Arc Raiders AI Voices Divided the Dialogue Into Human Lines and Licensed Ones

The callout comes before the loot does. You mark a crate or an enemy in Arc Raiders, and your Raider says something about it, in the voice of the actor hired to play that Raider. At launch, some of those callouts — the Arc Raiders AI voices at the centre of the argument — were not the actor speaking. They were a text-to-speech model speaking in the actor’s voice, trained on recordings the actor had made, saying lines the actor had never recorded. Players said some of the generated lines sounded flat, or short of the emotion the moment needed, and they heard the difference before Embark explained it.

Embark’s answer arrived in a video about sound and music, the fifth episode of the second season of its Evolution of Arc Raiders series. Audio director Bence Pajor said the problem went back to earlier studios, where a recorded line locked a gameplay decision in place and changing the decision left the line no longer fitting. He called that a true pain for writers, for the dialogue team and for actors whose other work made a return visit hard to arrange. The fix Embark built in Stockholm, with actors it could reach easily, was to record performances and use them as training data, so the team could try lines out while features were still moving. Every new item, weapon or feature needed words, because in Arc Raiders almost anything can be pinged, and Pajor said recording every possible version was impractical. The game kept changing. Actors, he said, would just go crazy if they had to read all of that.

Diana Gardner, who plays Celeste, called the arrangement freeing, because she could take other work without worrying that a recall would cancel a shoot. She also said the lines the system generated for Celeste were still hers, and Jo Rideout, who plays Shani, felt the same. Gardner said she hoped the voice-acting sector would find a middle ground.

Söderlund’s March interview shows where Embark drew the line: inside the script. In March, chief executive Patrick Söderlund told GamesIndustry.biz that Embark pays its actors for all their time in the booth and keeps bringing many of them back as the game is updated. For select usage, he said, the studio also pays for approval to license their voices through text-to-speech, for lines that aren’t as essential to the immersion of the experience, mostly ping system audio. The same interview had him saying some launch lines had already been re-recorded with real voices, and that there is a quality difference. A real professional actor is better than AI, he said, that’s just how it is. The text-to-speech system stayed on as a way to test fifteen different lines without recording them, and then know what to record.

Celeste, the unofficial leader of Speranza, was rebuilt around Gardner herself, with full motion capture used for the first time in the game, so that her face and voice would match in the Frozen Trail trailer on 6 October and in the update two days later. Embark said Celeste had never fully represented the depth of Gardner’s performance, and that her limited expressions broke immersion. Celeste and Diana, the studio wrote, are no longer hiding behind an imaginary face. More characters are to be brought in line with their real-world actors, and the studio wants human performances in every interaction.

No published list tells a player which callouts use the licensed voices. Embark changed how the Frigate, a machine that arrives in the game, can be marked: from a distance a ping labels the whole machine as it comes in, and up close or on board the same ping marks a particular spot. The patch did not say which of those callouts use the licensed voices. A fan news page told its followers that Embark had replaced most of the AI lines and re-recorded the quest lines with real actors. That claim is still unverified, and a report repeating it said so itself. What Embark has actually said, in March and again in November, is narrower: some lines re-recorded, and the rest replaced where it can be replaced.

At Nexon’s Capital Markets Briefing in April, Söderlund, newly appointed executive chairman of Nexon, described deciding what should be done by hand and what a machine could do more efficiently, and said two games had been built with significantly fewer people at a fraction of the cost expected for a AAA game. He said that thinking was now being brought to the rest of Nexon. Nexon chief executive Junghun Lee went further, calling Arc Raiders a Trojan Horse, a gift carrying a shift in mindset about how technology frees developers and live-service teams to spend more time thinking and less time typing, and presenting Mono Lake, a plan to put the company’s accumulated data behind every team. Lee said the method does not replace creative people, it frees them to create, with context.

Embark’s other game, The Finals, added more synthetic speech in October. Its commentators Scotty and June have used AI text-to-speech since launch, and Dynamic Commentary arrived in the Season 12 PRIME TIME update on 20 October: Scotty now names individual players after a team wipe, a cashout interception or a run of revives, in a beta feature aimed at a handful of such callouts per match, with more events and modes to follow. The system generates the lines live, alongside prerecorded commentary, and respects the game’s Anonymous and Streamer modes by not naming the player. One studio gave its central character her actor’s face in the same month its other game gave its synthetic commentator more to say. Embark had tried naming players before, in December 2024, when a bug made the commentators repeat one top player’s name until Embark fixed it the same month.

Equity, Britain’s performers’ union, warned members in October to proceed with extreme caution over a video game casting that asked for a full assignment of rights, AI and language-model training and speech-model use, and a buyout, with the full terms behind a non-disclosure agreement. Equity’s reading was that a client could use the voice in any way it wanted, wherever, forever, including lines the actor never said. That casting was not identified as Embark’s; Equity named no game or studio. Gardner’s contract was a private deal in Stockholm. In the United States, the 2025 Interactive Media Agreement that ended the video game performers’ strike treats a vocal digital replica as an extension of the performance itself: pay counted by lines generated, a line defined as roughly ten words, consent in a separate signed document describing the use in reasonably specific terms, and a usage report within ninety days of a game’s release.

Camenwolf, posting on Threads, put the question the record keeps answering with a price list: “Does the type of dialogue change where you draw the line?” Jerry Fontaine, commenting on Eurogamer’s Facebook post, wrote: “And yet they have been for two decades? Weird.” He invoked Jennifer Hale, the veteran voice actor, as someone who might have a comment or two. Alana Eways, on the same post, answered with a rewrite of Embark’s explanation: “We didnt want to pay them to read all that” — Fixed it for you. Kwinn Moore, on the fan page’s post about the re-recorded quest lines, asked why a shouted callout should count as an art form at all: “Why tho? Im sorry but yelling ‘Hey Raider’ isnt an art form”. Fontaine and Eways read the studio’s practicality as money and routine. Kwinn Moore reads the ping lines the way Söderlund priced them, as work that does not need a performance. Gardner counted the generated lines as hers. Embark has not said how many lines the licensed voices still speak.