Sound design

Separate spoken voice, environmental sound, action-linked effects and music. Decide which layer communicates the event and which layers should stay quiet. Russian terms: звук, атмосфера, музыка.

Studio example

Hear the greenhouse door creak, then light rain on the glass. The gardener does not speak. A separate gentle instrumental track enters after the door closes.

Use the video model’s audio controls and prompt for sounds that belong to the generated action. Prepare a separate Voiceover or Music asset when you need independent control. Check the inputs of the chosen assembly workflow; separate tracks do not mix themselves merely because they are present in the graph.

Understand and adapt existing audio

Use Audio Breakdown (Разбор аудио) to ask about a recording, and Music (Музыка) to generate a track influenced by it. These are separate operations; you can use either on its own.

1. Prepare the source: upload or choose a library track on a Music node, or use an existing Video result. Connect its output to Audio Breakdown’s аудио input. A video contributes its soundtrack.
2. Enter a question, for example: “Describe the instruments, rhythm, mood and any speech in this recording.” You can also connect text to вопрос; the connected text replaces the question field. Turn on подробный разбор if you want detailed analysis, then run the node. Its one model, Audio Understanding, returns the full answer through разбор as text. It does not produce separated audio stems.
3. Read the answer and listen to the original. Write a music brief describing the desired change, for example: “A quieter instrumental variation with soft percussion and more space for dialogue.” Enter it in a new Music node’s промпт field, or supply the brief through its промпт input. A connected analysis answer is passed as text; it is not automatically rewritten into a music brief.
4. Connect the original source separately to this Music node’s аудио input. The text answer carries no audio. Choose Stable Audio 3 or Stable Audio 2.5, set the duration and run. MiniMax Music 3 and Eleven Music 2.5 do not accept source audio in Vini. Keep лирика disconnected for Stable Audio: these models are instrumental.
5. Listen to the complete generated track. Connect its аудио output to Montage’s (Монтаж) музыка input when ready to use it. Check the assembled result with speech and the clip’s existing sound.

Audio Breakdown and Music can incur separate charges. Review each estimate before running. See Music models for compatible modes and prices, the node reference for inputs, and assemble and finish for the final mix.

Turn an analysis into a MiniMax Music 3 brief

After checking an Audio Breakdown answer against the recording, use its musical description as Music Caption Rewriter → Caption; a suitable разбор answer can be connected directly. Add the scenario, narrative development and your intended changes in дополнительные указания or a текст N input, for example: “Keep the gentle acoustic character; make it instrumental and leave space for dialogue.” The rewriter uses that text to prepare a structured music caption; it does not receive the source recording.

Connect the rewriter’s промпт output to Music → промпт, choose MiniMax Music 3, and review the caption before generating. Keep Music’s аудио input disconnected: MiniMax Music 3 does not accept source audio in this connection. To condition a track on the original sound itself, use the Stable Audio workflow above.

For vocals, supply only the sung words and bracketed section/control tags in the rewriter’s Lyrics field or input and connect its лирика output to Music → лирика. Keep the title, style paragraph and assistant commentary out of that text where you can. The built-in Lyrics generator can draft this sheet from a scenario; review it first. The rewriter turns them into a clean sheet and shows what it removed. It never composes or rewords lyrics. Leave that connection absent for an instrumental track. Audio analysis, caption rewriting and music generation are separate operations with separate estimates. See Music Caption Rewriter for its language, caption check and version controls.

Verify the result

Listen with the image and without it. Check speech intelligibility, clipping, unintended music, timing and the final duration. Native audio availability says a request can include sound; it does not establish accurate synchronization. Isolation and dubbing have separate contracts, and a background output is not a promise of a perfectly clean stem.

Availability and evidence

This technique is documented as an authoring or workflow practice. It is available only through the stated inputs and operations; it is not a new model parameter. An unrendered example is untested. A dated successful case is an observation for its exact model, version, quality and mode, not a guarantee for another scene. Choose a model by the required operation first.

Related guides: music, voice, silence.

See the model–technique applicability matrix for connected operations, dated outcomes and configurations that remain untested.

Reviewed 16 October 2026 against the current ports, question and lyrics inputs, and source-audio model selection. Mocked API tests cover transport and full-text delivery; no new paid render or listening benchmark was made for this example.