Specula user guide
Every task, view, setting and command-line flag, on one page.
The Silero speech detection model ships inside Specula, so speech-gated analysis works offline, with no download and no network connection.
The window has four parts:

From the left:
The selection fields and the zoom show only with a file, and without one the transport is greyed out. Everything sits on one row where it fits. In a narrower window the groups wrap onto further rows in the same order, the transport keeping the first row. A group drops its button names only when its own row is too narrow for them, and the output device drops its name first. Settings › General › Top bar › Icons only shows the task and view buttons as icons. The window stops shrinking where every bar has wrapped as far as it goes.
Stop stops playback and returns the playhead to the start of the file. Loop repeats the selection when there is one, else the whole file. With a selection, Play plays the selection: a playhead before it starts at its start, and without Loop playback stops at its end.
Hold play start changes what Pause does. Off (the default), Pause stops where it is. On, Pause returns the playhead to where Play last started, so Space plays the same passage again. In Compare, the offset nudges go back to that point too.
The seek bar under the top bar spans the whole file. Click or drag along it to place the playhead. It does not move the playhead while a take records. Its playback point is an HDR accent while playing (see HDR accents).
The canvas holds the task's views: lanes along the file's timeline (waveform, loudness, spectrogram, stereo width, stems, the chapter map) and panels (the spectrum and phase panel, Compare's matrix, the chapter table, Record's meters). The ruler sits over the first lane, and one playhead and one selection run across every lane. Drag the line between two views to share their space; double-click it to reset. See The views.
The inspector reads in four parts, top to bottom:
With no file the inspector says No file loaded.
| Task | Key | Colour | What it is for |
|---|---|---|---|
| Inspect | ⌘1 | white | Reading the file against its delivery |
| Edit | ⌘2 | amber | Editing the audio, plugins, Save and Save As |
| Compare | ⌘3 | sky blue | Up to six files side by side, residuals |
| Chapters | ⌘4 | green | Chapters, their readings and ACX checks |
| Speech | ⌘5 | rose | Correcting the speech regions, Level Dialogue |
| Spectral | ⌘6 | violet | Selecting and repairing in time and frequency |
| Record | ⌘7 | red | Recording a new file or into the loaded one |
Choose a task in the switcher, with its key or from the View menu. The task's key again, or Esc with nothing selected, returns to Inspect. Pressing the task in front in the switcher does nothing.
With no file open, the canvas shows:

With no file, Tab reaches the tiles and the question's buttons, and Space presses them.
Esc does one thing per press, in this order:
Esc never ends a recording in progress. In the Signal Generator, Settings and report windows, Esc does not touch the main window.
Every lane follows the same rules:
| Gesture | Action |
|---|---|
| Click | Place the playhead. A click outside the selection also clears it; a click inside keeps it. |
| Drag | Select a range |
| ⌘-drag | Select a range |
| ⌃-drag | Move the playhead (scrub), also over a selection, a speech region or a chapter boundary |
| Drag a selection's edge or body | Resize or move the selection |
| Click or drag in the ruler | Place or move the playhead |
| Pinch, or scroll up and down | Zoom time around the pointer |
| Swipe sideways, or ⇧-scroll | Pan |
| ⌥-scroll | Zoom the lane's own axis: amplitude in the waveform, the LUFS range in the loudness view, frequency in the spectrogram |
Each view's strip ends with a ? button that lists the view's gestures, worded for the current Selection gesture setting.
A selection is a range of time. It drives analysis (⌘Return), the Edit operations, Loop and the export of a selection's readings.
IN, OUT and LEN, the Compare slot card's offset and trim fields, and the loudness view's M, ST and TP fields take the keyboard only when you click them or reach them with Tab. Return commits a typed value; Esc in IN, OUT or LEN puts the value back. Number fields take a point or a comma as the decimal mark. A value that is not a number, such as a frequency typed as 1,000, is refused with the reason. Specula prints numbers with a point for decimals and no thousands separator, whatever the system's region.
On macOS 26 and later, on a display that shows HDR, a few things that come and go light brighter than white: the playback point on the seek bar while playing, the record light while a take records (the record button and the dot beside the Record meters' title), a clip on the Record input bars, a hot peak on the Channels meters, an over on the live meter, the brightest spots of the live phase scope, and the outline of a Spectral selection. They light only while Specula is the app in front, and on a display such as Studio Display only below full brightness. Otherwise they keep their usual colours.
⌘O (File › Open…) opens the file picker; you can pick several files. Other ways in:
open path/to/file.wav in Terminal.Opening a file always starts a clean session: any compare set is replaced. Open several files at once and they fill the compare slots in the Finder's file-name order ("Take 2" before "Take 10"): the first becomes slot A and the window switches to Compare. To add files to the current set, use File › Add Files to Compare… (⇧⌘O). Opening a file, from any path, asks first when the session holds unsaved edits, a take that is recording, or a recorded take not yet saved.
While a file opens, the window shows its name and how far the decode has got, with Cancel; Esc cancels too. The file in front stays as it was until the new one has decoded, so a cancelled open leaves it untouched, with its edits, its speech regions and its plugin chain. Nothing plays while a file opens.
Specula keeps the decoded audio in memory. When the file has loaded, it measures the whole file in the background and fills the inspector; see Readings.
A decoder under test emits raw samples, a logic analyser dumps an I2S capture, a test harness writes a bare array of floats. Nothing in such a file says how to read it. When a file cannot be read as a known audio format (any extension, through every open path above), Specula asks for its format instead of refusing it. Declare it once and the file works like any other: measure it, compare it, edit it, report on it.

The import sheet declares:
Rank likely formats decodes a few seconds under every byte format that fits and scores how much each result behaves like audio; the top candidates apply with a click. Ranking covers the byte formats only; the rate is yours to declare. The preview draws the decoded waveform, shows the same score for the current declaration and plays a few seconds. A wrong declaration is easy to spot: the wrong byte order is full-scale noise, and a wrong channel count combs the waveform.
After a raw file opens:
--raw-preset).Three services appear under Services when you right-click one or more audio files in the Finder:
If the entries do not appear after the first install, log out and back in.
Specula judges every file against one delivery lens: a delivery category and its checked targets, one of which is the primary target the verdict leads with.
The Deliver to chip names the primary target and the category, such as Deliver to Spotify · Music streaming. It sits in the inspector over the delivery card, and in the top bar while the inspector is hidden. Click it, or choose Analyze › Deliver To… (⇧⌘D), to open the picker:

Audiobook is Podcast with ACX leading: under Audiobook, ACX stays checked and primary (Audiobook leads with ACX. Choose Podcast to lead with another target.). Under Podcast, ACX cannot lead (Choose Audiobook to lead with ACX.), and unchecking the primary target passes the lead to another checked target, never to ACX.
| Category | Target | Measured on | Reference | Tolerance | True-peak ceiling | Checked at first |
|---|---|---|---|---|---|---|
| Music streaming | Spotify | Integrated | -14 LUFS | Platform gain, turns up and down | -1 dBTP | ✓ |
| Music streaming | Apple Music | Integrated | -16 LUFS | Platform gain, turns down only | -1 dBTP | ✓ |
| Music streaming | YouTube | Integrated | -14 LUFS | Platform gain, turns down only | -1 dBTP | ✓ |
| Music streaming | Tidal | Integrated | -14 LUFS | Platform gain, turns down only | -1 dBTP | ✓ |
| Music streaming | Amazon Music | Integrated | -14 LUFS | Platform gain, turns up and down | -2 dBTP | |
| Music streaming | Deezer | Integrated | -15 LUFS | Platform gain, turns up and down | -1 dBTP | |
| Music streaming | SoundCloud | Integrated | -14 LUFS | Platform gain, turns up and down | -1 dBTP | |
| Music streaming | AES TD1008 (streaming) | Integrated | -18 LUFS | Platform gain, turns up and down | -1 dBTP | |
| Podcast, Audiobook | Apple Podcasts | Dialogue | -16 LKFS | Platform gain, turns down only | -1 dBTP | ✓ Podcast |
| Podcast, Audiobook | Spotify (Podcast) | Dialogue | -14 LKFS | Platform gain, turns up and down | -1 dBTP | ✓ Podcast |
| Podcast, Audiobook | ACX (Audiobook) | RMS | between -23 and -18 dB RMS | noise floor less than -60 dB RMS | less than -3 dBTP | ✓ Audiobook, primary |
| Film and TV | Netflix | Dialogue | -27 LKFS | ±2 LU | -2 dBTP | ✓ |
| Film and TV | Prime Video | Dialogue | -27 LKFS | ±2 LU | -2 dBTP | ✓ |
| Film and TV | Apple TV+ | Dialogue | -27 LKFS | ±2 LU | -2 dBTP | ✓ |
| Film and TV | Disney+ | Dialogue | -27 LKFS | ±2 LU | -2 dBTP | |
| Film and TV | Max | Dialogue | -27 LKFS | ±2 LU | -2 dBTP | ✓ |
| Broadcast | EBU R128 (EU) | Integrated | -23 LUFS | ±0.5 LU | -1 dBTP | ✓ |
| Broadcast | EBU R128 (EU, live) | Integrated | -23 LUFS | ±1 LU | -1 dBTP | |
| Broadcast | EBU R128 S1 (EU, short-form) | Integrated | -23 LUFS | ±0.2 LU, max short-term at most -18 LUFS | -1 dBTP | |
| Broadcast | ATSC A/85 (CALM Act) | Dialogue | -24 LKFS | ±2 LU | -2 dBTP | ✓ |
| Broadcast | ARIB TR-B32 (Japan) | Integrated | -24 LUFS | ±1 LU | -1 dBTP | |
| Broadcast | OP-59 (Australia) | Integrated | -24 LUFS | ±1 LU | -1 dBTP |
Two kinds of target:
What each target measures:
A file with no detected speech has no dialogue reading, and a dialogue-gated target reads No reading for Netflix: no speech was found.
ACX asks for a volume between -23 and -18 dB RMS, peaks less than -3 dB and a noise floor less than -60 dB RMS, and Specula judges and words them so, on the delivery card, in the chapter table, in the room check and in the reports. A true peak of exactly -3 dBTP, or a noise floor of exactly -60 dB RMS, fails. Limit peaks for ACX limits to -3.1 dBTP, and Level chapters starts there. Every other target's ceiling is a maximum that the limit itself meets.
Under the chip, the delivery card judges the whole file (in Compare, the slot in front):

While the file is measured the card says Measuring the file…. After a change it shows the previous numbers dimmed with Measuring the file again. The fixes wait for it. With nothing checked in the category it says No target is checked for <category>. Choose one with Deliver to.
A reading is coloured by what the primary target does with it:
The inspector, the per-channel statistics window and the PDF and HTML reports use the same colours.
A fix appears only where it changes something:
The primary target's fixes are buttons; every other target's are in its Fix menu. A file within a target's tolerance, or a quiet master that a turn-down-only platform plays as recorded, gets no level fix. Every fix works from the whole file's readings. With more than one file in Compare, only Hear is offered.
The Normalize panel opened from a target shows its Basis (Integrated or Dialog-gated), Now (the whole file's reading), Target and Gain, and an editable Ceiling filled in with the target's ceiling as set. Apply lands the reading on the target with a true-peak limiter at the ceiling, as one undoable edit.
When the file fits another category better, a banner under the chip says so and offers the switch, such as This file is mostly speech. Switch to Podcast. The rules:

Mostly speech means speech in 60% or more of the file's non-silent blocks. The lens changes only when you press the switch. The banner's close button hides the suggestion for that file until Specula quits; the picker still shows it.
Under Music streaming, the delivery card says whether a stereo file folds to mono. It reads left against right as the Stereo width view does, over the whole file and over every second of it (blocks below -70 dBFS do not count). A file passes when its correlation stays above -0.2:

The codec check shows what Apple Music's AAC encode does to the master. Under Music streaming the card's CODEC CHECK row offers Check Apple Music's AAC encode; Analyze › Check the Apple Music Encode does the same under any category. Specula encodes the file as AAC with the encode Apple's afconvert makes (256 kbps, constrained VBR, quality 127; 128 kbps for mono), decodes it and measures the encode's true peak and its samples over full scale. Files above 48 kHz are resampled to 44.1 or 48 kHz first. The check takes mono and stereo files at 44.1 kHz and above.

The row then states the result, such as After Apple Music's AAC encode the true peak is +0.4 dBTP, with 12 samples over full scale., coloured against the primary target's true-peak ceiling, and offers:
After an edit the result reads dimmed with Checked before the last edit. and Check again. The export sheet's Codec check switch puts the check in the reports.
Settings › Delivery targets › Add Custom Target… defines your own delivery spec: a Name, a Mode (the list it joins: Music / Streaming, Podcast / Spoken Word, VOD or Broadcast; a Podcast / Spoken Word target appears under both Podcast and Audiobook), what to Measure (Integrated (full program), Dialog-gated (speech only) or RMS (un-weighted, ACX-style)), a Reference, a Tolerance and a True-peak ceiling. Fields take up to two decimals, and a target's levels print as you set them. A custom target is a compliance target: it passes when the reading sits within the tolerance and the true peak stays at or under the ceiling. It appears in its category's list in the picker and in Settings, and works everywhere a built-in target does: the verdict, the fixes and the reports.

As soon as a file has loaded, Specula measures the whole file in the background and fills the inspector: the delivery card, Loudness details, Speech, Music and Channels. These readings hold still while the file plays. They are measured again after every edit, Undo and Redo, after a speech region is corrected, when a file with saved speech corrections opens, and when the speech detection settings change, so the verdicts, the fixes and the reports always describe the audio and the speech regions in front. While a measurement runs the readings say Measuring the file….
The Loudness details group has a Whole file | Live switch, which starts on Whole file.
Whole file lists:
| Row | Meaning |
|---|---|
| Integrated (I) | Integrated loudness, LUFS |
| Loudness Range (LRA) | LU |
| Max Momentary (M↑) | The loudest 400 ms window, LUFS |
| Max Short-term (S↑) | The loudest 3 s window, LUFS |
| True Peak | dBTP, 4× oversampled |
| Sample Peak | The highest sample, dBFS |
| Headroom | 0 dBTP minus the true peak; negative when the true peak is above 0 dBTP (Music streaming and Broadcast) |
| Dynamic Range | DR (Music streaming and Broadcast) |
Live shows what is playing: momentary and short-term loudness, and the integrated loudness, loudness range and true peak of what has played since playback was stopped, the file was edited or another compare slot was chosen, with short bars. The target verdicts judge the whole file either way. Settings › General › Loudness readout sets how often the live numbers repaint.
Specula implements ITU-R BS.1770-4 and EBU R128:
Specula finds the speech in a file with two detectors:
From the speech blocks Specula measures:
Correct the speech regions by hand in the Speech task. Corrected regions feed every speech-gated reading, the verdicts, the fixes and the reports.
The Noise Floor is the 90th percentile of the 100 ms RMS of the mono mix over the non-speech blocks, in dB RMS. When the pauses are digital zeros it reads digital silence. ACX judges it; it shows for Podcast and Audiobook and for any target that checks the noise floor.
The PRE | POST switch beside the Channels meters chooses what they show: PRE the file's channels, POST the routed output (see Output and routing). The group names which it shows: File channels or Routed output. The same switch sits in the Output and Routing window.
⌘Return (Analyze › Analyse Selection, or the top bar's Analyse button) measures the selection on its own: loudness, true peak, speech detection over the selection, per-channel levels and an averaged spectrum. Without a selection it selects the whole file first.
The results land in a Selection card under the delivery card: its time range (0:12.000 to 0:20.000), Integrated, Dialogue (where speech applies), True Peak, Loudness Range and Max Short-term, and Per channel on request. The ✕ clears the card. The spectrum panel can show the selection's averaged spectrum while that range stays selected. The target verdicts keep judging the whole file.
Each task shows its own views and remembers them:
| Task | Opens on | Can also show |
|---|---|---|
| Inspect | Waveform, Loudness, Spectrogram | Spectrum, Stereo, Chapter map, Stems |
| Edit | Waveform, Loudness | Spectrogram, Spectrum, Stereo, Chapter map, Stems |
| Speech | Waveform, Loudness | Spectrogram, Spectrum, Stereo, Chapter map, Stems |
| Compare | Matrix, Waveform | Loudness, Spectrogram, Spectrum, Stereo, Chapter map, Stems |
| Chapters | Chapter map, Chapter table | Waveform, Loudness, Spectrogram, Spectrum, Stereo, Stems |
| Spectral | Spectral canvas | |
| Record | Meters, Last take | Waveform, Loudness |
The view buttons in the top bar show and hide views: the shared views in the order of their keys, then the task's own. A shown view is lit and a hidden one dimmed, and each tooltip names its key. ⌃1 to ⌃7 do the same from the keyboard (View menu):
| Key | View |
|---|---|
| ⌃1 | Waveform |
| ⌃2 | Loudness |
| ⌃3 | Spectrogram |
| ⌃4 | Spectrum and phase |
| ⌃5 | Stereo |
| ⌃6 | Chapter map |
| ⌃7 | Stems (macOS 27) |
Every view can be hidden but the task's last one. Maximize (the icon after the view buttons, or ⌃⌥1 to ⌃⌥7) lets one view fill the canvas; the icon turns into Restore, and the same key again or a task switch restores the task's views. Zoom and scroll stay the same when the task changes, and a new file resets them.
All channels of the file, each in its layout's colour. Clipped samples (at or above full scale) show as red bars, and dBFS reference lines adapt to the amplitude zoom. Hovering reads each channel's level at the pointer.
The strip under the waveform holds:
The time ruler runs over the first lane. A click or a drag in it places or moves the playhead. The label at its end names the unit it shows, Time, Samples or Bars, and opens a menu of the three with the current one checked; View › Ruler Unit holds the same three. Bars stay in the list, greyed out with the reason, until the music analysis has found a beat grid. IN, OUT, LEN and the waveform's hover readout follow the unit; with Bars they stay in time.
Once the music analysis has found a song, a band under the ruler shows its sections (S1, S2 and on), each with Specula's integrated loudness for that section and shaded from darker (quieter) to lighter (louder). Click a section to select it; Space then plays it and ⌘Return analyses it. See Songs.
The loudness view plots the momentary, short-term and integrated loudness over time, aligned with the other lanes, for the whole file.

The strip under it holds:
In a narrow window the thresholds move to a second row.
The spectrogram shows the whole file. It is computed when the view is shown, with its progress in the lane (Computing spectrogram…), clears at once when another file opens, and stays as it is while the file plays. Each column sits at the centre of the audio it analyses, so a transient lines up with the waveform and the playhead. On a multichannel file the lane names the channel it shows: one channel when only that channel is on in the spectrum panel's channel toggles, else the mix of all. Hover for time, frequency and level; ⌥-scroll zooms the frequency axis.
The strip under it holds two knobs, Floor (the level drawn black, -120 to -24 dB, default -96) and Ceiling (the level drawn brightest, -18 to 0 dB, default 0), and the gesture legend. Drag a knob up or down to turn it, with ⌥ for fine steps, and double-click it for its default. Spectral's canvas uses the same range.
Its look is set in Settings › Analysis:
| Overlap | Hop (4096 at 48 kHz) | Columns per second |
|---|---|---|
| 75% (default) | 21.3 ms | about 47 |
| 87.5% | 10.7 ms | about 94 |
| 93.75% | 5.3 ms | about 188 |
| 96.875% | 2.7 ms | about 375 |
| 98.4375% | 1.3 ms | about 750 |
The spectrum runs from 20 Hz to 20 kHz on a log frequency axis. While the file plays it is live. At rest it shows:
The label is a menu for switching between them. Hover for frequency, level and the nearest note with its cents (442.6 Hz · -18.2 dB · A4 +10 ¢ with NOTES on). Scroll zooms the frequency axis; double-click, or the full-range button, returns to the whole range.
The strip holds the channel toggles, the FFT size (2k to 32k), M/S, PEAKS with the number of peaks to label, NOTES, the band filter and the gesture legend.
The band filter. A drag on the spectrum (⌘-drag under the classic Selection gesture) selects a frequency band, and the band filters playback: BAND passes only the band, NOTCH cuts it. Drag the band's edges or body to adjust it while listening; ✕ or Esc removes it.
Window function (Settings › Analysis › FFT window function):
| Window | Side lobes | Use |
|---|---|---|
| Hann | -31 dB | General purpose |
| Hamming | slightly higher | Narrower main lobe than Hann, higher far side lobes |
| Blackman | -58 dB | Strong side-lobe rejection, slightly wider main lobe |
| Blackman-Harris | -92 dB | Best side-lobe rejection, widest main lobe; closely spaced harmonics |
| Flat Top | very wide main lobe | Most accurate amplitude; calibration and level measurement |
The averaged spectrum (the whole file's and a selection's) samples up to 65,536 FFT windows spread across the range, statistically equivalent to averaging every window. Settings › Analysis › Average every window (exact) accumulates every window instead, which can take minutes on hours of audio.
The phase panel plots left against right (L on the x axis, R on the y axis). At rest it shows where the whole file's samples fall, with their correlation; while playing it is a live scope, whose brightest spots are an HDR accent.
The correlation between left and right over time (the zero-lag normalized product, as a correlation meter reads it). The strip's Correlation | Width % switch chooses the scale:
ZONES colours the curve: Mono (0.95 and above), Narrow (0.5 to 0.95), Stereo (0 to 0.5), Wide (-0.5 to 0) and Out of phase (below -0.5). Hover for the time, correlation, width and zone.
The stems view (⌃7, macOS 27) holds Apple's music analysis of the file: the song's structure, the beat grid, the pace and four stem lanes. See Songs.
The chapters along the timeline, each with its readings: RMS (the ACX loudness check) and integrated loudness with their deviation from the book's median, true peak and noise floor, then LRA, dialogue loudness and Speech %. A narrow chapter shows what fits, down to its number alone, in red when it fails ACX or stands out from the median; hover over a chapter for all its readings, or zoom in. Click a chapter to select it and make its time range the selection, in every task that shows the map; Space then plays it, and the view stays where it is. Double-click a chapter's name to rename it, and double-click elsewhere in it to move the playhead to its start. See Chapters.
A view with nothing to show says why, such as Mono file: no stereo width, Digital silence: no loudness to draw or Shorter than one FFT window.
⌘1, or the Inspect button. Inspect is the home task: the file and its delivery.
A typical pass: open the file, read the delivery card, play the passages the loudness view or the thresholds point at, select one and press ⌘Return for its own readings, and export the report.
⌘2, or the Edit button. Entering Edit pauses playback.
Edit, Chapters, Spectral and Record share one editing session and one edit history: the edited audio, the history and the unsaved state stay with you across them, and one Save at the end writes everything. Changing tasks never discards edits; opening another file or File › Revert to Saved… does. A dot in the task bar marks unsaved edits.
| Button | What it does |
|---|---|
| Cut | Removes the selection and joins the audio on either side, with an equal-power crossfade at the join (10 ms to start; 0 for a hard cut; clamped to the audio kept on each side). Not when the selection is the whole file. |
| Trim | Keeps the selection and discards the rest. |
| Fade In, Fade Out | Open the Fade panel for the selection: the direction and the curve (Linear, Log or Equal Power). |
| Gain | Changes the level of the selection, or the whole file without one, by a typed amount in dB (quick values -6, -3, -1, +1, +3, +6). |
| Loudness | Normalize…, Limit True Peak… and Level Dialogue…. |
| More | Silence…, Invert Phase, Remove DC Offset, Swap Left and Right (stereo files) and Split to Mono Files… (two or more channels). |
| Plugins | The plugin drawer. |
| Save, Save As… | See Save and Save As. |
Cut, Trim and the fades need a selection; a disabled button's tooltip says why. Every typed value takes a point or a comma, and anything else is refused with the reason.
The operations in detail:
Over the waveform, a selection shows its length (the LEN field's text) with Cut, Trim, Fade in, Fade out, Gain -1 dB, Gain +1 dB and Analyse. It sits clear of the selection's edges, so a drag on an edge still reaches the waveform. The HUD and menu fades use the equal-power curve, and the HUD and menu Cut a 10 ms crossfade.
The inspector's History lists every edit of the session by name, up to 16 steps back. Click a step to go back or forward to it in one move; every step brings back the audio whole, to its last sample. The step that matches the file on disk says On disk, and going back to it leaves nothing unsaved. Saving starts the history again. ⌘Z and ⇧⌘Z step one at a time. The history is kept on disk, one temporary file per step, so deep history does not multiply memory. When a step cannot be written, Specula asks before applying the edit without Undo.

The Plugins button opens a strip of Audio Unit effects that process the audio in Edit, for mono and stereo files. Closed, it shows how many plugins the chain holds and how many play, and while a render runs its progress with Cancel.

The Edit menu reaches every drawer action: Show Plugin Chain, Add Plugin…, Restore Last Plugin Chain, Render Plugin Chain…, Cancel Plugin Render, and for the selected plugin Select Next Plugin, Bypass Selected Plugin, Move Selected Plugin Earlier or Later, Show Selected Plugin's Controls and Remove Selected Plugin.
Save As… (⇧⌘S) writes the audio to a new file and refuses any name that opens the source file, under other capitals or through a hard link. It writes the source's bit depth and sample rate (32-bit float for a source with no fixed PCM bit depth: compressed audio, raw PCM, a generated signal), in the container the file name's extension names. When an integer file would clip samples above full scale, Save As asks first and offers 32-bit float, which keeps every sample. The saved file becomes the working document: chapters, corrected speech regions and the playhead stay, and you stay in the task you were in.
Save (⌘S) writes over the loaded file once Settings › Editing and plugins › Allow Save to overwrite the original file is on (off to start); without it, ⌘S opens Save As. Save keeps the file's own format: bit depth, sample rate and channel layout, and the container read from the file's header, so a WAV named .WAV or .bwf stays a WAV. The audio is written to a temporary file next to the original and swapped in once complete, so a failed save leaves the original untouched. Specula asks before each overwrite; tick Don't ask again, or turn off Ask before overwriting, to stop it asking. Save works for WAV (including Broadcast Wave, RF64, BW64 and Wave64), AIFF and CAF. A compressed file (MP3, AAC, FLAC), a raw PCM file or a recorded take still in the takes folder goes through Save As; Save's tooltip says why. Once written, the original is gone: the history and Revert to Saved reach back only to that save.
Neither save carries metadata chunks over (BWF bext, iXML, cue points, markers). A multichannel file stored in a non-SMPTE channel order is written back in its own order.
Quitting or opening over unsaved work asks first: save (over the file where Save is allowed for it, otherwise as a new file), discard, or cancel. A take still recording, or a recorded take not yet saved out of the takes folder, is asked about too.
⌘3, or the Compare button. Load up to six files as slots A to F, switch between them while playing, match their loudness and listen to the difference between A and any other slot.
Adding files. File › Add Files to Compare… (⇧⌘O), or the task bar's Add. Adding a second file from Inspect or Record switches to Compare. Several files added together fill the slots in the Finder's file-name order. Each slot is measured when it loads.
Switching. Press 1 to 6, click a slot's chip, or click Listen over its column in the matrix; playback carries on from the same position. A file at another sample rate is resampled to slot A's rate once, when it is added: switching to it is instant, and its residual and Align work. Its readings and reports describe the file itself, at its own rate, and the inspector reads, for example, 44.1 kHz, plays at 48.0 kHz. Left alone after the others are removed, it loads again at its own rate. Switching to a slot with another channel count rebuilds the audio graph, with a short pause.
The Compare task opens on the metrics matrix over the waveform. Each slot has a column: its letter and name, a Listen (n) button naming its key, and its readings: Integrated, True peak, Sample peak, LRA, DR, Max momentary, Max short-term, Duration, and Dialogue when a slot holds speech or the primary target judges it. Each value shows its difference from A, worked out from the printed values (= A when both print the same). A Residual (dBFS) row shows the peak and RMS of A minus each slot once a residual exists.

The inspector's Slot card holds the slot in front:
Slot A's card says it is the reference: the other slots align to it and their residuals subtract from it.
The residual is A minus the slot in front, at the slot's offset, trim, polarity and level match. With level match on, two correct masters of the same source null close to silence; anything above silence is what differs. It is computed again whenever the slot's offset, gain, polarity or level match changes, also when the change comes while it is still being computed.

While a residual plays, the inspector shows its own Peak and RMS (over the stretch where both files have audio) in place of the delivery card, with only the Channels group below, and The two files null: their difference is digital silence. when they do. A residual gets no loudness reading or verdict. Press D again to hear the slot. File › Export Residual Audio… (⌃⌘E), the task bar's Export Residual… and the export sheet's Residual audio write the residual in front, and are unavailable when it has none.
Moving a file into slot A, or removing slot A, rebases every offset onto the new A, computes the residuals again and names the new reference: <name> is the reference now. Alignments follow it.
⌘4, or the Chapters button. Chapters splits a long file, such as an audiobook or a long-form podcast, at its silences and measures every chapter against ACX and against the book's own median.
First entry. Opened on a file without chapters, Chapters detects and measures them, once per load. It offers the chapters saved for this file in an earlier session instead, when there are some. For a song (the music analysis found a key), it offers Chapters from Sections or Detect from Silences.
Views: the chapter map over the chapter table, plus any lane on request.
| Button | What it does |
|---|---|
| Detect (⌘D) | Splits the file at its silences. Asks before replacing chapters you named or moved. |
| From Music (macOS 27) | One chapter per song section, from the music analysis. |
| Add (⌘B) | Adds a boundary at the playhead. |
| Remove (⌫) | Removes the selected chapter; its range joins the chapter before it. |
| Analyse (⌘L) | Measures every chapter. |
| Detection | The detection settings. |
| Recall | Restores the chapters saved for this file in an earlier session (shown when there are some). |
| Import…, Export… | Read or write the chapters (boundaries and names) as JSON. |
| Export CSV… | The chapter table as CSV, through the export sheet. |
| Save, Save As… | As in Edit. |
Detection holds Silence below (-90 to -30 dB, -55 to start) and For at least (0.5 to 10 s, 2 s to start), used at the next Detect; Compare by (RMS or Integrated LUFS, RMS to start) and Flag beyond (0.5 to 5, 2.0 to start), which apply at once; and Offer chapters saved in earlier sessions.

How detection works. Each silence below the threshold and longer than the minimum becomes a boundary, and the audio between two silences a chapter. A chapter with under 1 s of non-silent content is dropped. The first chapter starts at 0 and the last ends at the end of the file.
Boundaries. Chapter boundaries are drawn in every lane. Drag a boundary in the waveform to move it. Click a boundary to select the chapter that starts there; Remove then deletes that boundary. After Remove nothing is selected.
One row per chapter: #, Name, Time range and Duration, then three groups of columns:
Click a row to select the chapter, make its range the time selection and fit it in the lanes. Rename a chapter in the table, or by double-clicking its name in the chapter map. A value just past a limit prints with the decimals that show which side it is on, such as -17.96, not -18.0.
What each chapter measures. Every reading covers all the file's channels at the same BS.1770 channel weights as the whole file's, so a chapter's integrated loudness and true peak are what you would get from that span on its own.
After a change. An edit, Undo, Redo or a speech-region change outdates the readings. The chapters are measured again once the file's own measurement lands; meanwhile the map and the table say These readings are from before the last change. or Measuring the chapters…, show no pass, fail or outlier, Level chapters waits, and Export measures them first.
The inspector's ACX card counts the table's own marks: All 20 chapters pass ACX. or 3 of 20 chapters fail ACX., then one line per failing check (RMS: 2 chapters). Under Music streaming, Film and TV or Broadcast it reads Not judged against ACX under Music streaming.
Level chapters gives every chapter a uniform gain:
A uniform gain moves a chapter's room tone with its level. For a pass that also keeps the floor down, duck the room tone first with Level Dialogue, then level.
Chapters, with their names, are saved per file in ~/Library/Application Support/Specula/chapters/, matched by the file's name, length, sample rate and channel count; chapters of pure digital silence keep their names too. An automatic detection never replaces a saved set holding your work. Chapter JSON writes a silent chapter's readings as "-inf", as the JSON report does. Export CSV… writes the chapter table with every reading, sample peak and max momentary included.
The Chapters menu holds Analyse Chapter Loudness, Add Boundary at Playhead, Remove Selected Chapter, Level Chapters, Chapters from Music Sections, Chapters from Music Segments (a finer split), Recall Saved Chapters, Chapter Detection…, Import Chapters…, Export Chapters… and Export Chapters CSV….
⌘5, or the Speech button. Speech corrects the speech regions the detection found. The regions drive the dialogue loudness, Speech %, the noise floor, the verdicts, the fixes, Level Dialogue and the reports.
Views: the waveform and the loudness view to start with. The regions are always drawn on the waveform, tinted by where they came from: teal detected, amber corrected by hand, blue imported.
On the waveform: click a region to select it, drag its body to move it, drag an edge to resize it. ⌃-drag moves the playhead over a region.
| Button | Key | What it does |
|---|---|---|
| Add Region / Add at Playhead | A | Makes the selection a region, or adds a 1 s region at the playhead without one. |
| In [ | I | With no region selected, marks the start of a new region at the playhead; with one selected, moves its start to the playhead. |
| Out ] | O | Makes a region from the In to the playhead; with a region selected, moves its end to the playhead. |
| Split | S | Splits the selected region at the playhead. |
| Delete | ⌫ | Deletes the selected region. |
| Reset to Detected | ⌘R | Discards your corrections and detects speech again. |
| Detection | The detection settings. | |
| Export…, Import… | Write the regions to a JSON file, or replace them with the ones in a file. |
The bar also shows a pending In (In at 0:12.340), the number of regions with a dot in their tint, and why saved regions were not applied.

⌘Z and ⇧⌘Z in Speech step through Speech's own changes, newest first: the region edits (up to 50 steps) and Level Dialogue's Apply. Edits made in other tasks stay with the edit history. Region undo survives leaving Speech and is cleared by opening a file or by an edit that moves the timeline.
Detection holds the three speech detection settings, which apply to every file:
Over detected regions, a change detects speech again, then measures the file again. Over corrected or imported regions, a change is kept for the next detection: Your corrected regions are kept. Detect Again to replace them with these settings. Detect Again… then asks Replace the corrected speech regions?; Undo in Speech brings them back. Analyze › Detect Speech Again runs the detection again only over detected regions; with corrected or imported regions it is unavailable, and Reset to Detected is the way.
Corrected and imported regions are saved next to the audio file as <audio>.dlg.json and apply when the file opens again, as long as they were made for a file of the same length (within 0.1 s). Regions made for another length are not applied, and the bar says so (Saved speech regions were made for a file 2.3 s longer. They were not applied.). Importing such regions asks first. Save As takes corrected regions to the new file. Saved regions made for another length are never replaced without asking Replace the saved speech regions? (Keep Saved Regions is the default), and Reset to Detected, region undo or a save never removes them. The command line, Shortcuts and the Finder services use saved regions made for the file's length too.
The inspector's Level dialogue panel:
Speech › Level Dialogue applies the panel's settings. Level Dialogue is an audio edit: it lands in the edit history, which Edit's History shows.
The Speech menu holds Add Region, Mark In at Playhead, Mark Out at Playhead, Split Region at Playhead, Delete Selected Region, Reset to Detected, Level Dialogue, Speech Detection…, Import Speech Regions… and Export Speech Regions….
⌘6, or the Spectral button. Spectral turns the canvas into a zoomable, full-resolution spectrogram of the file, where you select regions in time and frequency and process only those: a cough inside a piano chord, a chair squeak under speech, a bite of mains hum. It works on any channel count, and its edits share the edit history and the Save commands with Edit.

The tools sit in a column on the canvas's left edge, each with its key:
With the rectangle and lasso a drag starts a new selection, Shift adds to it and Option subtracts; a plain click clears it. A brush stroke adds and an Option-stroke paints away. A wand click replaces the selection, Shift adds and Option subtracts. Clear in the task bar, or Esc, clears the selection.
A selection has a white outline (an HDR accent) and keeps its own colours; the rest of the canvas is dimmed while it exists, and the feather's roll-off shows as a soft edge around it. Its time span and frequency range print beside it and in the inspector.
The active tool's options, Clear, the analysis size (1024 to 8192, 4096 to start; larger resolves frequency finer, smaller time), Feather, the Floor and Ceiling knobs, the channels shown, Wave, and Save and Save As…. Changing the analysis size with a selection asks first, since the selection belongs to the current size. Feather sets how far each operation rolls off into the audio around the selection, in time and in frequency, and Erase's depth. Spectral › Contrast… opens the Floor and Ceiling as sliders, for the keyboard.
The inspector's Spectral edit panel:

The operations change the level of the frequency bins inside the selection and leave their phase alone. Selections longer than 10 minutes are refused; select a shorter range and work in passes.
Playback in Spectral. With a selection, Space plays its time span. Placing the playhead by hand (the ruler, the waveform strip, the arrow keys, Go to Start) makes Space play from the playhead until the selection changes. One thing plays at a time: starting Solo or Hear Result pauses the file; Space stops an audition and starts nothing; Play and Stop end it too. While an audition plays, the Play button shows pause.
The Spectral menu holds the four tools, Bigger Brush and Smaller Brush, Reset Zoom, Fit Time to Window, Reset Frequency Zoom, Clear Selection, Include Harmonics, Attenuate, which applies the panel's gain and names it (Attenuate (-18 dB)), Erase to Floor (⌫), Repair, Play Selection Only, Hear the Result, Feather…, Contrast… and Show Waveform Strip. Erase to Floor and Repair also set the panel's operation.
⌘7, the Record button, or the Record audio tile with no file open. Record captures from any input device: a new file, a take appended to the loaded file, an insert at the playhead that pushes the rest later, or a punch that records over existing audio. It is the one task that works with no file open. The first time, macOS asks for microphone permission. Record checks the license when you enter it and again at every take.
Take goes to: New file, Append, Insert at playhead or Punch in; all but New file need a loaded file. Splice sets the equal-power crossfade at each join (0 to 100 ms, 10 ms to start; 0 is a hard splice). Save and Save As… as in Edit. The record button is in the top bar's transport, and R starts and stops a take.
The Meters view names the input (Input · <device>) and shows, under LIVE INPUT, each channel's level from -60 to 0 dBFS with its peak (a clip lights it as an HDR accent), the momentary loudness and the true peak. While the input is armed it adds the room floor (below). While a take records the title reads Recording · 0:12.3 with the record light, and the Last take view draws the take as it records, growing and scrolling, so a dead input shows within a second. When the input is not live, the view says why, such as No microphone access with the way to turn it on in System Settings › Privacy & Security › Microphone.

Specula records the signal as the device delivers it: no noise suppression, no automatic gain, no EQ. Set the recording level with the input volume in System Settings › Sound › Input.
Where a take lands is fixed when recording starts. Each landing is one edit: Undo (⌘Z, not while a take records) restores the audio exactly, and Redo brings the take back. After an insert or a punch the landed take becomes the selection, one Space away from listening back. Speech is detected again inside the punched stretch, also after Undo and Redo across a punch, and corrected regions elsewhere stay. Appending, or punching past the end, keeps the chapters and the corrected speech regions.
A take records at the input device's sample rate and is converted to the file's rate when it lands. A mono take goes to every channel of the file, and a stereo take folds to mono for a mono file; other channel mismatches are refused, so a 5.1 file takes a 5.1 take.
Takes are written as they record to ~/Library/Application Support/Specula/Recordings, as CAF files, so a take is bounded by free disk space, not memory. A take not saved elsewhere is kept for seven days and removed at the next launch after that. Save Selected Take As… (and the Takes card's Save As…) saves the take, as 32-bit float in WAV, AIFF or CAF. To Compare sends a take to a Compare slot: record a part twice and compare the takes level-matched, or listen to their residual.
The Record menu holds Start Recording, the four places a take goes, Splice Crossfade…, Select Next Take, Add Selected Take to Compare, Save Selected Take As… and Move Selected Take to Trash.
macOS 27 and later. Specula reads a song's structure with Apple's on-device music analysis. Nothing is uploaded: the analysis runs on the Mac in the background, at about 0.6 seconds per minute of audio, plus a few seconds the first time after launch while the model loads.
Settings › Analysis › Run music analysis:
Analyze › Analyse Music runs it on any file, and reads Cancel Music Analysis while it runs. The Stems view's Analyse button does the same.
A key found means the file is a song. Then:
When no key is found, nothing opens.
The Stems view (⌃7) holds the detail, aligned with the other lanes. Click in it to place the playhead.

The header reads the key, the tempo in BPM, the bar and section counts, and the bar and beat under the playhead.
Speech and other non-musical material. The model still returns a beat grid and a structure for speech, and they mean nothing. It finds no key on such material, so without a key the view hides the beat and pace rows, dims the tempo and says no key detected: tempo and structure unreliable. The stem lanes stay useful: speech lights the Vocal lane.
The Music group lists the Key and Tempo, the Loudest and Quietest sections and their Spread, and the stem-gated loudness: Vocal (integrated over the stretches where the vocal is present), Bed (no vocal) and Vocal balance (the difference in LU; positive means the vocal stretches run louder than the bed), then Drums, No drums and Drums balance. These are Specula's own BS.1770-4 measurements over the file's 100 ms blocks, gated by Apple's presence spans. A side with less than 10 seconds of audio above the absolute gate reads -.
The Chapters task's From Music, or Chapters › Chapters from Music Sections, makes one chapter per section; Chapters from Music Segments makes a finer split. Analyse then gives the per-section table.

The JSON report carries the full result under musicAnalysis (stem spans and activity, beats, bars, tempo, sections, segments, phrases, keys, pace) and the stem-gated readings under stemGatedLoudness. The PDF and HTML reports get a Music summary. Apple's analysis is labelled as such. Every loudness figure in Specula is its own BS.1770-4 measurement; the model's loudness is not used.
The results come from a closed model and can differ between macOS releases and between sample rates of the same audio. Read them as a map of the song, not as a measurement. Below macOS 27 the Stems view and Analyse Music are unavailable, and the command-line tool and Shortcuts do not run the analysis.
File › Check a Folder…, or the Check a folder tile with no file open, opens the Check a Folder window, which checks every audio file in a folder against Deliver to.

Each file gets a row with its Verdict, Integrated, True peak, LRA and Speech (%), measured with the app's speech detection settings and any saved speech regions made for the file. A file's reports are written as it finishes, and the CSV (<folder> check.csv) at the end or on Cancel. The CSV starts with the folder, the delivery category, the targets and the speech settings when they are not the defaults, then one row per file: File, Verdict, Integrated (LUFS), True Peak (dBTP), LRA (LU), Speech (%), Dialog (LUFS), Speech map and Note.
After a check, a change of lens made with Change… judges the rows again at once, from their readings. A change made elsewhere offers Judge Again. The reports and the CSV keep the targets they were written with. Show in Finder shows the CSV.
File › Export… (⌘E), or Export in the top bar, opens one sheet that writes into one folder in one step:
| Item | Available |
|---|---|
| PDF report | always |
| HTML report | always |
| JSON report | always |
| Chapters CSV | with chapters |
| Edited audio | with unsaved edits (then ticked) |
| Residual audio | with a residual for the compare slot in front |

An item that cannot be exported says why. ⇧⌘E (Export JSON Report…) opens the sheet with only the JSON report ticked, and ⌥⌘E (Export PDF Report…) with only the PDF.
Window › Report Preview shows the report before you export it; while the file is measured it says Measuring the file….
The PDF and HTML reports open with the file's facts (the format read from its header, a headerless file's declaration) and Judged against: the delivery category and its targets with the primary first, such as Judged against: Spotify (primary), Apple Music, YouTube, Tidal. Then:
The JSON report has sorted keys, so runs diff cleanly. It carries a delivery block (the category, the primary target and the checked targets), the full-file readings including speechGatedNoiseFloorDBRMS and speechMap, and per-channel statistics. Each entry of loudnessTargets.targets[] has an inline verdict object: status (compliant, penalty, notCompliant or unavailable), actualLUFS, penaltyGainDB, penaltyApplies and truePeakClean for platform targets, and nonComplianceReason, nonComplianceActual and nonComplianceLimit for fails. codecCheck, musicAnalysis and stemGatedLoudness appear when they apply, and a headerless file's declaration under rawFormat. Non-finite values read "-inf"; JSON keeps a silent noise floor as a number.
The chapters CSV is the chapter table with every reading. The command line's specula report --format csv writes a flat table of metric and value, with a Speech map row.
Window › Output and Routing (⌥⌘O), or the output device button in the top bar, opens the Output and Routing window. It is a separate window that remembers its size and position.
Routing changes playback only; the file is untouched. A 5.1 master with the wrong channel order is a routing problem, not an EQ problem: reroute channels for playback to check L, R, C, LFE, Ls and Rs against your monitors.
The preset bar applies a complete mapping in one click. Presets show only where they fit the file's and the device's channel counts:
A preset replaces every route, custom gains included, and mutes the outputs it does not address. Lo/Ro follows the ITU-R BS.775 stereo downmix: the front channel, plus Center at -3 dB, plus the same-side surround at -3 dB (1/sqrt(2), the equal-power gain for summing correlated channels).
Saved presets sit above the built-in ones. Save stores the current routing (sources, gains and phase) under a name; a saved preset shows when the file and device have the same channel counts. Import reads a routing presets JSON file and adds the presets not in your library yet; Export writes all your presets to a file. Right-click a saved preset to rename it, export it alone or delete it. Saved presets persist across launches.
The List | Matrix switch shows the same routing two ways:

The ±90° states are a broadband Hilbert shift, not a polarity flip, so a quadrature relationship survives a fold-down built by hand. Save such a matrix as a preset to recall it.
With the keyboard, Tab reaches every control; a gain takes Return or Space to type a value and the arrow keys to step it by 1 dB. With VoiceOver, each cell is a button named for its route, such as L to R Out, with its state.
Render to File… in the Routing header renders the routing to a new audio file: every route's gain, phase and ±90° shift, computed offline. Only the routed outputs are written, so a 5.1 to stereo downmix on an eight-channel device renders a two-channel file. The file is a WAV, or a CAF when the channel count needs a layout WAV cannot store, and never replaces the source. It bakes the routing only: the Hear level and the channel mutes are not baked. To drop a channel from the render, unroute it in the matrix.
Window › Signal Generator (⌘G), or the Generate a test signal tile with no file open, opens the Signal Generator. It builds test signals from scratch and writes them to a file, so it needs nothing loaded. It keeps your signal across closing the window and relaunching.

A signal is made of modules. A module is one sound with a level, a start time and a length:
A module's level is peak dBFS for the tones, sweeps and the impulse, and RMS dBFS for noise.
Modules that overlap in time sum: three sines at once make a chord, a tone plus noise a tone in noise. Modules placed one after another form a sequence: a second of silence, an impulse, a 1 kHz tone, a sweep.
Each module can add click-free fade in and fade out edges, a Level sweep (from one level to another, linear or in dB) and Bursts (the module on and off: set the on and off lengths and the count; for an impulse, on plus off set the spacing).
Same signal on all channels builds one signal for every channel, with noise decorrelated per channel, so stereo noise is real stereo. Turn it off to give each channel its own modules.
Set the sample rate, the bit depth (16-bit or 24-bit integer, or 32-bit float) and the container (WAV, AIFF or CAF; AIFF stores integers only). Set total length fixes the file's length, padding or trimming the modules; without it the file ends where the last module does. Peak ceiling scales the whole render so its loudest sample sits at the level you set.
A live waveform and spectrum of the signal sit at the top of the window and follow your edits, with peak, RMS and integrated-loudness readouts; long signals preview their first 30 seconds. Preview in Specula renders the signal to a temporary file and opens it in the main window.
Generate… (⌘Return), at the bottom right, asks where to save, writes the file and reports its peak, RMS and integrated loudness. Reveal shows it in the Finder; Open in Specula opens it in the main window. The same generator runs on the command line as specula generate.
For batch QC, scripts and Shortcuts, Specula ships a command-line tool and App Intents on the same measurement engine as the app: the same numbers, without a window.
The tool ships inside the app. Choose Specula › Install Command-Line Tool… (or Settings › General › Install Command-Line Tool…) and Specula links specula into /usr/local/bin, with one admin prompt if that folder needs it. The link points into Specula.app, so app updates keep the command current. By hand:
sudo ln -sfh "/Applications/Specula.app/Contents/Helpers/specula" /usr/local/bin/specula
Check it with specula --version.
The tool shares the app's license and trial. During the trial it prints the days left on stderr; after it, it needs the license activated in the app (Specula › Manage License…), and an unlicensed install exits with code 77 and an explanation on stderr. --help and --version always work. The first licensed run asks once for access to the license in the Keychain: click Always Allow. Do that once in a local Terminal window before using the tool over SSH, where the prompt cannot appear.
specula analyze <FILE>The headline numbers as JSON: integrated loudness, true peak, sample peak, loudness range, speech-gated loudness, stereo correlation, and speech.map (detected, corrected or none).
specula analyze mix-v2.wav
specula analyze mix-v2.wav --no-vad # skip speech detection
specula analyze mix-v2.wav --no-stereo # skip the correlation pass
specula analyze mix-v2.wav --no-sidecar # ignore saved speech regions
--raw-*)analyze, report and edit take headerless raw PCM when you declare the format (see Headerless audio):
specula analyze dump.pcm --raw-rate 48000 --raw-channels 2 --raw-format s24le
specula report capture.bin --raw-rate 96000 --raw-channels 1 --raw-format f32le --format pdf --out capture.pdf
specula edit take.raw --raw-rate 48000 --raw-channels 2 --raw-format s16le --gain=-3 --out take-quiet.wav
specula analyze dump.pcm --raw-preset "Apollo capture"
--raw-rate, --raw-channels and --raw-format are all required for a raw input. There is no default rate.--raw-format tokens: u8, s16le, s16be, s24le, s24be (packed 3 bytes), s24_32le, s24_32be (24 bits low-aligned in 4 bytes), s32le, s32be, f32le, f32be, f64le, f64be.--raw-skip N ignores N header bytes. --raw-tolerance N allows N trailing bytes after the last whole frame. --raw-planar reads one block per channel.--raw-preset NAME uses a preset saved in the app's import sheet; other --raw-* flags refine it.rawFormat.specula compare <A> <B>Two files side by side, with B minus A differences for every metric, and speech.map for each.
specula compare master-v1.wav master-v2.wav
specula compare master-v1.wav master-v2.wav --match-loudness
specula compare master-v1.wav master-v2.wav --no-vad
--match-loudness subtracts the integrated-loudness difference from B's peak and speech-gated readings, so the differences show changes of shape or tone, not of level.--no-sidecar ignores saved speech regions.--out-diff <PATH> also writes the residual, A minus gainB·B, as a 32-bit float WAV file, whatever the name's extension. Pair it with --match-loudness to remove a pure level difference first. A file B at another sample rate is resampled to A's rate; the channel counts must match. The JSON gains a residual block with the residual's peak, RMS and the gain applied; for two identical files both readings are left out.specula edit <FILE> --out <OUT> …Applies edit operations and writes the result to --out (-o). The extension picks the container, in any letter case: .wav (also .bwf, .rf64, .bw64, .w64), .aiff, .aifc or .caf; any other is refused. --out writes 32-bit float and refuses any path that opens the input, the rule Save As follows. To write over the source deliberately, pass --in-place instead of --out: it works for WAV (including Broadcast Wave and RF64), AIFF and CAF, keeps the file's own container, bit depth and channel layout, and writes atomically, so a failure leaves the source untouched. Raw PCM inputs cannot be rewritten in place.
specula edit mix.wav --out mix-norm.wav --normalize-lufs=-14 --lufs-ceiling=-1
specula edit mix.wav --out mix-trimmed.wav --trim 5.0:30.5
specula edit mix.wav --out mix-louder.wav --gain 3
specula edit mix.wav --in-place --dc-remove
Negative values need the --option=value form (--gain=-3, --normalize-lufs=-14); otherwise the leading - reads as an option name.
The operations, applied in this order when several are passed:
--trim START:END: keep only that range, in seconds.--dc-remove: subtract each channel's mean.--invert-phase: flip the polarity of every channel.--gain DB: a uniform gain.--normalize-peak DBFS, or --normalize-lufs LUFS with --lufs-ceiling DBTP (default -1); not both.--limit-tp DBTP: the true-peak limiter at this ceiling, without a loudness change; it runs last, to catch peaks an earlier normalize left.LUFS normalizing and --limit-tp use the same 5 ms two-pass true-peak limiter with 4× oversampled detection.
specula generate --out <OUT> …Synthesizes a test signal: one source per run, across N identical channels.
specula generate --sine 1000 --duration 5 --channels 2 --out tone.wav
specula generate --noise pink --level=-20 --duration 10 --out pink.wav
specula generate --sweep 20:20000 --sweep-shape log --duration 3 --bit-depth 24 --out sweep.wav
Pick one source: --sine HZ, --square HZ, --saw HZ, --triangle HZ, --noise white|pink|brown, --sweep START:END, --impulse or --silence. --level is peak dBFS for tones and RMS dBFS for noise (default -6 and -20). Format: --sample-rate (default 48000), --bit-depth 16|24|32f (default 24), --container wav|aiff|caf (else from the --out extension) and --peak-ceiling.
For layered or sequenced multichannel signals, describe the signal in JSON and pass it with --spec-file. The JSON follows the window's builder (channels holding timed modules); omitted fields take the builder's defaults:
{"channels": [{"name": "L", "modules": [{"source": {"sine": {"frequency": 1000}}, "duration": 5}]}]}
Module fields: source (required: sine, square, sawtooth or triangle with frequency, noise with a colour, sweep with start, end and shape, impulse, silence), start, duration, gainDB, fadeInSeconds, fadeOutSeconds, levelSweep, burst. Spec fields: sampleRate, bitDepth (int16, int24, float32), container, channels, duration, peakCeilingDB. The receipt on stdout gives the format written and the measured peak, RMS and integrated loudness.
specula report <FILE>A full Specula report through the same pipeline, in your choice of format.
specula report mix.wav --mode music # JSON to stdout
specula report mix.wav --mode podcast --out report.json # JSON to a file
specula report mix.wav --mode music --format html --out report.html
specula report mix.wav --mode music --format pdf --out report.pdf
specula report mix.wav --mode broadcast --no-curve
--mode music|podcast|vod|broadcast chooses the targets that judge the file.--format json|html|pdf|csv: JSON is the default. PDF needs --out, and prints a receipt with the input, the output, the bytes written and the mode. CSV is a flat metric and value table.--no-vad skips speech detection, --no-stereo the correlation pass, --no-curve the loudness curve.--vad-threshold (0.1 to 0.9, default 0.5), --vad-min-speech (0.05 to 2 s, default 0.1) and --vad-min-silence (0.05 to 2 s, default 0.1) set the speech detection as the app's Detection settings do.--no-sidecar ignores saved speech regions.analyze, compare and report use the speech regions saved beside a file (its .dlg.json) when they were made for the file's length (within 0.1 s), as the app does, unless --no-sidecar.
Every subcommand prints pretty-printed JSON with sorted keys, so runs diff cleanly. Non-finite values (silence reads minus infinity) print as "-inf".
Specula offers eight actions to the Shortcuts app and Siri, in two variants each, one taking a file and one taking a path as text:
| Action | File input | Path input | Returns |
|---|---|---|---|
| Measurements | Get Measurements | Get Measurements (from Path) | Specula Measurements (I, M, S, LRA, TP, SP, DR, Speech-Gated, Speech %) |
| Numeric compare | Compare Files | Compare Files (from Paths) | Specula Comparison (both sides, B minus A differences) |
| Audio compare | Get Compare Diff | Get Compare Diff (from Paths) | the residual WAV (A minus gainB·B) |
| Report | Get Report | Get Report (from Path) | the report file (JSON, HTML or PDF) |
--match-loudness does.--mode.Speech detection is off by default in the measurement, compare and report actions; turn on Enable Speech Detection in the step when you need speech-gated readings. The two Get Compare Diff actions have no speech detection. The measurement, report and compare actions use saved speech regions made for the file's length.
Binding the file input. A file parameter shows Choose and a small … menu. Choose picks one file. The … menu holds the variable picker: choose Shortcut Input to take the file from a Quick Action, the Share Sheet or the previous step. Do not choose Clear, which removes the binding.
The (from Path) variants take the file's path as text, from the clipboard, Ask for Text or a shell. They accept bare paths (/Users/me/foo.wav), paths with spaces (/Users/me/My Mix.wav), tilde paths (~/Music/foo.wav), single-quoted, double-quoted and curly-quoted paths, and file:// URLs (file:///Users/me/My%20Mix.wav). Quotes are stripped and a tilde expands to your home folder, so Get Clipboard followed by Get Report (from Path) makes a report of the path on the clipboard.
Spotlight phrases: "Measure with Specula", "Measure path in Specula", "Compare files in Specula", "Compare paths in Specula", "Get diff with Specula", "Get diff from paths with Specula", "Build a report with Specula" and "Build a report from path with Specula". The file actions also answer to "Get measurements from Specula", "Compare audio in Specula" and "Get Specula report".
Every action checks the license, as the command-line tool does.
Specula › Settings… (⌘,) has five tabs. Settings holds what applies to every file; speech detection is set in the Speech task, chapter detection in the Chapters task, and the report contents in the export sheet. Each tab scrolls.


One list per category: Music streaming, Podcast and Audiobook, Film and TV and Broadcast, each target with its reference and how a platform plays a master (penalty, turns down only or penalty), the primary target marked, and the number checked out of four. The lists keep the Deliver to picker's rules: under Audiobook ACX stays checked, and under Podcast unchecking the primary target passes the lead to another target, never to ACX. The category and the primary target are chosen with Deliver to.

Custom targets lists yours with Edit and a delete button, and Add Custom Target…. See Custom targets.
The open-source components Specula is built with, and their full license texts.
Help › Keyboard Shortcuts lists every key. The menus declare every key except Space, 1 to 6, D and Esc, which act in the main window only. A key marked with a task works only in that task.
| Key | Action | Where |
|---|---|---|
| Space | Play or pause (with a file; without one Space presses the focused control) | Main window |
| 1 to 6 | Bring a compare slot to the front | Compare |
| D | Listen to the residual, or the slot again | Compare |
The menu items name these keys in their titles: Play (Space), Slot 2 (2), Listen to the Residual (D).

| Key | Command |
|---|---|
| ⌘O | Open… |
| ⇧⌘O | Add Files to Compare… |
| ⌘S | Save (over the file with Settings › Editing and plugins' option, else Save As) |
| ⇧⌘S | Save As… |
| ⌘E | Export… |
| ⇧⌘E | Export JSON Report… |
| ⌥⌘E | Export PDF Report… |
| ⌃⌘E | Export Residual Audio… |
Also: Open Recent, Check a Folder…, Revert to Saved… and Reinterpret Raw Format….
| Key | Command | Task |
|---|---|---|
| ⌘Z | Undo (the edit history; in Speech, its region edits and Level Dialogue) | |
| ⇧⌘Z | Redo | |
| ⌘A | Select All (the whole file, when no text field is being edited) | |
| Esc | Deselect | |
| ⌫ | Cut Selection | Edit |
| ⌥⌫ | Silence Selection | Edit |
| ⌘T | Trim to Selection | Edit |
| ⌘⌫ | Cut to Start (Playhead) | Edit |
| ⇧⌘⌫ | Cut to End (Playhead) | Edit |
| ⌘F | Fade In (Selection) | Edit |
| ⌥⌘F | Fade Out (Selection) | Edit |
| ⌥[ | Fade In to Playhead | Edit |
| ⌥] | Fade Out from Playhead | Edit |
| ⌥↑ | Gain +1 dB (Selection) | Edit |
| ⌥↓ | Gain -1 dB (Selection) | Edit |
In Edit the menu also holds Normalize…, Limit True Peak…, Level Dialogue…, Insert Silence…, Invert Phase, Remove DC Offset, Swap Left and Right, Split to Mono Files… and the plugin commands. While IN, OUT or LEN is being typed in, ⌥↑ and ⌥↓ move the caret.
| Key | Command |
|---|---|
| ⌘1 to ⌘7 | Inspect, Edit, Compare, Chapters, Speech, Spectral, Record |
| ⌃1 to ⌃7 | Show or hide Waveform, Loudness, Spectrogram, Spectrum and phase, Stereo, Chapter map, Stems |
| ⌃⌥1 to ⌃⌥7 | Maximize that view, or restore |
| ⌘I | Show Inspector |
| ⌘= | Zoom In |
| ⌘- | Zoom Out |
| ⌘0 | Zoom to Fit (Reset Zoom in Spectral) |
Also: Ruler Unit (Time, Samples, Bars).
| Key | Command |
|---|---|
| Space | Play / Pause |
| ⌘. | Stop (and return to the start) |
| [ | Slower |
| ] | Faster |
| \ | Normal Speed |
| ← / → | Back / Forward 5 Seconds |
| ⇧← / ⇧→ | Back / Forward 1 Second |
| ⌘← | Go to Start |
| L | Loop (not in Spectral, where L is the Lasso) |
Also: Hold Play Start, Hear at the target's level, and in Compare Listen to the Residual (D).
| Key | Command |
|---|---|
| ⌘Return | Analyse Selection, or Analyse File without a selection |
| ⌘D | Detect Chapters (in Chapters) |
| ⇧⌘D | Deliver To… |
Also: Analyse Music (Cancel Music Analysis while it runs), Detect Speech Again, Check the Apple Music Encode and Compare with the Apple Music Encode.
| Key | Command |
|---|---|
| ⌘L | Analyse Chapter Loudness |
| ⌘B | Add Boundary at Playhead |
| ⌫ | Remove Selected Chapter |
| ⇧⌘L | Level Chapters |
| Key | Command |
|---|---|
| A | Add Region |
| I | Mark In at Playhead |
| O | Mark Out at Playhead |
| S | Split Region at Playhead |
| ⌫ | Delete Selected Region |
| ⌘R | Reset to Detected |
| Key | Command |
|---|---|
| R | Rectangle Tool |
| L | Lasso Tool |
| B | Brush Tool |
| W | Wand Tool |
| ⇧] / ⇧[ | Bigger Brush / Smaller Brush |
| ⌘0 | Reset Zoom |
| ⇧⌘0 | Fit Time to Window |
| ⌥⌘0 | Reset Frequency Zoom |
| Esc | Clear Selection |
| ⌫ | Erase to Floor |
| ⇧Space | Play Selection Only (Solo) |
| ⌥Space | Hear the Result |
| Space | Play the selection's span, or stop an audition |
| Key | Command |
|---|---|
| R | Start Recording / Stop Recording |
| Key | Command |
|---|---|
| ⌥⌘O | Output and Routing |
| ⌘G | Signal Generator |
Also: Channel Statistics and Report Preview. In the Signal Generator, ⌘Return is Generate….
| Gesture | Lanes | Spectrum panel |
|---|---|---|
| Click | Place the playhead | Clear a band the click falls outside |
| Drag | Select a range (Drag setting) or move the playhead (classic) | Select a frequency band (Drag setting) |
| ⌘-drag | Select a range | Select a frequency band |
| ⌃-drag | Move the playhead | |
| Drag an edge or the body | Resize or move the selection | Resize or move the band |
| Pinch, or scroll up and down | Zoom time | Scroll zooms the frequency axis |
| Swipe sideways, or ⇧-scroll | Pan | |
| ⌥-scroll | The lane's own axis (amplitude, LUFS range, frequency) | |
| Double-click | The loudness view: reset its LUFS range | Reset the frequency zoom |
| In 1.4 | In 2.0 |
|---|---|
| The Mode picker (Music, Podcast, VOD, Broadcast) | Deliver to: five categories (Music streaming, Podcast, Audiobook, Film and TV, Broadcast) and a primary target |
| No mode | Inspect, ⌘1 |
| Edit mode, ⌘1 | Edit task, ⌘2 |
| Compare mode, ⌘2 | Compare task, ⌘3 |
| Chapter mode, ⌘3 | Chapters task, ⌘4 |
| Dialogue mode, ⌘4 | Speech task, ⌘5 |
| Spectral mode, ⌘5 | Spectral task, ⌘6 |
| Record mode, ⌘6 | Record task, ⌘7 |
| The Live / Offline switch | Whole-file readings at rest, and Loudness details' Whole file / Live switch. The spectrogram always shows the whole file. |
| Loudness target rows with three icons | The delivery card: the verdict in words, fixes as buttons, a Fix menu for the other targets |
| Settings › Targets | The Deliver to picker and Settings › Delivery targets; chapter detection in the Chapters task's Detection |
| Settings › Speech | The Speech task's Detection |
| Settings › Report | The export sheet's report contents, remembered per category |
| Settings › Layout, sections shown on launch | Each task's own views, remembered per task |
| Section toggles ⌃1 to ⌃7 and focus ⌃⌥1 to ⌃⌥7 | View toggles ⌃1 to ⌃7 and Maximize ⌃⌥1 to ⌃⌥7, per task |
| The file dock's nudge and gain rows, the metrics popover | Compare's slot chips, the metrics matrix and the inspector's slot card |
| The Song detected button | The lens suggestion under Deliver to |
| Export JSON and PDF as separate commands | One export sheet, File › Export… (⌘E); ⇧⌘E and ⌥⌘E open it on the JSON or the PDF |
| Discard Edits | File › Revert to Saved… |
| A drag moves the playhead, ⌘-drag selects | A drag selects, ⌃-drag moves the playhead; Settings › General › Pointer brings the old way back |
| Gain ±1 dB on ⌃↑ and ⌃↓ | ⌥↑ and ⌥↓ |
| The chapter ribbon | The chapter map |
| The Music module | The Stems view |
| ACX: a true peak of exactly -3 dBTP or a noise floor of exactly -60 dB RMS passes | Both fail; Limit peaks for ACX aims at -3.1 dBTP |
Help › Send Feedback…, or Send Feedback… next to the version in Settings › General, opens two questions, what you opened Specula for and whether it did that, plus a free-text field, Anything else? Every field is optional. Send sends it to Headroom Studio; without a connection it is saved and sent automatically later, for up to 14 days. If the server refuses it, Email it opens a mail to [email protected] with the same text.
Feedback is anonymous. It carries your answers, the Specula and macOS versions, whether Specula is in its trial, licensed or expired, and the day of the trial, and nothing that identifies the Mac. An email address goes with it only when you tick You can reply to me about this feedback; it is used only to reply and deleted after 90 days.
On the third day you use Specula during the trial (days of use, not calendar days), a slim banner under the top bar asks two questions about Specula. Answer opens them; the cross dismisses the banner. Either way it does not come back. It never shows with a license.
When the trial has ended, Specula asks Need more time? before it shows the license window, at every launch and on opening a file while it stays open, with the buy button and Enter License… above the questions. What's the extra week for? takes any that apply, plus Anything else?. Send and get one more week sends the answer and runs the trial for one more week from that moment; Settings › General › License and the license window then read "Trial: N days left (extended)", and the command-line tool runs for that week too. Not now goes on to the license window, and the question comes back the next time. The extra week is given once: when it ends, the license window shows on its own.
| Standard | Where it appears |
|---|---|
| ITU-R BS.1770-4 | K-weighting, 100 ms gating blocks, dual gating, 4× true-peak oversampling |
| EBU R128 | Momentary, short-term and integrated loudness, LRA, true peak, programme loudness |
| ITU-R BS.775 | The Lo/Ro stereo downmix presets |
| Silero VAD (MIT) | Speech detection model |
| FluidAudio (Apache 2.0) | Swift wrapper for Silero on Apple platforms |
EBU R 128Specula's loudness measurement is validated against the EBU R128 loudness test set (EBU Tech 3341 and 3342), to the EBU Tech 3341 tolerance of ±0.1 LU. The underlying algorithm is ITU-R BS.1770-4, which EBU R128 builds on.
Specula reads the audio files you give it. It sends no audio anywhere, and needs no network connection to analyse audio: the Silero speech detection model ships inside the app, so speech-gated analysis runs offline, and Apple's music analysis runs on the Mac. Record listens to your input device only while the Record task is open, after macOS grants microphone permission, and writes recordings to your Mac and nowhere else. Specula checks for app updates (Sparkle), validates your licence and sends the feedback you choose to send (see Feedback) over the network; those are its only outbound connections, and none sends any audio.
↑ Back to top