Stable Audio

Instrumental beds, loops and effects with a duration you can set.

Model profile

Vendor: Stability AI
Form: 网页服务
Input: 文本 · 参考音频
Output: 音频
Licence: 商用范围随服务条款与订阅档位变化，训练数据来源以官方说明为准
Verified: 2026-10-30

What it lets you finish

Describe an atmosphere or an instrumental texture and it returns instrumental audio with no voice, at a length you specify, from a few-second transition to a bed over a minute long. It can also transform existing audio, keeping the motion and changing the timbre. It suits video beds, game ambience layers and sample material. It does not handle lyrics or singing; that is another class of tool.

Inputs and outputs

Input is text; some tiers accept reference audio for audio-to-audio conversion. Output is stereo audio, usually without stems. The duration you set fixes the length but does not make the result seamlessly loopable. Each run from the same description differs noticeably in timbre and texture. The listening examples and instances below are references for this kind of output, not this model's generations.

Strengths

Ambient beds and atmospheric textures are its comfort zone: sustained tones, grain and space hold up.
Duration is controllable, so transitions, idents and fixed-length beds need no trimming.
Audio-to-audio keeps the rhythmic skeleton of source material while swapping the instrumentation.
With no voice involved, there is one less layer of vocal and likeness terms when scoring commercially.

Weaknesses and failure modes

Loops expose the seam: head and tail levels and reverb do not match, and looping the file as-is gives an audible click. A reproducible test is to loop it ten times and listen at the join.
Form requests ("two choruses then a bridge") get almost no response; it returns continuous texture, not a piece with sections.
Vocal requests become indistinct vocal-like texture with no intelligible words, so do not expect singing.
Low-end level varies widely between runs from one description, so a set of cues needs per-file normalisation.
The strength of audio-to-audio is hard to steer; push it and the source loses its identifying features.
Very short clips often carry fades at both ends, so percussion samples need hand-trimming.

Terms, price and alternatives

Confirm: commercial scope for your tier, the conditions for re-licensing generated audio, the official statement on training data, and whether output may be redistributed as a sample library. Pricing is tiered; exact figures and regional availability belong on the official page. Alternatives: MusicGen for local runs and fine-tuning on your own data; Suno when you need a song with vocals. Neither is necessarily better at fine atmospheric texture, so choose by whether you need offline use or singing.

Listen in the library

creepyloop v2 aC10:57An ambient loop, to hear the layering of bed material; a reference for this kind of output, not a Stable Audio generation (C1)Detailsfunk-groove2 · 105 BPM · beat 4/4C10:28A 105 BPM funk kit, to hear the density of a loop; a reference for this kind of output, not a Stable Audio generation (C1)Detailsbeach 01C10:32A short beach cue, to hear ambience mixed with instruments; a reference for this kind of output, not a Stable Audio generation (C1)DetailsNew BritainC10:43A hymn tune, to hear a motif that works in two bars; a reference for this kind of output, not a Stable Audio generation (C1)Details

All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).

Sources

产品能力、时长设置与导出形态，依官方公开说明与条款页核对，未抄原文
失败模式依公开用户报告与创作实践归纳，未引用具体评测数据

Used for fact-checking only; all prose is original.