scenario-minimax-music

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Scenario MiniMax Music

Scenario MiniMax 音乐

Overview

概述

MiniMax's music family on Scenario splits in two: Music 3.0 composes and performs a complete song from a style prompt plus a lyric sheet, and Music Cover rebuilds an uploaded song in a new style while keeping its melody. Discover both with

search

and treat

model_schema_get

as the contract: the members share

prompt

sampleRate

, and

bitrate

and agree on nothing else.

Connection and the core loop: see the

scenario

skill; model-agnostic audio work: the

scenario-audio

skill. If a sibling skill named here is missing from your available skills, ask the user to install it (

npx skills add scenario-labs/skills --skill <name>

); unattended, proceed from tool schemas and flag the gap.

Scenario平台上的MiniMax音乐系列模型分为两类：Music 3.0可根据风格提示词和歌词表创作并演绎完整歌曲，Music Cover则可在保留旋律的前提下，将上传的歌曲以全新风格重构。可通过

search

工具查找这两类模型，将

model_schema_get

视为接口约定：两类模型共有的参数仅有

prompt

、

sampleRate

和

bitrate

，其余参数均不通用。

连接方式与核心流程：参见

scenario

技能；与模型无关的音频处理工作：参见

scenario-audio

技能。如果你的可用技能中缺少此处提到的关联技能，请提示用户安装（命令：

npx skills add scenario-labs/skills --skill <name>

）；若处于无人值守模式，可基于工具Schema继续操作并标注缺失的技能。

Quick reference

快速参考

Four modes across two members (names from the live schema):

Mode	Member	Inputs	Behavior
Vocal song	Music 3.0	`prompt` + `lyrics`	sings your words; tags shape the arrangement
Auto lyrics	Music 3.0	`prompt` + brief in `lyrics` + `lyricsOptimizer: true`	writes the structured sheet for you
Instrumental	Music 3.0	`prompt` + `isInstrumental: true`	no lyrics; the prompt is the whole brief
Cover	Cover	`audioUrl` + `prompt`	keeps the melody, restyles everything else

prompt

is required on both members (10 to 2,000 characters at authoring time) and carries style only: genre, mood, tempo in BPM, key, vocal timbre, instrumentation. The words go in

lyrics

(up to 3,500 characters), which also sets the bill: about 700 characters per billable minute at authoring time, one to five minutes. Jobs run minutes, so launch with

wait=false

两类模型共包含四种模式（名称来自实时Schema）：

模式	模型成员	输入参数	功能说明
人声歌曲	Music 3.0	`prompt` + `lyrics`	演唱你提供的歌词；结构标签决定编曲方式
自动歌词	Music 3.0	`prompt` + `lyrics` 中传入简介 + `lyricsOptimizer: true`	自动为你生成结构化歌词表
纯音乐	Music 3.0	`prompt` + `isInstrumental: true`	无歌词；提示词即为全部需求说明
翻唱	Cover	`audioUrl` + `prompt`	保留旋律，重新生成其余所有部分的风格

两类模型都必须传入

prompt

参数（编写本文时要求长度为10至2000字符），且仅用于描述风格：流派、情绪、BPM速度、调式、人声音色、配器。歌词内容需放在

lyrics

参数中（最多3500字符），该参数也决定了费用：编写本文时，每计费分钟约对应700字符，歌曲时长为1至5分钟。任务运行需要数分钟，因此请使用

wait=false

参数启动任务。

Writing the lyric sheet

歌词表编写规范

The lyric sheet is an arrangement script, not plain text. Structure tags go on their own line in square brackets ([Intro], [Verse], [Pre Chorus], [Chorus], [Hook], [Drop], [Bridge], [Inst], [Outro]; the schema lists more). Newlines separate lines, a blank line between sections reads as a pause, and two to four lines per section sing better than a dense paragraph. Parentheses carry performance cues rather than sung words: (whispered), (belted), or a short direction inside an [Inst] or [Solo] passage. Close with [Outro] so the track resolves instead of cutting off. English and Mandarin sing the most consistently at authoring time.

With

lyricsOptimizer: true

the lyrics field flips meaning: pass a plain-language brief ("indie folk breakup song, walking alone at night") and the model writes the structured sheet itself. With it off, it sings whatever the field holds, briefs included.

歌词表是编曲脚本，而非普通文本。结构标签需单独占一行，放在方括号内（如[Intro]、[Verse]、[Pre Chorus]、[Chorus]、[Hook]、[Drop]、[Bridge]、[Inst]、[Outro]；Schema中列出了更多可用标签）。换行表示歌词分句，段落间的空行视为停顿，每个段落2-4行歌词的效果优于密集的大段文字。圆括号内的内容是演唱提示而非演唱歌词：例如(whispered)、(belted)，或是[Inst]或[Solo]段落中的简短说明。请以[Outro]结尾，这样音轨会自然收尾而非突然截断。编写本文时，英文和普通话的演唱效果最稳定。

当设置

lyricsOptimizer: true

时，

lyrics

字段的含义会发生变化：你只需传入白话简介（例如"indie folk breakup song, walking alone at night"），模型会自动生成结构化歌词表。关闭该参数时，模型会演唱字段中的所有内容，包括简介文本。

Covers keep the melody

翻唱功能保留旋律

Music Cover takes its source through

audioUrl

upload_asset

the MP3 or WAV first (see the

scenario

skill) and pass the asset id. It keeps the melodic line and regenerates the rest from the prompt (genre, vocal character, instruments, production), so it fits genre swaps, not remixes that change the tune. It works best on sources with clear vocals and melody, and there is no lyrics field, so the prompt changes the style, never the words. Its cost sat flat per run at authoring time while 3.0 scales with the sheet, so

dry_run

both when choosing.

Music Cover通过

audioUrl

参数获取源音频：请先使用

upload_asset

上传MP3或WAV文件（参见

scenario

技能），然后传入资产ID。该模型会保留旋律线，其余部分均根据提示词重新生成（包括流派、人声特点、乐器、制作方式），因此适合流派转换场景，不适用于改变曲调的混音场景。它在人声和旋律清晰的源音频上效果最佳，且没有歌词字段，因此提示词只会改变风格，不会修改歌词。编写本文时，该模型每次运行的费用是固定的，而Music 3.0的费用随歌词表长度变化，因此选择模型时请先使用

dry_run

参数测试。

Worked example: a trailer song with vocals

示例：制作带人声的预告片音乐

```
search
```
with
```
target="models"
```
,
```
query="minimax music"
```
,
```
public=true
```
. Prefer the newest non-deprecated hit, e.g.
```
model_minimax-music-3-0
```
(a live hit at authoring time: re-discover each session); the cover member lists
```
audio2audio
```
in capabilities.
```
model_schema_get
```
with that id: fields, caps, and defaults before anything else.

model_run

with that

model_id

dry_run=true

, and the exact

parameters={"prompt": "epic orchestral trailer, 140 BPM, D minor, female vocals, taiko and strings", "lyrics": "[Intro]\n(Low strings)\n\n[Verse]\nWe were born below the thunder\n...\n\n[Outro]\n(Choir fades)"}

. Re-estimate after lyric edits: character count moves the price.

Repeat
```
model_run
```
with
```
wait=false
```
, then
```
jobs_wait
```
with the returned job id, re-called with
```
pending_job_ids
```
on timeout, never a second
```
model_run
```
.
```
asset_display
```
the output and listen before laying it under a cut.

使用
```
search
```
工具，设置
```
target="models"
```
、
```
query="minimax music"
```
、
```
public=true
```
。优先选择最新的未弃用匹配结果，例如
```
model_minimax-music-3-0
```
（编写本文时的可用结果：每次会话都需重新查找）；翻唱模型的能力列表中会包含
```
audio2audio
```
。
使用该模型ID调用
```
model_schema_get
```
：先获取字段、能力和默认值，再进行其他操作。

使用该

model_id

调用

model_run

，设置

dry_run=true

，并传入精确的

parameters={"prompt": "epic orchestral trailer, 140 BPM, D minor, female vocals, taiko and strings", "lyrics": "[Intro]\n(Low strings)\n\n[Verse]\nWe were born below the thunder\n...\n\n[Outro]\n(Choir fades)"}

。修改歌词后请重新估算费用：字符数会影响价格。

再次调用
```
model_run
```
，设置
```
wait=false
```
，然后使用返回的任务ID调用
```
jobs_wait
```
，若超时则使用
```
pending_job_ids
```
参数重新调用，切勿重复发起
```
model_run
```
。
使用
```
asset_display
```
展示输出结果，先试听再将其用于剪辑。

Common mistakes

常见错误

Lyrics in
```
prompt
```
or genre words in
```
lyrics
```
: the model then describes a song, or sings your style brief.
Asking for "no vocals" in the prompt: unreliable; set
```
isInstrumental: true
```
.
Passing finished lyrics with
```
lyricsOptimizer: true
```
: they get rewritten; the flag expects a brief.
Faking instrumental passages with filler words: use [Inst] or [Solo] with a parenthetical cue.
Sending a local path or external URL as
```
audioUrl
```
: upload first, pass the asset id.
Covering a track with no clear melody: the cover keeps only what it can hear.

将歌词写在
```
prompt
```
中，或将流派描述写在
```
lyrics
```
中：这会导致模型只是描述一首歌，或是演唱你的风格简介。
在提示词中要求"no vocals"（无人声）：这种方式不可靠；请设置
```
isInstrumental: true
```
。
传入完整歌词的同时设置
```
lyricsOptimizer: true
```
：歌词会被重写；该参数的预期输入是简介。
使用填充词假装器乐段落：请使用[Inst]或[Solo]标签搭配括号内的提示。
将本地路径或外部URL作为
```
audioUrl
```
传入：请先上传文件，再传入资产ID。
翻唱没有清晰旋律的音轨：翻唱模型只会保留它能识别出的内容。