The same voice in every video

What goes wrong, and the fix

Before you start

• A character on the Characters page. Creating one needs a name and at least one photo: press Add reference images, upload the photo, name the character and press Create character. Ours is a generated portrait, so no real person is involved.
• A speech model with named voices. The Voice dialog lists only those: the ElevenLabs models and Kokoro.
• The clips you want narrated, in your Library. Ours are three five second clips without sound: the runway and the rooftop on Kling v3, the night drive on Seedance 2.5.

Save the voice on the character

On the Characters page, open the ... menu on the character's card and choose Voice. The dialog offers two kinds: Catalog voice and Voice clip. Keep Catalog voice. Pick the Speech model, then the Catalog voice; a Play preview button plays a recorded sample of that voice for free. Add a language if you want one, then Save. The card now shows the voice under the description.

Narrate with Speak as

Open the audio composer on the Voiceover task and press Speak as. Pick the character. The model picker switches to the saved model and the Voice list to the saved voice, and a line under the button confirms it: Speaking as Isla Marr · Sarah. Type the line and press Generate audio. We did this five times in a row for two guides, and the pair never moved.

This is the sound of silk at midnight.

Nothing here was built to stand still.

You already know the ending. Wear it anyway.

Put a bed and a cue under the voice

A line on its own sounds bare. Make a music bed in the audio composer's Music task and one sound cue in Sound effect, then build each clip's timeline in Studio: Timelines, New timeline, and in the left panel's Library tab the clip with its +, the take at the playhead where the line starts, the bed, and the cue.

Sleek fashion-show runway track, 120 bpm, deep kick and a crisp clicking hi-hat, glossy synth stabs, instrumental, no vocals. Strong dynamics: two seconds of the full groove, then four seconds of a hushed, near-silent filtered pulse, then the full groove slams back in, and that shape repeats through the track.

Dark synthwave night-drive track, 100 bpm, driving analog bass arpeggio, gated snare, cold pads, instrumental, no vocals. Strong dynamics: two seconds of the full drive, then four seconds of a sparse, hushed passage with only a low filtered bass, then the full drive returns, and that shape repeats through the track.

Warehouse rave build for a rooftop party at night, 128 bpm, instrumental, no vocals: a soft muffled kick and a quiet rising filtered riser that builds for five seconds, then a huge drop with open hi-hats and a big supersaw lead, and that quiet build and big drop repeat through the track.

The cues are one short line each on ElevenLabs Sound Effects v2, trimmed with Mark Out so they stay clear of the voice: camera shutters for the first second of the runway, the engine and spray after the drive's line, and a crowd cheer as the rooftop confetti falls. Press play: the voice is the same across all three because the model and the voice were the same request each time, and it sits above the music.

We measured each export: no gap of silence longer than half a second, and the voice at least 3 dB above everything else while she speaks. Studio mixes every lane at the same level, so these numbers come from the passages we chose, not from a setting.

Three ways to run it

The API path is documented under Characters and locations: a character's voice is used by audio generation with the character id, and by video generation when you ask for the character's voice. We ran the composer path for this guide, not the other two.

Where you still regenerate

• Delivery still varies a little. Same voice, same model, but each take is a new read. If a line lands flat, run it again; the voice will not move.
• No volume or fade in Studio yet. A bed plays at full level. If a passage buries the line, move Mark In to a quieter one or ask the music model for more dynamics.
• No cloning here. You cannot make a catalog voice from your own recording. A voice clip on a character is a reference for models that take one, not a text to speech voice.
• Speed is API only. The composer has no speed control today.

What it costs

Saving a voice costs nothing; previewing one costs nothing. Each line bills the model's per character rate with a minimum. Each music bed and each cue is one flat charge. The clips were priced by length, and the Studio exports are not billed.

Try it yourself

Fashion film, one continuous shot. A model in a sculpted emerald green floor-length gown with long sleeves walks down a mirrored runway toward the camera, camera strobes flashing from the crowd in silhouette on both sides, white haze, slow motion, low tracking shot moving backwards, anamorphic flares, deep blacks. No text, no logo.

Premium car commercial, one continuous shot. An original concept coupe that resembles no production car: a low, wide, wedge-shaped matte white body with a single thin continuous light bar across the nose, no grille, no badges, a smooth flush glass canopy, enclosed rear wheel arches and a sharp ducktail rear. It slides sideways through a wet night street lined with bare concrete walls and plain vertical neon tubes in teal and magenta, no shop windows, no signs, no lettering anywhere. Rain streaks in the air, neon reflections stretched across the asphalt, low side tracking shot, spray from the tyres, motion blur on the passing lights, cinematic contrast.

Music video, one continuous shot. A rooftop rave at night: a dense crowd in silhouette wearing long dark coats, hoodies and bucket hats, arms raised, under sharp green laser beams fanning through a thick haze layer, confetti bursting in slow motion, a city skyline of blurred lights behind them, low handheld camera looking up, backlit, high contrast. No text, no logo.

Your clips and beds will not match ours; a prompt gives the model the scene or the mood, not the take, so find your own hushed passage for the Mark In. The voice will match, because the character carries it.

Checklist

1. Character exists with a name and a photo.
2. Voice dialog: Catalog voice, a model with named voices, preview played, saved.
3. Audio composer: Speak as shows the character and the note names the voice.
4. One generation per line; check each take before you cut.
5. A music bed asked for with strong dynamics, and one cue per clip.
6. Studio: clip, take at the playhead, a five-second bed window from a hushed passage, the cue clear of the voice, Export video.