Skip to main content
Ask for sound the way you’d brief an editor. The agent finds or makes each piece, places it on the timeline and mixes it, with the music dipping under the voice. Sound on: narration, the music ducking under it, and effects landing with the action.

​
Ask for it

In the Studio app or your own agent, in one message or one at a time:
Quote the script when the words are approved; otherwise give the points and let the agent write it. Name what you don’t want, too: “no music” and “no sound effects” are followed exactly.

​
Free, or HeyGen quality

The Studio app uses HeyGen when you add a HeyGen API key (Home says Add a HeyGen API key to make voiceovers, avatars and clips), and a local voice when you don’t. With your own agent, sign in to the HeyGen CLI or set an API key. The words to use for each are on Voice, music and sound.

​
Use your own recording

Add the file and say what it is (“this is the narration”, “keep the interview audio”). To get timed words from it yourself, transcription runs locally and costs nothing:
Pass --language when the speech isn’t English. It also imports an existing .srt or .vtt. For a talking-head clip that needs captions or a tighter cut, see Captions and recuts.

​
Read a reference soundtrack

To understand an existing recording before making a new soundtrack, run:
With a HeyGen credential, the recording is separated into six stems: vocals, drums, bass, guitar, piano and other. The reading describes each audible stem, then gives an overall description. voice-free mix: points to a mix without the vocal stem. bgm prompt: describes the music for a music generator, or says there is no music. These descriptions use HeyGen; a separate Google or Gemini API key is not used. Without a HeyGen credential, or if the description service fails, a warning explains why descriptions are unavailable instead of guessing. Timed effects are named effect and grouped as S1, S2 and so on, without assigning sound kinds. Each keeps its source stem, time, frequency band, attack, tail, peak level and level above the surrounding music. Model descriptions can explain what a sound resembles; their timing is approximate. The timed lines carry the measurements. Sign in with npx hyperframes auth login, or set HEYGEN_API_KEY, before using the HeyGen services. See audio analysis for transcript, range and local-analysis options.

​
Get captions right

Read the transcript before it goes on screen: names, product terms and numbers are where recognition slips. Keep lines short and broken on meaning, keep them off faces and buttons, and emphasize the two or three words that carry the point. To adjust them by hand, open Captions in the browser editor.

​
Repeat music inside a slot

Use the native loop attribute to repeat a short track for a longer slot:
The track repeats until the 30-second slot ends in preview and export. If you set data-media-start, every repeat starts at that source in-point. Playback speed and rate automation apply to the repeated audio. Looping videos with data-has-audio="true" repeat their sound too, including any silence before the next cycle.

​
Check the mix

Watch once with sound and once muted, then listen once without watching.
  • The voice stays clear through the loudest part of the music.
  • Each effect lands with the thing it belongs to.
  • Nothing cuts off a word or a breath at a scene change.
  • The music ends on purpose, not at the end of the file.
The audio analysis gives each sound a first guess of voice, music or sfx. Voice needs both a speech-like level pattern and voice-range pitch, so a hit with a speech-like tail is less likely to be mistaken for narration. These are guesses, not transcript labels. When you inspect a time range, kind uses the whole source span played by the clip; loudness, hits and the other measurements still describe the selected range. A short slice of a music bed therefore keeps the bed’s kind. Repeated range queries reuse that guess until the source or clip changes. If the whole clip cannot be analyzed, the result warns that kind uses only the selected range; the range’s other measurements are still returned. See audio analysis for the command and output. To dip the music only where the voice sits instead of turning it all down, use a voiceover carve. Music can start after narration begins: the dip then starts at the right level for that moment, and speech that has already ended doesn’t make a new dip.

Practice it: Sound

Free 20-minute lesson with a short video, a quiz and a brief to try. Say the sound in your brief: the music’s mood, the voice, and where the music starts and stops.