Dialogue And Voice

Dialogue, voice and audio

Audio is generated whether you direct it or not. Leaving it blank is the single most common mistake: the model invents speech, usually in English, and burns the clip.

1. The audio hierarchy

Direct these four, in this order of priority:

1. Dialogue — foreground, most important
2. SFX — tied to a visible action
3. Ambience — the background bed
4. Music — add in post-production only (see §6)

2. Dialogue format

Dialogue in <LANGUAGE>: <Speaker> speaks in <tone/delivery>. He/She says: <exact words>
Voice: <gender>, <age> years old, <language + accent>, <delivery>, <energy>, <emotional quality>.

Three rules that each cost a generation to learn:

Quotation marks — the contested one

[CONTESTED] Google's Veo prompt guide shows attributed dialogue with quotation marks (A woman says, "We have to leave now."). Separate Google guidance says to use a colon after the speaker's action and avoid quotation marks, because quotes push the model toward rendering the text visually in the video.

Field testing sided with the no-quotes form, and it composes with the known caption bug — anything that nudges the model toward rendering text makes burned-in subtitles more likely. Default to the colon form. If speech fails to trigger at all, try the quoted form as a fallback and note which one your footage came from.

Prefer:   Man says: Bom dia!
Fallback: A woman says, "We have to leave now."

No ellipsis, no mid-sentence periods

An ... or a full stop in the middle of a line stops the voice and the lip-sync dead. Replace mid-sentence periods with commas; use a period or question mark only at the very end. For genuinely separate thoughts, split into two attributed statements.

Wrong:  She says: I want to get so full that... what is it called?
Wrong:  She says: And the next day, I wake up wrecked. That is what I want.
Right:  She says: And the next day I wake up wrecked, cooked like an old puppet,
        that is what I want, and you?
Right:  He says in Brazilian Portuguese: Eu vou querer tangerina.
        He says in Brazilian Portuguese: Só não tá gelada, tá?

Always name the language

Omitting the language causes audio-generation failures and random-language output. Every spoken line — and every background murmur — states its language.

Wrong:  She says: Hm, abacaxi.
Right:  She speaks in Brazilian Portuguese with deadpan delivery. She says: Hm, abacaxi.

Non-English dialogue keeps its accents — nao → não, e → é. Accents drive pronunciation; never ASCII-strip a dialogue line. Spell numbers out in words.

3. Line length and lip-sync

One or two short attributed lines per clip. Crowding garbles the lip-sync and the mix. Do not write "precise lip sync" and expect it — exact word-level sync is documented as unreliable across model families. Ask for the line, accept the approximation, or dub in post.

4. Voice profiles

Define one per speaking character, and reuse it verbatim:

Voice: Male, 25-28 years old, deadpan monotone Brazilian Portuguese, casual understated
delivery, low energy, sounds like he just woke up, zero emotional reaction.

Voice: Male, 40-45 years old, warm friendly Brazilian Portuguese, relaxed casual
delivery, medium energy, slight smile in the voice, genuine helpfulness.

Naming the speaker in every line prevents voice drift between clips.

5. Ambience — concrete audible sounds only

Use 3–5 concrete sounds. Specify the language of any voices. Be specific: "keyboard clacking", not "office sounds".

Wrong: Ambience: city sounds, people talking, background noise
Right: Ambience: car engine hum, a single motorcycle passing, muffled conversation in
       Brazilian Portuguese, footsteps on concrete

What is not an ambient sound — this is where prompts quietly rot:

Reliable concrete beds: soft room tone · clock ticking quietly · distant traffic hum · fabric rustling · silverware clinking · a glass setting down · a gentle exhale · air circulation · a chair creaking.

One bed per clip. Layering muddies the mix.

6. SFX and music

SFX: exactly one per clip, tied to a visible action.

SFX: door handle click as the door opens
SFX: two hollow knocks on the wooden door
SFX: None.

Write sound events as verbs, not nouns — "the mask tearing free with a sharp rip of fabric", not "fabric, ripping". Where a model exposes named layers, use all three:

Audio. Foreground: <the main event>. Bed: <ambience>. Background: <the distant layer>.

Music: None. — always. The model generates a different track for every clip, and the assembly sounds disjointed. Lay one track across the whole timeline in post, ducked −8 dB to −12 dB under dialogue.

7. Silence

Speech cues are the primary trigger for both unwanted audio and burned-in captions. To get a genuinely silent clip:

• Remove every quoted line and every speech cue.
• Never mention the mouth, lips or speaking — "his mouth opens slightly as if about to speak" and "lips press together" both generate actual speech. Convey emotion through eyebrows, jaw and eyes only.
• State the silence positively and explicitly: Dialogue: None. He does NOT speak. Mouth closed. Silent scene. plus No dialogue. No music. Ambient room tone only.
• A thin ambient bed is safer than total dead air, which can read as a glitch.
• Never add an Audio language: line to a silent clip — it forces speech.
• In Google Flow there is also an explicit toggle: View Settings → Return silent videos: On.

8. Failure modes worth knowing

• Depicting people under 18 has been observed to kill audio entirely ("audio generation failed"). Age the character up, or accept a silent clip.
• Upscaling 720p → 1080p strips the audio. Plan the resolution before you commit.
• Re-voicing a clip's spoken audio is restricted on Omni Flash — you cannot reliably chat-edit speech after the fact.

More hard-won failures: docs/veo3-field-lessons.md.