AI by department · Sound and post sound
The tools are good here, and the sharpest consent line in the industry sits here.
Source separation, dialogue isolation, noise and reverb removal and speech-to-text logging are in daily professional use, and they work. It is also the department where the industrial fight happened, because a voice is a person.
Last checked September 2026. Nothing on this page is legal advice.
What works
What works.
| Tool | What it does | Why it matters |
|---|---|---|
| Dialogue isolation and lav rustle | iZotope RX 12’s Dialogue Isolate and De-rustle. | Rescues performances that would otherwise need ADR. iZotope’s published prices in September 2026: Elements $99, Standard $399, Advanced $1,399. |
| Source separation and recovery | Music Rebalance and De-bleed; Spectral Recovery rebuilds content above 4kHz lost to compression. | Archive with no surviving stems becomes usable, and compressed contributor audio sits in a broadcast mix. |
| Separation inside the edit | Adobe’s Enhanced Audio, in beta since September 2026, splits a mix into dialogue, music, ambience and effects. | The decisions start being made in an editor’s application, before they reach post sound. |
| Speech-to-text logging | Transcribes everything, searchably. | The biggest time saving in the post chain — and transcription still hallucinates, so it is a starting point, not a record. |
Almost none of it is generative: it reads what you already recorded rather than inventing, which keeps it clear of most rights arguments. Heavy processing has a sound, and a track rescued at high settings carries an artefact a dubbing mixer can hear.
The consent line
A voice is a person, and a general release almost certainly does not cover cloning it.
Any synthesised or cloned voice needs the performer’s specific, informed, written consent for that use. Not a general clause, not one signed years ago for a different purpose, and not an assumption that owning the recording means owning the voice. Cloning starts at $6 a month on ElevenLabs’ published price in September 2026, so cost is not a barrier to anybody. And in AI & Work in the Media, Arts & Entertainment Sector in Europe, published in July 2026 by FIA, FIM, UNI MEI and EFJ, voice is the worst-affected category of all: respondents reported speaking engagements down by 80 to 90 per cent, and that “dubbing jobs have practically disappeared”.
- List item text
- List item text
- List item text
- List item text
Before you agree to anything
Ask before you agree to anything.
| Ask | Why |
|---|---|
| Is any voice work on this production synthetic | Including narration, crowd, looping and temporary lines that survive into the final mix. |
| Whose consent exists, and for exactly what use | A named person, a dated document, a described use. Anything vaguer is an assumption. |
| Is uploaded audio used to train the vendor’s model | Check the processing agreement and the retention period. |
| Can this run on my own machine | Several of the strongest tools do, and most of the confidentiality problem goes with it. |
