peak sample energy. Helps reject near-silent recordings.
notify_on_no_speech	boolean	true	When [notifications].enabled is true, show a desktop notification if gates reject a clip.

Tuning gates

• Increase min_rms_energy or min_peak_energy in noisy environments to reduce false triggers.
• Decrease them if legitimate speech is rejected (skald doctor and failed toggles with no_speech).
• Run skald calibrate mic to measure ambient microphone noise and print recommended gate values.
• Run skald calibrate mic --apply only when you want Skald to write the recommended gate values to config.
• Setup wizard recording uses the same gates as normal dictation.

Recalibrate after changing microphones, input gain, room noise, desk position, or audio backend settings. Calibration records a short ambient sample without starting a dictation job and does not write captured audio to disk.

Notes

• Stereo input is mixed to mono before resampling.
• Preview uses a separate RMS threshold in [preview].min_rms_energy for chunk gating.