Anything to do with finishing an audio file the user already has -- .wav, .flac, .aiff, .mp3. Reach for it when the ask sounds like "make this sound finished", "this is too quiet for YouTube", "get rid of the harsh S sounds", "the room sounds...
Full description and source
Anything to do with finishing an audio file the user already has -- .wav, .flac, .aiff, .mp3. Reach for it when the ask sounds like "make this sound finished", "this is too quiet for YouTube", "get rid of the harsh S sounds", "the room sounds boxy", "make this podcast match last week's episode", "level these three files to each other", "tighten the low end", "hit -14 LUFS without clipping", "trim the silences", "get rid of the umms and ehms", or "tighten up this recording". It measures what is actually there (loudness, true peak, spectral balance, crest factor, sibilance, ambience), FINDS things in it (onsets, silent spans, filler words with their timings), then runs an EDITING and MASTERING chain over it: cut listed regions, strip or shorten silences, de-ess, de-verb, parametric EQ, EQ-match against a reference recording, MULTIBAND compression and multiband dynamic range control, saturation, controlled ambience, loudness targeting and true-peak brickwall limiting. Every CHAIN STAGE (cut, strip-silence, gate, expand, deess, dereverb, eq, eq-match, compress, saturate, reverb, stretch, pitch, loudness, limit, downmix, resample) appends to a plan, and one render applies the whole chain in a single pass, so a chain can be inspected and re-run, and the whole job is ONE shell command rather than a round trip per stage. The read-only/reporting verbs (analyze, detect, verify, check, config, manifest) and the plan-lifecycle verbs (plan, preset) do not append to a plan themselves -- see the verb table below for which is which. This is mastering and cleanup, NOT mixing: it works on a finished stereo or mono programme, not on multitrack stems. It uses speech recognition internally to locate filler words, but it does NOT produce transcripts -- do NOT use it for video files, for transcription as an output, or for music generation.
Manifest name: aud
- Measure a file honestly before touching it -- LUFS, true peak, crest factor, spectral balance
- Bring a podcast or music file to a loudness target without clipping or inter-sample peaks
- Trim the silences out of a spoken-word recording, or shorten them to a fixed length
- Trim the pauses without chopping the start of a word -- edits are padded, snapped to a safe cut point and crossfaded, rather than falling wherever the detector put the boundary
- Get rid of the "umms" and "ehms" and long hesitations in an interview or a voice-over
- Tighten up a rambling recording -- find the pauses and fillers, cut them, master the result
- Find where the onsets are in a recording, to decide where an edit should land
- Tame harsh sibilance on a vocal or spoken-word recording
- Reduce a boxy or reverberant room on a recording made in the wrong space
- Make one recording sit in the same tonal balance as a reference recording
- Extract the EQ curve of a recording you like and apply it to another file
- Even out dynamics per frequency band with multiband compression, not one blunt full-band squeeze
- Retime or re-pitch a programme without changing the other
- Verify that a finished render actually meets the loudness and ceiling it was asked for
- Collapse a stereo file down to mono for a mono-only destination
- Deliver a file at a target sample rate, e.g. 48000 Hz, or downsample for a speech model
Snapshot refreshed: 2026-10-25T03:34:20Z
Source commit: 5e26cba11e92fb6329b7d8fe0566b938bdf8f92c
Ref: main / Path: .
