You have a clip you like, and it is silent. Maybe it has adult themes, maybe it is a flirty scene that a mainstream tool would flag. You want it to say a line, you want the mouth to move with the line, you want a few room sounds, and you want a music bed that ends when the picture ends.
Short answer: do the audio after the picture is final, in this order: voice line, mouth match, sound effects, music, mix. Voice first, because the mouth match needs the finished audio; music last, because its length has to match the final cut. We tested each step on its own, not the whole chain on one clip, so the order and the mix levels below are a recipe built from those tests, not a recorded session. For a 10-second clip the whole pass comes to about $0.51.
The 30-second version
- Re-render or add afterwards? Add afterwards when a specific line has to be spoken, or when you want to change one word without re-rendering the picture. Built-in video audio is fine for loose ambience, but it is baked into the render: a wrong line means paying for the whole clip again.
- Voice: start with MiniMax Speech 2.8 HD; round-bracket tags such as
(sighs)are performed. - Mouth: Lip Sync v1 for a video of a face, Talking Avatar v1 for a still.
- Effects: Foley v1.
- Music: "Score this cut" in the Studio video editor.
- Mix: voice loudest, effects under it, music quietest.
A recipe for a 10-second clip
This is a recipe, not a report of a run. Every step uses settings we describe later in the guide.
- Line. About 2.5 words per second is normal speech, so a 10-second clip holds roughly 22 words. Lip Sync makes the finished video as long as the audio, so a line much shorter than the clip cuts the picture short. Example: "You're late again. (sighs) I almost gave up waiting. Come here, sit with me a minute, and tell me about your whole awful day."
- Voice. MiniMax Speech 2.8 HD, the line above, one of its seven emotions if you want one. Export MP3 or WAV and trim the silence at both ends. Cost: about $0.02.
- Mouth. Lip Sync v1 with the clip and the voice file. Cost: $0.012 per second of audio, about $0.12.
- Effects. Foley v1 on the lip-synced clip, duration 10. Prompt: "soft fabric rustle, a chair creaking, quiet room tone". Negative prompt: "music, speech, voices". Cost: about $0.012.
- Music. In the editor, "Score this cut", mood Sultry, note "slow, warm bass and soft electric piano". Cost: $0.36. Through the API, the same thing is one Seed Audio 1.0 request whose prompt asks for an instrumental of about 10 seconds.
- Mix in the editor: voice 100%, effects about 50%, music 20 to 30%. Fade the music in and out.
Total: about $0.02 + $0.12 + $0.012 + $0.36 = $0.51. Twenty clips would be about $10.
No-code route in Studio
You do not need the API for any step. In Studio:
| Step | Where |
|---|---|
| Voice line | Voiceover tool in the Tools panel |
| Mouth match | Lip sync in the Tools panel, or on a video in the viewer |
| Effects | Add sound in the Tools panel, or on a video in the viewer; the editor also offers it on your own generated clip |
| Music | "Score this cut" in the video editor |
| Mix and export | The editor: up to eight audio tracks and fades, exported in your browser |
If you made a voice or a music track through the API, use "Save to My assets", then in the editor open the Audio tab and pick it under Assets.
Step 0: lock the picture
Cut the clip to its final length first. If you join several clips, score after the join (see the last-frame chain for joining). Lip Sync's output length follows the audio: a clip longer than the line is cut to the line, and a clip shorter than the line is looped, with a visible jump at the loop point. So write the line to fit the clip, and make the source clip at least as long as the line.
Step 1: the voice line
Each model below read explicit scripts in full in our tests, and all are marked unrestricted on SpicyAPI. What differs is how each takes direction, which is where lines go wrong:
| Model | How you direct delivery | What happened with cues |
|---|---|---|
| MiniMax Speech 2.8 HD (and Turbo) | Seven emotions, round-bracket tags | (sighs)-style tags were performed |
| Gemini 2.5 Pro TTS | 30 voices, style, accent, instructions note | [sighs]-style cues were acted out |
| Gemini 3.8 Flash TTS | Square-bracket cues | Breathy cues were usually performed and occasionally read aloud (we did not count how often) |
| Seed Speech 2.0 | 41 voices, separate instructions field | Put delivery in the instructions, not the script |
| Chatterbox v1 | A reference voice you have the right to use | <laugh>-style tags were read aloud as words |
Chatterbox is a poor fit for cue tags; use it when you need a specific voice and leave the tags out. The lighter Gemini 3.8 Flash-Lite model reads tag words aloud too, so use the Flash model for tagged lines. If a model reads a tag aloud, your clip says "sighs" in a flat voice: test one short line before committing a long script. The TTS comparison has costs per paragraph.
Step 2: make the mouth match
A video of a face: Lip Sync v1. It redraws only the mouth, so the footage decides the result: one person, face reasonably large, nothing covering the mouth, one continuous shot. The audio must be 3 to 60 seconds, MP3 or WAV. A line shorter than 3 seconds, such as a single breath, is rejected, so pad it with a moment of silence. What comes back is the same clip with the new line as its sound track, as long as the audio, at up to about 720p. It is billed by the second of audio, rounded up. The lip-sync guide has the footage checks.
Only a still: Talking Avatar v1. Image plus voice file gives back a speaking clip. The output is a compact talking view, not a full scene, so for a moving scene use Lip Sync on video instead. Portrait and audio limits are in the Talking Avatar guide.
Skip it if nobody on screen is speaking.
Step 3: sound effects
Foley v1 takes an MP4 (up to 90 MB) and returns an MP4 with a generated soundtrack attached. Set duration_seconds (1 to 30) to your clip's length, because it is billed. Name the sources in the prompt and keep speech and music out with the negative prompt, as in the recipe: prompt "soft fabric rustle, a chair creaking, quiet room tone", negative prompt "music, speech, voices". We did not test whether voice-like sounds appear when the negative prompt leaves them out, so include them.
Treat Foley's soundtrack as one track in the mix. Keep your voice file separate and put it on its own track; if the Foley MP4 still carries the spoken line and you hear it twice, turn that clip's own audio down to zero in the editor and let your voice track carry the line. The Foley guide covers timing checks and pulling the soundtrack out of the MP4.
Step 4: music that ends with the clip
For a 5 to 15 second clip: "Score this cut" in the editor. Pick a mood (Tender, Playful, Sultry, Tense, Epic, Bittersweet), add an optional one-line note, and turn on "Keep it under voices" if there is a spoken line; that switch shapes the score's dynamics so it leaves room for the voice, it does not lower the track volume. It writes an instrumental track sized to your cut, at least 10 seconds and up to about two minutes, and places it on an audio track from 0:00. The length is written into the request rather than set exactly, so expect it to run a little long or short: trim the end on the audio track and fade it out. The price is on the button ($0.36 at the time of writing).
Other routes:
- Through the API: Seed Audio 1.0. It is the model behind "Score this cut". One prompt returns one MP3 up to about two minutes for a flat $0.36; ask for an instrumental and state the length you want. The same prompt can also describe speakers and room sound when you want a whole scene from text alone. The Seed Audio guide has prompts.
- Short films: "Score this film" in the short-film tool writes an instrumental score from the film's logline, look and mood, mixed quietly under dialogue.
Keep scores instrumental. MiniMax Music 3.0 writes whole songs and decides their length itself, so it is the wrong tool for a bed under a clip.
Step 5: mix
Put the voice file on track 1, Foley's soundtrack on track 2, the music on track 3. The editor has up to eight audio tracks, fades, and takes audio from local files or My assets. Starting levels, untested: voice 100%, effects about 50%, music 20 to 30%. Listen on phone speakers as well as headphones, and raise the effects until you can hear them in the pauses between words. Fade the music in and out so it does not click in or stop mid-note, and do not fade through effect transients like a contact sound. If you change the cut later, regenerate the music at the new length.
What the audio pass costs
| Step | Billed by | For a 10-second clip |
|---|---|---|
| Voice (TTS) | Per second of audio or per 1,000 characters, by model | about $0.02 |
| Lip Sync v1 | $0.012 per second of audio, rounded up | about $0.12 |
| Foley v1 | $0.0012 per second of requested duration | about $0.012 |
| Score this cut / Seed Audio 1.0 | Flat per request | $0.36 |
Music is the largest single line, about twice the rest of the pass combined. Voice and effects are cheap enough to retry; Lip Sync is where retries add up, so settle the voice first.
FAQ
Do I need lip sync? Only if a face on screen is meant to be speaking. For a voiceover over a scene, skip it.
How do I get music that ends with my clip? Use "Score this cut" in the editor: it writes the score to your cut's length, and you trim or fade the last second.
Can I add sound to a silent AI video? Yes: Foley generates effects from the video, and the voice and music steps add a line and a score.
