Specula app icon

Workflow: Podcast

Measure the talking, not the music bed.

Podcast platforms judge dialogue loudness, not the whole file's. When an episode opens, Specula finds the speech with a neural speech detector, measures the loudness over those blocks and the noise floor over the rest, and judges the episode against Apple Podcasts and Spotify under Deliver to Podcast. Where the detector is wrong, the Speech task corrects the regions by hand, and every reading follows.

7-day trial, one-time purchase, macOS 14 and later

Specula's Speech task on a podcast episode: detected and corrected speech regions on the waveform, the region tools in the task bar, and Level dialogue with its prediction in the sidebar.
The Speech task. Detected and corrected speech regions on the waveform, the region tools in the task bar, and Level dialogue with its prediction in the sidebar.

Why whole-file LUFS is the wrong number

Loudness over the whole episode mixes the voice with the bed, the stingers and the pauses, so podcast platforms judge dialogue loudness.

An integrated reading over the entire episode averages the talking with everything that is not talking. A loud cold-open bed, a sponsor stinger and a long quiet pause land in the same average, so the file can read on target while the voice sits a couple of LU off. Specula measures the loudness only over the speech blocks, so the number checked against Apple Podcasts or Spotify is the dialogue's.

The same separation matters at the quiet end. The room tone between phrases is invisible on a whole-file meter and inaudible on headphones, but the noise floor is measured over every block that is not speech, so a noisy capture shows up as a number before it ships.

Episode speech music · pause music speech speech bed whole-file LUFS the bed drags the average dialogue-gated LKFS metered on speech only Episode whole-file LUFS the bed drags the average dialogue-gated LKFS metered on speech only speech music · pause
What the gate keeps. Whole-file LUFS averages the bed and the pauses in with the voice; the dialogue-gated number meters only the blocks marked as speech, and that is the figure the platforms judge.
Apple Podcasts−16 LKFS dialogue-gated, true peak at or below −1 dBTP. Turns a loud episode down and plays a quiet one as recorded.
Spotify (Podcast)−14 LKFS dialogue-gated, true peak at or below −1 dBTP. Turns an episode up or down to its reference.
ACXBetween −23 and −18 dB RMS, peaks less than −3 dB and a noise floor less than −60 dB RMS, for audiobooks: it leads under Audiobook and can be checked under Podcast. See the audiobook workflow.

From episode to verdict

The detector does the first pass, the Speech task corrects it, the verdicts read off the result.

1. Deliver to Podcast
Open the episode. Deliver to Podcast judges it against Apple Podcasts and Spotify (Podcast), with one of them primary. A file that is mostly speech under another category gets a note under Deliver to, "This file is mostly speech.", with a switch to Podcast; the category changes only when the switch is pressed.
2. Speech, found on load
The Silero neural speech detector runs when the file opens, on a 16 kHz mono copy, and classifies every 100 ms block as speech or not. The Speech group in the sidebar reads the Dialogue loudness in LKFS, Speech % and the Noise Floor over the whole file, and the loudness graph's Dialogue overlay shows where the speech is.
3. Correct it in the Speech task
The Speech task (⌘5) puts the regions on the waveform: green as detected, amber where corrected, blue where added by hand. Drag a region to move it or its edges to resize it; Add (A), In and Out (I and O), Split at the playhead (S) and Delete (⌫) do the rest, with 50 steps of undo. The corrections are saved beside the file and come back when it reopens at the same length, Save As carries them to the new file, and Reset to Detected (⌘R) discards them and detects again.
4. Read the verdict
A correction measures the file again, so the dialogue loudness, Speech % and noise floor, the verdicts, the fixes and the reports follow it. The card says what each platform does: "Apple Podcasts plays this as recorded, 8.7 dB under its −16 LKFS reference.", and "Spotify (Podcast) turns this up by 10.7 dB." With the TP threshold at −1 dBTP, every breach is marked on the waveform: an intro sting, a clipped laugh, a hot plosive.
5. Export the receipt
File › Export… (⌘E) writes the PDF, HTML and JSON reports into one folder, and they say whether speech was detected or came from corrected regions; ⇧⌘E and ⌥⌘E open the sheet with the JSON or the PDF ticked. A season goes through File › Check a Folder… in one run.

Speech detection, and where the corrections go

The detector is Silero, a neural model (MIT-licensed, run through FluidAudio) that ships inside the app, so detection works offline. Where it cannot run, a spectral fallback classifies on four features (a high-frequency gate, the 300 to 3,400 Hz band-energy ratio, spectral flatness and a harmonicity-plus-flux test), so the dialogue reading still has a result.

Tune it in the Speech task. Detection in the task bar (Speech › Speech Detection…) holds three settings that apply to every file: Threshold (0.1 to 0.9, default 0.5; lower picks up quieter speech), Shortest region (0.05 to 2.0 s, default 0.10; longer rejects clicks and short interjections) and Merge gap (0.05 to 2.0 s, default 0.10; longer fuses regions through breath pauses). A change detects again over the detector's own regions only. Corrected or imported regions stay until Analyze › Detect Speech Again, which asks first.

Corrections reach everything that measures speech. The dialogue loudness, the noise floor and the verdicts in the app; the reports; and specula analyze, compare and report, the Shortcuts actions and the Finder services, which read the saved regions when they were made for the file's length. specula report takes the app's detection settings as --vad-threshold, --vad-min-speech and --vad-min-silence.

Then fix the level

Level dialogue, in the Speech task's sidebar and in the Edit task's Loudness menu, works off the same speech regions as the measurement. It ducks the room tone between phrases by the amount set and can lift the dialogue to a target in the same pass, so an episode comes up in level while the pauses stay quiet. Before Apply it shows the dialogue loudness and the noise floor now and predicted after leveling, with the floor against ACX's limit; after Apply, the file's own readings. It buys a few dB on a borderline floor; hiss from the room or the preamp needs fixing at the source. It is offered once detection has finished, on files with 10 % speech or more.

For a straight level move, the card's Normalize dialogue to −14 LKFS shifts the whole episode by the gain that lands its dialogue loudness on Spotify's reference. It is offered where the target changes the level: Apple Podcasts plays a quiet episode as recorded, so it offers no level fix. Limit peaks to −1.0 dBTP caps the peaks and leaves the loudness. Each is an edit in the Edit task's History, and Save or Save As writes the file. Editing

Before the take. The Record task's room check reads the room's floor over the last five seconds, measured the way the file's noise floor will be, against ACX's limit. Recording

In practice

Speech regions on a spoken-word waveform in the Speech task, the corrected set saved beside the file.
Speech regions. Silero marks them when the file opens, offline; corrected regions are saved beside the file and come back with it.
The Level dialogue panel in the Speech task's sidebar: the room tone duck, the lift to a target, and the dialogue loudness and noise floor now and predicted.
Level dialogue. The dialogue loudness and the noise floor now and predicted, before Apply.
The sidebar's Speech group: dialogue loudness in LKFS, Speech % and the noise floor in dB RMS.
The Speech group. Dialogue loudness, Speech % and the noise floor, the floor measured over the pauses.
The delivery card under Podcast: Apple Podcasts plays the episode as recorded, and Spotify (Podcast) turns it up.
The Podcast card. Apple Podcasts plays a quiet episode as recorded; Spotify turns it up.

Ship the episode on target

The episode's number is the dialogue's: measured over the speech as corrected, judged against Apple's and Spotify's specs, and fixed in the same window. Level dialogue for the floor, Normalize dialogue for the target, Limit peaks for the peaks.