Choose whether the voice will be generated with the video or produced separately, then approve the performance before final captions and delivery. Narration, off-screen dialogue, and visible speech need different timing and synchronization work.
The AI market now offers both audio-video generation and conversational speech tools. That makes “add a voiceover” an incomplete instruction. The real question is which parts of the performance you need to control or replace independently.
Choose the audio route from the revision you expect
Route
Good reason to use it
Main dependency
Voice generated with video
The scene depends on coordinated timing and visible performance
A voice correction may require changing picture too
Speech first, picture second
Exact wording, casting, or localization matters
Visible speakers still need synchronization
Picture first, narration later
The story can use an external narrator or off-screen voice
Narration must fit the available action and pacing
For a whispered clue whose exact name matters, separate speech may provide a cleaner approval path. For an interruption with both people reacting continuously, an integrated route may be worth testing.
Do not compare a text-to-speech demo with a complete synchronized conversation as if they were the same output. One supplies audio; the other must also make the character’s body, mouth, and reaction fit the exchange.
Write a dialogue specification with playable intentions
Use this original practice exchange:
MARA: You said the room was empty.
ION: It was.MARA: Then whose coat is that?
The performance depends on intention. Mara first tests Ion’s account, Ion tries to close the question, and Mara introduces visible evidence. Giving every line the direction “dramatic and emotional” can flatten those differences.
Prepare a line table:
ID
Speaker
Intention
Timing dependency
L01
Mara
Invite an explanation without revealing the evidence
Leave room for Ion to decide how to answer
L02
Ion
End the discussion quickly
Do not overlap the next revelation accidentally
L03
Mara
Force attention toward the coat
Coordinate the final words with the visual reveal
Keep the exact spoken text separate from directions. A service may read unsupported instructions aloud. Use only the direction syntax documented for the selected voice model.
Generate an exchange, then create controlled pickups
Conversational context matters. Generating “It was” alone may produce a cheerful, explanatory, or uncertain delivery that does not fit Ion’s attempt to end the discussion.
Start with a contextual audition of the exchange. Compare whether the voices remain distinct, the pauses support the intentions, and the final question points the listener toward the reveal. Then replace individual lines where the service and workflow allow it.
ElevenLabs’ current Text to Dialogue documentation supports attributed conversational turns and model-specific delivery tags. It also notes that multiple generations may be needed. Use those controls as a way to direct delivery, not as a guarantee of deterministic acting.
Save the chosen voice identifiers, model, supported settings, and text with each take. A similar-sounding voice name in another service does not establish the same licensed voice or performance behavior.
For a pickup, play the preceding and following lines first. Approve the replacement inside the exchange, not only as a standalone audio file.
Use the edit to support performance
A J-cut lets the next line begin before the picture cuts to its speaker. An L-cut lets a voice continue over the listener’s reaction. Both can make generated coverage feel like a conversation rather than three unrelated talking portraits.
For this scene, Ion’s “It was” might play over Mara holding his gaze. Then the picture can reveal the coat as her final question lands. That gives the listener information through reaction and object coverage, rather than requiring a flawless mouth close-up for every word.
The cut must still make speaker attribution clear. If both voices sound alike and neither mouth is visible, off-screen delivery can become confusing.
Leave some silence for intention. Removing every pause can make the accusation feel mechanically assembled. Conversely, a long gap with no visible thought may simply slow the scene. Time the silence against the performance and story, not a generic pause preset.
Treat lip synchronization as its own production task
If the mouth is visible, inspect the opening, consonant closures, phrase endings, and facial motion during the line. Check whether the selected speech file is actually the one used by the synchronization process.
An important market distinction is whether a tool animates from audio, transfers a recorded performance, or generates speech and picture jointly. Those routes require different inputs.
Runway’s published Act-Two conversation workflow uses recorded driving performances and compositing to build multi-character dialogue. It is not a camera-free text-to-speech button.
If you want no new filming, choose a suitable synthetic route or use off-screen coverage that serves the story. Do not discover the recording requirement after writing an episode around unsupported facial performances.
Diagnose timing drift before regenerating the voice
Compare the same line near its start and end:
- A constant offset suggests an alignment problem: the audio begins too early or late relative to the picture.
- A mismatch that grows suggests different durations, a different take, or a timing conversion somewhere in the chain.
- Correct timing but wrong expression is a performance problem, not a synchronization offset.
These are diagnostic clues, not proof of a single cause. Inspect source duration and the actual project settings before applying a correction.
Do not fix drift by arbitrarily stretching an emotionally important line. A small technical adjustment may be appropriate, but changing speech speed can alter emphasis and breath. When the performance changes materially, reapprove the line and update its captions.
Mix for intelligibility and preserve replacement options
Keep dialogue, ambience, effects, and music separate where possible. If every sound arrives in one mixed file, replacing a word may also remove the room tone or music underneath it.
Listen to the scene quietly through ordinary speakers. If the plot depends on a whispered name that cannot be understood, lower competing sound or reconsider the performance. A subtitle should support the scene, not conceal an unusable dialogue mix.
Track the rights for each selected voice. ElevenLabs’ current publication policy distinguishes free-plan and commercial permissions and creation-time conditions; do not assume an upgrade clears older material.
ElserStudio documents character Voice Profiles as part of asset organization. Use that association to retain casting decisions, while checking the actual provider’s speech and synchronization features. ElserStudio guide.
FAQ
Should I balance the voices before adding room effects?
Establish intelligible, balanced dialogue first, then apply the scene treatment. Strongly different reverb on individual lines can expose joins and make later pickups difficult to match.
Can I keep the original voice when translating?
Only if the service, voice permissions, and language capability support that use. Recheck pronunciation and acting with a competent speaker rather than assuming voice similarity means a good translation.
What if the best visual take contains an unwanted voice?
Check whether usable separate audio is available. If it is one mixed track, replacing speech may require reconstruction of ambience or a different shot. Do not assume a clean dialogue stem exists.