How do solo filmmakers use AI voice conversion for their movies?

A solo filmmaker records every character's dialogue themselves, then runs the recordings through speech-to-speech conversion tools like ElevenLabs to turn that single performance into a cast of different voices. This is the method used for the AI feature Boss, where director Jenny Krakovsky performed all the dialogue in a Liverpool accent and converted it into cloned voices for every character.

What does speech-to-speech voice conversion actually change about a performance?

A recorded line carries two things: the instrument (timbre, pitch range, grain of a particular throat) and the reading (where the stress falls, the pauses, the breath, whether a threat is meant or bluffed). Voice conversion discards the instrument and preserves the reading. So every converted character in a film reads their lines the way one person decided they should be read - the timbres differ, but the thinking does not.

Can one person make a good AI feature film entirely alone?

One person can generate sets, images, music, and voices from a laptop, but cannot import the second opinions a real cast provides - the disagreements and accidental discoveries a director could never generate alone. The better approach is to spend money on the one input the laptop cannot produce: ask other people to read the parts, keep your own performance for the character who is actually you, and play scenes back with your eyes closed to check every character doesn't pause in the same place.

Nobody important -- CinePrompt Field Notes

In the autumn of 1948, on a soundstage at Ealing, a cinematographer named Douglas Slocombe locked a camera to the floor and refused to let anyone near it. The script for Kind Hearts and Coronets called for a funeral in a country church with six members of the D'Ascoyne family seated in the pews, and all six were Alec Guinness. Slocombe would not trust the optical printers of the day, so he did it in the camera. He divided the frame into seven parts, exposed one, blacked out the rest, wound the film back to the first frame, and waited while Guinness spent three hours becoming the next relative. The shot took two days. "If even the studio cat had tripped over the tripod," he said later, "we would've been knocked for six. I slept overnight beside the camera."

The shot runs eleven seconds. Audiences laughed at it in 1949, and they still do, because the trick is meant to be seen. One man is playing eight people, the film knows you know, and the pleasure is watching him make each of them a different idea. The banker is not the parson. The parson is not the admiral who goes down saluting with his ship. Guinness wore one face under the makeup and found eight ways of thinking with it.

Hold that shot in mind. Nearly eight decades later, someone made the same multiplication, and this time the point was that you would not see it.

One woman, every part

In May, a London commercials director named Jenny Krakovsky announced she would make a feature with AI, on her laptop, in thirty days. It took fifty, and she puts the cost at around seventeen thousand pounds. The film is Boss, a crime caper about women turning gangster in 1960s Liverpool, a period piece she could not get financed the conventional way. She built the sets from archive photographs and generated the images with Seedance. The Merseybeat soundtrack came out of Suno. Roger Avary has hailed her as "the vanguard of indie filmmaking," and she has a line for the skeptics that she clearly enjoys: "I'd rather make fake films for real people than real films for fake people."

The voices are where the film becomes interesting. Krakovsky recorded every character's dialogue herself, in a Liverpool accent, and ran the recordings through ElevenLabs to turn her performance into a cast of cloned voices. Asked on BBC Radio 4 where the voices came from, she said: "I cloned Scouse voices that I found on the internet. Nobody important or whatever." Asked whether the owners would find out: "They're not going to know, are they?"

Set the ethics to one side for a few paragraphs. They will come back on their own. Look first at what the method does to a film, because it is the method every solo filmmaker with a laptop is about to be offered.

What conversion keeps

A recorded line carries two things. One is the instrument: the timbre, the pitch range, the grain of a particular throat. The other is the reading: where the stress falls, how long the pause runs before the second clause, whether the character means the threat or is bluffing, where the breath goes. Speech-to-speech conversion separates these two and keeps exactly one of them. It discards your instrument and preserves your reading, which is the whole reason it exists. You perform, the software dresses you.

So every gangster, every mother, every copper and barmaid in Boss reads her lines the way one woman decided they should be read. The timbres differ. The thinking does not. I have seen only clips and what others have written. A trade newsletter that watched scenes from it described "voices that don't quite belong to those faces," and that seam has a precise location. The faces were generated from a crowd. The throats were taken from a crowd. Underneath both, one person is deciding every inflection for twenty people, and the ear can hear twenty instruments playing one part.

Guinness ran the trick the other way round. Eight characters, one actor, and the multiplication lived in the readings. He changed how each man thought, then let the makeup department handle the rest. Boss kept one reading and multiplied the makeup.

What a cast is for

A cast is usually described as a set of faces and voices. Its real function on a set is less glamorous. A cast is the mechanism by which a director imports decisions she would never have made. The actor who reads the line wrong and turns out to be right. The supporting player who finds a laugh the writer did not know was in the scene. The day player who mumbles a throwaway and makes it the line people quote. The lead who refuses a beat on the third day and forces a rewrite that saves the act. Every one of those is a person disagreeing with the director in a way the director could not have generated alone, because she was the one holding the original idea.

Kierkegaard published most of his major work under invented names, then announced in an appendix to one of them that "in the pseudonymous books there is not a single word by me." Nobody quite believes him, and yet the pseudonyms work, because he labored to make them disagree. They argue with each other across books. One thinks the aesthetic life is the only life worth living; another thinks it is despair wearing a nice coat. The names were the easy part. The disagreement was the work.

A solo AI feature has, for the first time, solved every problem in that list except the one that matters. Seedance can fill a Liverpool courtyard with a thousand extras, and voice conversion can supply twenty timbres from one microphone. What none of it can supply is the second opinion. The models with native audio will offer one if you ask: Veo, Kling, Seedance, Grok Imagine, and recent WAN all generate a reading in the same pass as the picture. That reading is the statistical average of every line of its kind the model has heard. So the solo filmmaker's choice comes down to two options, her own reading or the crowd's, and neither of those is a cast.

The part that was taken

Now the ethics come back, and they arrive carrying the argument with them.

A human voice is the shape of one particular throat, worn in by one particular life. The vowels of someone who grew up off Scotland Road carry the docks and forty years of shouting over traffic. On the scale of the species there are eight billion of these instruments, no two alike, and the internet is now full of recordings of them, posted to say something to someone. To a model they look like a crowd. Each one belonged to an individual, and Kierkegaard's old warning applies with unusual force: the crowd is untruth, because the moment you treat a thousand people as a crowd you stop seeing that each of them is somebody.

"Nobody important" is the crowd talking. And the irony is exact. From those strangers, Krakovsky took the one thing that could not improve her film, their timbre, and left behind the one thing that could, which was their reading. The Scouse voices she found online belonged to people who know how a Liverpool woman says a threat. That knowledge was the only part of them a film about Liverpool women actually needed, and it was the part the software was built to throw away.

She was right that the throats were unimportant to the performance. She was wrong about the people.

The cheap cast

None of this means a person alone with a laptop cannot make a film. It means she should spend her savings on the one input the laptop cannot produce, and that input has never been cheaper.

Ask other people to read. A friend, a neighbor, a stranger from the city the film is set in, paid a fair day rate for an afternoon. Give them the scene and no direction for the first pass, then listen for the moment they do something you would never have done. If the project needs voice conversion, convert someone else's reading into a voice you have the rights to, and keep your own performance for the one character who is actually you. Before you lock the cut, play a dialogue scene back with your eyes closed. If every character pauses in the same place, the film has one character.

Slocombe slept beside his camera for a single eleven-second shot, because the trick only worked if the six men in the pews were really six different performances by the same man. The audience could see the seams, and loved him for them. The new trick hides the seams and has nothing behind them.

Guinness gave Ealing eight men. Boss has one.

Bruce Belafonte is an AI filmmaker at Light Owl. He reads all his own scratch dialogue aloud, which is exactly why he pays someone else to read it again.